Dataset Viewer
Auto-converted to Parquet
id
stringlengths
10
10
title
stringlengths
14
156
abstract
stringlengths
279
1.89k
full_text
sequence
qas
sequence
figures_and_tables
sequence
1909.00694
Minimally Supervised Learning of Affective Events Using Discourse Relations
Recognizing affective events that trigger positive or negative sentiment has a wide range of natural language processing applications but remains a challenging problem mainly because the polarity of an event is not necessarily predictable from its constituent words. In this paper, we propose to propagate affective polarity using discourse relations. Our method is simple and only requires a very small seed lexicon and a large raw corpus. Our experiments using Japanese data show that our method learns affective events effectively without manually labeled data. It also improves supervised learning results when labeled data are small.
{ "section_name": [ "Introduction", "Related Work", "Proposed Method", "Proposed Method ::: Polarity Function", "Proposed Method ::: Discourse Relation-Based Event Pairs", "Proposed Method ::: Discourse Relation-Based Event Pairs ::: AL (Automatically Labeled Pairs)", "Proposed Method ::: Discourse Relation-Based Event Pairs ::: CA (Cause Pairs)", "Proposed Method ::: Discourse Relation-Based Event Pairs ::: CO (Concession Pairs)", "Proposed Method ::: Loss Functions", "Experiments", "Experiments ::: Dataset", "Experiments ::: Dataset ::: AL, CA, and CO", "Experiments ::: Dataset ::: ACP (ACP Corpus)", "Experiments ::: Model Configurations", "Experiments ::: Results and Discussion", "Conclusion", "Acknowledgments", "Appendices ::: Seed Lexicon ::: Positive Words", "Appendices ::: Seed Lexicon ::: Negative Words", "Appendices ::: Settings of Encoder ::: BiGRU", "Appendices ::: Settings of Encoder ::: BERT" ], "paragraphs": [ [ "Affective events BIBREF0 are events that typically affect people in positive or negative ways. For example, getting money and playing sports are usually positive to the experiencers; catching cold and losing one's wallet are negative. Understanding affective events is important to various natural language processing (NLP) applications such as dialogue systems BIBREF1, question-answering systems BIBREF2, and humor recognition BIBREF3. In this paper, we work on recognizing the polarity of an affective event that is represented by a score ranging from $-1$ (negative) to 1 (positive).", "Learning affective events is challenging because, as the examples above suggest, the polarity of an event is not necessarily predictable from its constituent words. Combined with the unbounded combinatorial nature of language, the non-compositionality of affective polarity entails the need for large amounts of world knowledge, which can hardly be learned from small annotated data.", "In this paper, we propose a simple and effective method for learning affective events that only requires a very small seed lexicon and a large raw corpus. As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event.", "We trained the models using a Japanese web corpus. Given the minimum amount of supervision, they performed well. In addition, the combination of annotated and unannotated data yielded a gain over a purely supervised baseline when labeled data were small." ], [ "Learning affective events is closely related to sentiment analysis. Whereas sentiment analysis usually focuses on the polarity of what are described (e.g., movies), we work on how people are typically affected by events. In sentiment analysis, much attention has been paid to compositionality. Word-level polarity BIBREF5, BIBREF6, BIBREF7 and the roles of negation and intensification BIBREF8, BIBREF6, BIBREF9 are among the most important topics. In contrast, we are more interested in recognizing the sentiment polarity of an event that pertains to commonsense knowledge (e.g., getting money and catching cold).", "Label propagation from seed instances is a common approach to inducing sentiment polarities. While BIBREF5 and BIBREF10 worked on word- and phrase-level polarities, BIBREF0 dealt with event-level polarities. BIBREF5 and BIBREF10 linked instances using co-occurrence information and/or phrase-level coordinations (e.g., “$A$ and $B$” and “$A$ but $B$”). We shift our scope to event pairs that are more complex than phrase pairs, and consequently exploit discourse connectives as event-level counterparts of phrase-level conjunctions.", "BIBREF0 constructed a network of events using word embedding-derived similarities. Compared with this method, our discourse relation-based linking of events is much simpler and more intuitive.", "Some previous studies made use of document structure to understand the sentiment. BIBREF11 proposed a sentiment-specific pre-training strategy using unlabeled dialog data (tweet-reply pairs). BIBREF12 proposed a method of building a polarity-tagged corpus (ACP Corpus). They automatically gathered sentences that had positive or negative opinions utilizing HTML layout structures in addition to linguistic patterns. Our method depends only on raw texts and thus has wider applicability.", "" ], [ "" ], [ "", "Our goal is to learn the polarity function $p(x)$, which predicts the sentiment polarity score of an event $x$. We approximate $p(x)$ by a neural network with the following form:", "${\\rm Encoder}$ outputs a vector representation of the event $x$. ${\\rm Linear}$ is a fully-connected layer and transforms the representation into a scalar. ${\\rm tanh}$ is the hyperbolic tangent and transforms the scalar into a score ranging from $-1$ to 1. In Section SECREF21, we consider two specific implementations of ${\\rm Encoder}$.", "" ], [ "Our method requires a very small seed lexicon and a large raw corpus. We assume that we can automatically extract discourse-tagged event pairs, $(x_{i1}, x_{i2})$ ($i=1, \\cdots $) from the raw corpus. We refer to $x_{i1}$ and $x_{i2}$ as former and latter events, respectively. As shown in Figure FIGREF1, we limit our scope to two discourse relations: Cause and Concession.", "The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event. We expect the model to automatically learn complex phenomena through label propagation. Based on the availability of scores and the types of discourse relations, we classify the extracted event pairs into the following three types.", "" ], [ "The seed lexicon matches (1) the latter event but (2) not the former event, and (3) their discourse relation type is Cause or Concession. If the discourse relation type is Cause, the former event is given the same score as the latter. Likewise, if the discourse relation type is Concession, the former event is given the opposite of the latter's score. They are used as reference scores during training.", "" ], [ "The seed lexicon matches neither the former nor the latter event, and their discourse relation type is Cause. We assume the two events have the same polarities.", "" ], [ "The seed lexicon matches neither the former nor the latter event, and their discourse relation type is Concession. We assume the two events have the reversed polarities.", "" ], [ "Using AL, CA, and CO data, we optimize the parameters of the polarity function $p(x)$. We define a loss function for each of the three types of event pairs and sum up the multiple loss functions.", "We use mean squared error to construct loss functions. For the AL data, the loss function is defined as:", "where $x_{i1}$ and $x_{i2}$ are the $i$-th pair of the AL data. $r_{i1}$ and $r_{i2}$ are the automatically-assigned scores of $x_{i1}$ and $x_{i2}$, respectively. $N_{\\rm AL}$ is the total number of AL pairs, and $\\lambda _{\\rm AL}$ is a hyperparameter.", "For the CA data, the loss function is defined as:", "$y_{i1}$ and $y_{i2}$ are the $i$-th pair of the CA pairs. $N_{\\rm CA}$ is the total number of CA pairs. $\\lambda _{\\rm CA}$ and $\\mu $ are hyperparameters. The first term makes the scores of the two events closer while the second term prevents the scores from shrinking to zero.", "The loss function for the CO data is defined analogously:", "The difference is that the first term makes the scores of the two events distant from each other.", "" ], [ "" ], [ "" ], [ "As a raw corpus, we used a Japanese web corpus that was compiled through the procedures proposed by BIBREF13. To extract event pairs tagged with discourse relations, we used the Japanese dependency parser KNP and in-house postprocessing scripts BIBREF14. KNP used hand-written rules to segment each sentence into what we conventionally called clauses (mostly consecutive text chunks), each of which contained one main predicate. KNP also identified the discourse relations of event pairs if explicit discourse connectives BIBREF4 such as “ので” (because) and “のに” (in spite of) were present. We treated Cause/Reason (原因・理由) and Condition (条件) in the original tagset BIBREF15 as Cause and Concession (逆接) as Concession, respectively. Here is an example of event pair extraction.", ". 重大な失敗を犯したので、仕事をクビになった。", "Because [I] made a serious mistake, [I] got fired.", "From this sentence, we extracted the event pair of “重大な失敗を犯す” ([I] make a serious mistake) and “仕事をクビになる” ([I] get fired), and tagged it with Cause.", "We constructed our seed lexicon consisting of 15 positive words and 15 negative words, as shown in Section SECREF27. From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO. We randomly selected subsets of AL event pairs such that positive and negative latter events were equal in size. We also sampled event pairs for each of CA and CO such that it was five times larger than AL. The results are shown in Table TABREF16." ], [ "We used the latest version of the ACP Corpus BIBREF12 for evaluation. It was used for (semi-)supervised training as well. Extracted from Japanese websites using HTML layouts and linguistic patterns, the dataset covered various genres. For example, the following two sentences were labeled positive and negative, respectively:", ". 作業が楽だ。", "The work is easy.", ". 駐車場がない。", "There is no parking lot.", "Although the ACP corpus was originally constructed in the context of sentiment analysis, we found that it could roughly be regarded as a collection of affective events. We parsed each sentence and extracted the last clause in it. The train/dev/test split of the data is shown in Table TABREF19.", "The objective function for supervised training is:", "", "where $v_i$ is the $i$-th event, $R_i$ is the reference score of $v_i$, and $N_{\\rm ACP}$ is the number of the events of the ACP Corpus.", "To optimize the hyperparameters, we used the dev set of the ACP Corpus. For the evaluation, we used the test set of the ACP Corpus. The model output was classified as positive if $p(x) > 0$ and negative if $p(x) \\le 0$.", "" ], [ "As for ${\\rm Encoder}$, we compared two types of neural networks: BiGRU and BERT. GRU BIBREF16 is a recurrent neural network sequence encoder. BiGRU reads an input sequence forward and backward and the output is the concatenation of the final forward and backward hidden states.", "BERT BIBREF17 is a pre-trained multi-layer bidirectional Transformer BIBREF18 encoder. Its output is the final hidden state corresponding to the special classification tag ([CLS]). For the details of ${\\rm Encoder}$, see Sections SECREF30.", "We trained the model with the following four combinations of the datasets: AL, AL+CA+CO (two proposed models), ACP (supervised), and ACP+AL+CA+CO (semi-supervised). The corresponding objective functions were: $\\mathcal {L}_{\\rm AL}$, $\\mathcal {L}_{\\rm AL} + \\mathcal {L}_{\\rm CA} + \\mathcal {L}_{\\rm CO}$, $\\mathcal {L}_{\\rm ACP}$, and $\\mathcal {L}_{\\rm ACP} + \\mathcal {L}_{\\rm AL} + \\mathcal {L}_{\\rm CA} + \\mathcal {L}_{\\rm CO}$.", "" ], [ "", "Table TABREF23 shows accuracy. As the Random baseline suggests, positive and negative labels were distributed evenly. The Random+Seed baseline made use of the seed lexicon and output the corresponding label (or the reverse of it for negation) if the event's predicate is in the seed lexicon. We can see that the seed lexicon itself had practically no impact on prediction.", "The models in the top block performed considerably better than the random baselines. The performance gaps with their (semi-)supervised counterparts, shown in the middle block, were less than 7%. This demonstrates the effectiveness of discourse relation-based label propagation.", "Comparing the model variants, we obtained the highest score with the BiGRU encoder trained with the AL+CA+CO dataset. BERT was competitive but its performance went down if CA and CO were used in addition to AL. We conjecture that BERT was more sensitive to noises found more frequently in CA and CO.", "Contrary to our expectations, supervised models (ACP) outperformed semi-supervised models (ACP+AL+CA+CO). This suggests that the training set of 0.6 million events is sufficiently large for training the models. For comparison, we trained the models with a subset (6,000 events) of the ACP dataset. As the results shown in Table TABREF24 demonstrate, our method is effective when labeled data are small.", "The result of hyperparameter optimization for the BiGRU encoder was as follows:", "As the CA and CO pairs were equal in size (Table TABREF16), $\\lambda _{\\rm CA}$ and $\\lambda _{\\rm CO}$ were comparable values. $\\lambda _{\\rm CA}$ was about one-third of $\\lambda _{\\rm CO}$, and this indicated that the CA pairs were noisier than the CO pairs. A major type of CA pairs that violates our assumption was in the form of “$\\textit {problem}_{\\text{negative}}$ causes $\\textit {solution}_{\\text{positive}}$”:", ". (悪いところがある, よくなるように努力する)", "(there is a bad point, [I] try to improve [it])", "The polarities of the two events were reversed in spite of the Cause relation, and this lowered the value of $\\lambda _{\\rm CA}$.", "Some examples of model outputs are shown in Table TABREF26. The first two examples suggest that our model successfully learned negation without explicit supervision. Similarly, the next two examples differ only in voice but the model correctly recognized that they had opposite polarities. The last two examples share the predicate “落とす\" (drop) and only the objects are different. The second event “肩を落とす\" (lit. drop one's shoulders) is an idiom that expresses a disappointed feeling. The examples demonstrate that our model correctly learned non-compositional expressions.", "" ], [ "In this paper, we proposed to use discourse relations to effectively propagate polarities of affective events from seeds. Experiments show that, even with a minimal amount of supervision, the proposed method performed well.", "Although event pairs linked by discourse analysis are shown to be useful, they nevertheless contain noises. Adding linguistically-motivated filtering rules would help improve the performance." ], [ "We thank Nobuhiro Kaji for providing the ACP Corpus and Hirokazu Kiyomaru and Yudai Kishimoto for their help in extracting event pairs. This work was partially supported by Yahoo! Japan Corporation." ], [ "喜ぶ (rejoice), 嬉しい (be glad), 楽しい (be pleasant), 幸せ (be happy), 感動 (be impressed), 興奮 (be excited), 懐かしい (feel nostalgic), 好き (like), 尊敬 (respect), 安心 (be relieved), 感心 (admire), 落ち着く (be calm), 満足 (be satisfied), 癒される (be healed), and スッキリ (be refreshed)." ], [ "怒る (get angry), 悲しい (be sad), 寂しい (be lonely), 怖い (be scared), 不安 (feel anxious), 恥ずかしい (be embarrassed), 嫌 (hate), 落ち込む (feel down), 退屈 (be bored), 絶望 (feel hopeless), 辛い (have a hard time), 困る (have trouble), 憂鬱 (be depressed), 心配 (be worried), and 情けない (be sorry)." ], [ "The dimension of the embedding layer was 256. The embedding layer was initialized with the word embeddings pretrained using the Web corpus. The input sentences were segmented into words by the morphological analyzer Juman++. The vocabulary size was 100,000. The number of hidden layers was 2. The dimension of hidden units was 256. The optimizer was Momentum SGD BIBREF21. The mini-batch size was 1024. We ran 100 epochs and selected the snapshot that achieved the highest score for the dev set." ], [ "We used a Japanese BERT model pretrained with Japanese Wikipedia. The input sentences were segmented into words by Juman++, and words were broken into subwords by applying BPE BIBREF20. The vocabulary size was 32,000. The maximum length of an input sequence was 128. The number of hidden layers was 12. The dimension of hidden units was 768. The number of self-attention heads was 12. The optimizer was Adam BIBREF19. The mini-batch size was 32. We ran 1 epoch." ] ] }
{ "question": [ "What is the seed lexicon?", "What are the results?", "How are relations used to propagate polarity?", "How big is the Japanese data?", "What are labels available in dataset for supervision?", "How big are improvements of supervszed learning results trained on smalled labeled data enhanced with proposed approach copared to basic approach?", "How does their model learn using mostly raw data?", "How big is seed lexicon used for training?", "How large is raw corpus used for training?" ], "question_id": [ "753990d0b621d390ed58f20c4d9e4f065f0dc672", "9d578ddccc27dd849244d632dd0f6bf27348ad81", "02e4bf719b1a504e385c35c6186742e720bcb281", "44c4bd6decc86f1091b5fc0728873d9324cdde4e", "86abeff85f3db79cf87a8c993e5e5aa61226dc98", "c029deb7f99756d2669abad0a349d917428e9c12", "39f8db10d949c6b477fa4b51e7c184016505884f", "d0bc782961567dc1dd7e074b621a6d6be44bb5b4", "a592498ba2fac994cd6fad7372836f0adb37e22a" ], "nlp_background": [ "two", "two", "two", "two", "zero", "zero", "zero", "zero", "zero" ], "topic_background": [ "unfamiliar", "unfamiliar", "unfamiliar", "unfamiliar", "unfamiliar", "unfamiliar", "unfamiliar", "unfamiliar", "unfamiliar" ], "paper_read": [ "no", "no", "no", "no", "no", "no", "no", "no", "no" ], "search_query": [ "", "", "", "", "", "", "", "", "" ], "question_writer": [ "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4", "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4", "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4", "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4", "258ee4069f740c400c0049a2580945a1cc7f044c", "258ee4069f740c400c0049a2580945a1cc7f044c", "258ee4069f740c400c0049a2580945a1cc7f044c", "258ee4069f740c400c0049a2580945a1cc7f044c", "258ee4069f740c400c0049a2580945a1cc7f044c" ], "answers": [ { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "a vocabulary of positive and negative predicates that helps determine the polarity score of an event", "evidence": [ "The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event. We expect the model to automatically learn complex phenomena through label propagation. Based on the availability of scores and the types of discourse relations, we classify the extracted event pairs into the following three types." ], "highlighted_evidence": [ "The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event.", "It is a " ] }, { "unanswerable": false, "extractive_spans": [ "seed lexicon consists of positive and negative predicates" ], "yes_no": null, "free_form_answer": "", "evidence": [ "The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event. We expect the model to automatically learn complex phenomena through label propagation. Based on the availability of scores and the types of discourse relations, we classify the extracted event pairs into the following three types." ], "highlighted_evidence": [ "The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event." ] } ], "annotation_id": [ "31e85022a847f37c15fd0415f3c450c74c8e4755", "95da0a6e1b08db74a405c6a71067c9b272a50ff5" ], "worker_id": [ "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4", "2cfd959e433f290bb50b55722370f0d22fe090b7" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "Using all data to train: AL -- BiGRU achieved 0.843 accuracy, AL -- BERT achieved 0.863 accuracy, AL+CA+CO -- BiGRU achieved 0.866 accuracy, AL+CA+CO -- BERT achieved 0.835, accuracy, ACP -- BiGRU achieved 0.919 accuracy, ACP -- BERT achived 0.933, accuracy, ACP+AL+CA+CO -- BiGRU achieved 0.917 accuracy, ACP+AL+CA+CO -- BERT achieved 0.913 accuracy. \nUsing a subset to train: BERT achieved 0.876 accuracy using ACP (6K), BERT achieved 0.886 accuracy using ACP (6K) + AL, BiGRU achieved 0.830 accuracy using ACP (6K), BiGRU achieved 0.879 accuracy using ACP (6K) + AL + CA + CO.", "evidence": [ "FLOAT SELECTED: Table 3: Performance of various models on the ACP test set.", "FLOAT SELECTED: Table 4: Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data.", "As for ${\\rm Encoder}$, we compared two types of neural networks: BiGRU and BERT. GRU BIBREF16 is a recurrent neural network sequence encoder. BiGRU reads an input sequence forward and backward and the output is the concatenation of the final forward and backward hidden states.", "We trained the model with the following four combinations of the datasets: AL, AL+CA+CO (two proposed models), ACP (supervised), and ACP+AL+CA+CO (semi-supervised). The corresponding objective functions were: $\\mathcal {L}_{\\rm AL}$, $\\mathcal {L}_{\\rm AL} + \\mathcal {L}_{\\rm CA} + \\mathcal {L}_{\\rm CO}$, $\\mathcal {L}_{\\rm ACP}$, and $\\mathcal {L}_{\\rm ACP} + \\mathcal {L}_{\\rm AL} + \\mathcal {L}_{\\rm CA} + \\mathcal {L}_{\\rm CO}$." ], "highlighted_evidence": [ "FLOAT SELECTED: Table 3: Performance of various models on the ACP test set.", "FLOAT SELECTED: Table 4: Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data.", "As for ${\\rm Encoder}$, we compared two types of neural networks: BiGRU and BERT. ", "We trained the model with the following four combinations of the datasets: AL, AL+CA+CO (two proposed models), ACP (supervised), and ACP+AL+CA+CO (semi-supervised). The corresponding objective functions were: $\\mathcal {L}_{\\rm AL}$, $\\mathcal {L}_{\\rm AL} + \\mathcal {L}_{\\rm CA} + \\mathcal {L}_{\\rm CO}$, $\\mathcal {L}_{\\rm ACP}$, and $\\mathcal {L}_{\\rm ACP} + \\mathcal {L}_{\\rm AL} + \\mathcal {L}_{\\rm CA} + \\mathcal {L}_{\\rm CO}$." ] } ], "annotation_id": [ "1e5e867244ea656c4b7632628086209cf9bae5fa" ], "worker_id": [ "2cfd959e433f290bb50b55722370f0d22fe090b7" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "based on the relation between events, the suggested polarity of one event can determine the possible polarity of the other event ", "evidence": [ "In this paper, we propose a simple and effective method for learning affective events that only requires a very small seed lexicon and a large raw corpus. As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event." ], "highlighted_evidence": [ "As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event." ] }, { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "cause relation: both events in the relation should have the same polarity; concession relation: events should have opposite polarity", "evidence": [ "In this paper, we propose a simple and effective method for learning affective events that only requires a very small seed lexicon and a large raw corpus. As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event.", "The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event. We expect the model to automatically learn complex phenomena through label propagation. Based on the availability of scores and the types of discourse relations, we classify the extracted event pairs into the following three types." ], "highlighted_evidence": [ "As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event.", "The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event. We expect the model to automatically learn complex phenomena through label propagation." ] } ], "annotation_id": [ "49a78a07d2eed545556a835ccf2eb40e5eee9801", "acd6d15bd67f4b1496ee8af1c93c33e7d59c89e1" ], "worker_id": [ "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4", "2cfd959e433f290bb50b55722370f0d22fe090b7" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "7000000 pairs of events were extracted from the Japanese Web corpus, 529850 pairs of events were extracted from the ACP corpus", "evidence": [ "As a raw corpus, we used a Japanese web corpus that was compiled through the procedures proposed by BIBREF13. To extract event pairs tagged with discourse relations, we used the Japanese dependency parser KNP and in-house postprocessing scripts BIBREF14. KNP used hand-written rules to segment each sentence into what we conventionally called clauses (mostly consecutive text chunks), each of which contained one main predicate. KNP also identified the discourse relations of event pairs if explicit discourse connectives BIBREF4 such as “ので” (because) and “のに” (in spite of) were present. We treated Cause/Reason (原因・理由) and Condition (条件) in the original tagset BIBREF15 as Cause and Concession (逆接) as Concession, respectively. Here is an example of event pair extraction.", "We constructed our seed lexicon consisting of 15 positive words and 15 negative words, as shown in Section SECREF27. From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO. We randomly selected subsets of AL event pairs such that positive and negative latter events were equal in size. We also sampled event pairs for each of CA and CO such that it was five times larger than AL. The results are shown in Table TABREF16.", "FLOAT SELECTED: Table 1: Statistics of the AL, CA, and CO datasets.", "We used the latest version of the ACP Corpus BIBREF12 for evaluation. It was used for (semi-)supervised training as well. Extracted from Japanese websites using HTML layouts and linguistic patterns, the dataset covered various genres. For example, the following two sentences were labeled positive and negative, respectively:", "Although the ACP corpus was originally constructed in the context of sentiment analysis, we found that it could roughly be regarded as a collection of affective events. We parsed each sentence and extracted the last clause in it. The train/dev/test split of the data is shown in Table TABREF19.", "FLOAT SELECTED: Table 2: Details of the ACP dataset." ], "highlighted_evidence": [ "As a raw corpus, we used a Japanese web corpus that was compiled through the procedures proposed by BIBREF13. ", "From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO. We randomly selected subsets of AL event pairs such that positive and negative latter events were equal in size. We also sampled event pairs for each of CA and CO such that it was five times larger than AL. The results are shown in Table TABREF16.", "FLOAT SELECTED: Table 1: Statistics of the AL, CA, and CO datasets.", "We used the latest version of the ACP Corpus BIBREF12 for evaluation. It was used for (semi-)supervised training as well.", "Although the ACP corpus was originally constructed in the context of sentiment analysis, we found that it could roughly be regarded as a collection of affective events. We parsed each sentence and extracted the last clause in it. The train/dev/test split of the data is shown in Table TABREF19.", "FLOAT SELECTED: Table 2: Details of the ACP dataset." ] }, { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "The ACP corpus has around 700k events split into positive and negative polarity ", "evidence": [ "FLOAT SELECTED: Table 2: Details of the ACP dataset." ], "highlighted_evidence": [ "FLOAT SELECTED: Table 2: Details of the ACP dataset." ] } ], "annotation_id": [ "36926a4c9e14352c91111150aa4c6edcc5c0770f", "75b6dd28ccab20a70087635d89c2b22d0e99095c" ], "worker_id": [ "2cfd959e433f290bb50b55722370f0d22fe090b7", "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "negative", "positive" ], "yes_no": null, "free_form_answer": "", "evidence": [ "Affective events BIBREF0 are events that typically affect people in positive or negative ways. For example, getting money and playing sports are usually positive to the experiencers; catching cold and losing one's wallet are negative. Understanding affective events is important to various natural language processing (NLP) applications such as dialogue systems BIBREF1, question-answering systems BIBREF2, and humor recognition BIBREF3. In this paper, we work on recognizing the polarity of an affective event that is represented by a score ranging from $-1$ (negative) to 1 (positive)." ], "highlighted_evidence": [ "In this paper, we work on recognizing the polarity of an affective event that is represented by a score ranging from $-1$ (negative) to 1 (positive)." ] } ], "annotation_id": [ "2d8c7df145c37aad905e48f64d8caa69e54434d4" ], "worker_id": [ "c1018a31c3272ce74964a3280069f62f314a1a58" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "3%", "evidence": [ "FLOAT SELECTED: Table 4: Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data." ], "highlighted_evidence": [ "FLOAT SELECTED: Table 4: Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data." ] } ], "annotation_id": [ "df4372b2e8d9bb2039a5582f192768953b01d904" ], "worker_id": [ "c1018a31c3272ce74964a3280069f62f314a1a58" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "by exploiting discourse relations to propagate polarity from seed predicates to final sentiment polarity", "evidence": [ "In this paper, we propose a simple and effective method for learning affective events that only requires a very small seed lexicon and a large raw corpus. As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event." ], "highlighted_evidence": [ "In this paper, we propose a simple and effective method for learning affective events that only requires a very small seed lexicon and a large raw corpus. As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive)." ] } ], "annotation_id": [ "5c5bbc8af91c16af89b4ddd57ee6834be018e4e7" ], "worker_id": [ "c1018a31c3272ce74964a3280069f62f314a1a58" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "30 words", "evidence": [ "We constructed our seed lexicon consisting of 15 positive words and 15 negative words, as shown in Section SECREF27. From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO. We randomly selected subsets of AL event pairs such that positive and negative latter events were equal in size. We also sampled event pairs for each of CA and CO such that it was five times larger than AL. The results are shown in Table TABREF16." ], "highlighted_evidence": [ "We constructed our seed lexicon consisting of 15 positive words and 15 negative words, as shown in Section SECREF27. " ] } ], "annotation_id": [ "0206f2131f64a3e02498cedad1250971b78ffd0c" ], "worker_id": [ "c1018a31c3272ce74964a3280069f62f314a1a58" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "100 million sentences" ], "yes_no": null, "free_form_answer": "", "evidence": [ "As a raw corpus, we used a Japanese web corpus that was compiled through the procedures proposed by BIBREF13. To extract event pairs tagged with discourse relations, we used the Japanese dependency parser KNP and in-house postprocessing scripts BIBREF14. KNP used hand-written rules to segment each sentence into what we conventionally called clauses (mostly consecutive text chunks), each of which contained one main predicate. KNP also identified the discourse relations of event pairs if explicit discourse connectives BIBREF4 such as “ので” (because) and “のに” (in spite of) were present. We treated Cause/Reason (原因・理由) and Condition (条件) in the original tagset BIBREF15 as Cause and Concession (逆接) as Concession, respectively. Here is an example of event pair extraction.", "We constructed our seed lexicon consisting of 15 positive words and 15 negative words, as shown in Section SECREF27. From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO. We randomly selected subsets of AL event pairs such that positive and negative latter events were equal in size. We also sampled event pairs for each of CA and CO such that it was five times larger than AL. The results are shown in Table TABREF16." ], "highlighted_evidence": [ "As a raw corpus, we used a Japanese web corpus that was compiled through the procedures proposed by BIBREF13. ", "From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO." ] } ], "annotation_id": [ "c36bad2758c4f9866d64c357c475d370595d937f" ], "worker_id": [ "c1018a31c3272ce74964a3280069f62f314a1a58" ] } ] }
{ "caption": [ "Figure 1: An overview of our method. We focus on pairs of events, the former events and the latter events, which are connected with a discourse relation, CAUSE or CONCESSION. Dropped pronouns are indicated by brackets in English translations. We divide the event pairs into three types: AL, CA, and CO. In AL, the polarity of a latter event is automatically identified as either positive or negative, according to the seed lexicon (the positive word is colored red and the negative word blue). We propagate the latter event’s polarity to the former event. The same polarity as the latter event is used for the discourse relation CAUSE, and the reversed polarity for CONCESSION. In CA and CO, the latter event’s polarity is not known. Depending on the discourse relation, we encourage the two events’ polarities to be the same (CA) or reversed (CO). Details are given in Section 3.2.", "Table 1: Statistics of the AL, CA, and CO datasets.", "Table 2: Details of the ACP dataset.", "Table 5: Examples of polarity scores predicted by the BiGRU model trained with AL+CA+CO.", "Table 3: Performance of various models on the ACP test set.", "Table 4: Results for small labeled training data. Given the performance with the full dataset, we show BERT trained only with the AL data." ], "file": [ "2-Figure1-1.png", "4-Table1-1.png", "4-Table2-1.png", "5-Table5-1.png", "5-Table3-1.png", "5-Table4-1.png" ] }
1908.06606
Question Answering based Clinical Text Structuring Using Pre-trained Language Model
Clinical text structuring is a critical and fundamental task for clinical research. Traditional methods such as taskspecific end-to-end models and pipeline models usually suffer from the lack of dataset and error propagation. In this paper, we present a question answering based clinical text structuring (QA-CTS) task to unify different specific tasks and make dataset shareable. A novel model that aims to introduce domain-specific features (e.g., clinical named entity information) into pre-trained language model is also proposed for QA-CTS task. Experimental results on Chinese pathology reports collected from Ruijing Hospital demonstrate our presented QA-CTS task is very effective to improve the performance on specific tasks. Our proposed model also competes favorably with strong baseline models in specific tasks.
{ "section_name": [ "Introduction", "Related Work ::: Clinical Text Structuring", "Related Work ::: Pre-trained Language Model", "Question Answering based Clinical Text Structuring", "The Proposed Model for QA-CTS Task", "The Proposed Model for QA-CTS Task ::: Contextualized Representation of Sentence Text and Query Text", "The Proposed Model for QA-CTS Task ::: Clinical Named Entity Information", "The Proposed Model for QA-CTS Task ::: Integration Method", "The Proposed Model for QA-CTS Task ::: Final Prediction", "The Proposed Model for QA-CTS Task ::: Two-Stage Training Mechanism", "Experimental Studies", "Experimental Studies ::: Dataset and Evaluation Metrics", "Experimental Studies ::: Experimental Settings", "Experimental Studies ::: Comparison with State-of-the-art Methods", "Experimental Studies ::: Ablation Analysis", "Experimental Studies ::: Comparisons Between Two Integration Methods", "Experimental Studies ::: Data Integration Analysis", "Conclusion", "Acknowledgment" ], "paragraphs": [ [ "Clinical text structuring (CTS) is a critical task for fetching medical research data from electronic health records (EHRs), where structural patient medical data, such as whether the patient has specific symptoms, diseases, or what the tumor size is, how far from the tumor is cut at during the surgery, or what the specific laboratory test result is, are obtained. It is important to extract structured data from clinical text because bio-medical systems or bio-medical researches greatly rely on structured data but they cannot obtain them directly. In addition, clinical text often contains abundant healthcare information. CTS is able to provide large-scale extracted structured data for enormous down-stream clinical researches.", "However, end-to-end CTS is a very challenging task. Different CTS tasks often have non-uniform output formats, such as specific-class classifications (e.g. tumor stage), strings in the original text (e.g. result for a laboratory test) and inferred values from part of the original text (e.g. calculated tumor size). Researchers have to construct different models for it, which is already costly, and hence it calls for a lot of labeled data for each model. Moreover, labeling necessary amount of data for training neural network requires expensive labor cost. To handle it, researchers turn to some rule-based structuring methods which often have lower labor cost.", "Traditionally, CTS tasks can be addressed by rule and dictionary based methods BIBREF0, BIBREF1, BIBREF2, task-specific end-to-end methods BIBREF3, BIBREF4, BIBREF5, BIBREF6 and pipeline methods BIBREF7, BIBREF8, BIBREF9. Rule and dictionary based methods suffer from costly human-designed extraction rules, while task-specific end-to-end methods have non-uniform output formats and require task-specific training dataset. Pipeline methods break down the entire process into several pieces which improves the performance and generality. However, when the pipeline depth grows, error propagation will have a greater impact on the performance.", "To reduce the pipeline depth and break the barrier of non-uniform output formats, we present a question answering based clinical text structuring (QA-CTS) task (see Fig. FIGREF1). Unlike the traditional CTS task, our QA-CTS task aims to discover the most related text from original paragraph text. For some cases, it is already the final answer in deed (e.g., extracting sub-string). While for other cases, it needs several steps to obtain the final answer, such as entity names conversion and negative words recognition. Our presented QA-CTS task unifies the output format of the traditional CTS task and make the training data shareable, thus enriching the training data. The main contributions of this work can be summarized as follows.", "We first present a question answering based clinical text structuring (QA-CTS) task, which unifies different specific tasks and make dataset shareable. We also propose an effective model to integrate clinical named entity information into pre-trained language model.", "Experimental results show that QA-CTS task leads to significant improvement due to shared dataset. Our proposed model also achieves significantly better performance than the strong baseline methods. In addition, we also show that two-stage training mechanism has a great improvement on QA-CTS task.", "The rest of the paper is organized as follows. We briefly review the related work on clinical text structuring in Section SECREF2. Then, we present question answer based clinical text structuring task in Section SECREF3. In Section SECREF4, we present an effective model for this task. Section SECREF5 is devoted to computational studies and several investigations on the key issues of our proposed model. Finally, conclusions are given in Section SECREF6." ], [ "Clinical text structuring is a final problem which is highly related to practical applications. Most of existing studies are case-by-case. Few of them are developed for the general purpose structuring task. These studies can be roughly divided into three categories: rule and dictionary based methods, task-specific end-to-end methods and pipeline methods.", "Rule and dictionary based methods BIBREF0, BIBREF1, BIBREF2 rely extremely on heuristics and handcrafted extraction rules which is more of an art than a science and incurring extensive trial-and-error experiments. Fukuda et al. BIBREF0 identified protein names from biological papers by dictionaries and several features of protein names. Wang et al. BIBREF1 developed some linguistic rules (i.e. normalised/expanded term matching and substring term matching) to map specific terminology to SNOMED CT. Song et al. BIBREF2 proposed a hybrid dictionary-based bio-entity extraction technique and expands the bio-entity dictionary by combining different data sources and improves the recall rate through the shortest path edit distance algorithm. This kind of approach features its interpretability and easy modifiability. However, with the increase of the rule amount, supplementing new rules to existing system will turn to be a rule disaster.", "Task-specific end-to-end methods BIBREF3, BIBREF4 use large amount of data to automatically model the specific task. Topaz et al. BIBREF3 constructed an automated wound information identification model with five output. Tan et al. BIBREF4 identified patients undergoing radical cystectomy for bladder cancer. Although they achieved good performance, none of their models could be used to another task due to output format difference. This makes building a new model for a new task a costly job.", "Pipeline methods BIBREF7, BIBREF8, BIBREF9 break down the entire task into several basic natural language processing tasks. Bill et al. BIBREF7 focused on attributes extraction which mainly relied on dependency parsing and named entity recognition BIBREF10, BIBREF11, BIBREF12. Meanwhile, Fonferko et al. BIBREF9 used more components like noun phrase chunking BIBREF13, BIBREF14, BIBREF15, part-of-speech tagging BIBREF16, BIBREF17, BIBREF18, sentence splitter, named entity linking BIBREF19, BIBREF20, BIBREF21, relation extraction BIBREF22, BIBREF23. This kind of method focus on language itself, so it can handle tasks more general. However, as the depth of pipeline grows, it is obvious that error propagation will be more and more serious. In contrary, using less components to decrease the pipeline depth will lead to a poor performance. So the upper limit of this method depends mainly on the worst component." ], [ "Recently, some works focused on pre-trained language representation models to capture language information from text and then utilizing the information to improve the performance of specific natural language processing tasks BIBREF24, BIBREF25, BIBREF26, BIBREF27 which makes language model a shared model to all natural language processing tasks. Radford et al. BIBREF24 proposed a framework for fine-tuning pre-trained language model. Peters et al. BIBREF25 proposed ELMo which concatenates forward and backward language models in a shallow manner. Devlin et al. BIBREF26 used bidirectional Transformers to model deep interactions between the two directions. Yang et al. BIBREF27 replaced the fixed forward or backward factorization order with all possible permutations of the factorization order and avoided using the [MASK] tag which causes pretrain-finetune discrepancy that BERT is subject to.", "The main motivation of introducing pre-trained language model is to solve the shortage of labeled data and polysemy problem. Although polysemy problem is not a common phenomenon in biomedical domain, shortage of labeled data is always a non-trivial problem. Lee et al. BIBREF28 applied BERT on large-scale biomedical unannotated data and achieved improvement on biomedical named entity recognition, relation extraction and question answering. Kim et al. BIBREF29 adapted BioBERT into multi-type named entity recognition and discovered new entities. Both of them demonstrates the usefulness of introducing pre-trained language model into biomedical domain." ], [ "Given a sequence of paragraph text $X=<x_1, x_2, ..., x_n>$, clinical text structuring (CTS) can be regarded to extract or generate a key-value pair where key $Q$ is typically a query term such as proximal resection margin and value $V$ is a result of query term $Q$ according to the paragraph text $X$.", "Generally, researchers solve CTS problem in two steps. Firstly, the answer-related text is pick out. And then several steps such as entity names conversion and negative words recognition are deployed to generate the final answer. While final answer varies from task to task, which truly causes non-uniform output formats, finding the answer-related text is a common action among all tasks. Traditional methods regard both the steps as a whole. In this paper, we focus on finding the answer-related substring $Xs = <X_i, X_i+1, X_i+2, ... X_j> (1 <= i < j <= n)$ from paragraph text $X$. For example, given sentence UTF8gkai“远端胃切除标本:小弯长11.5cm,大弯长17.0cm。距上切端6.0cm、下切端8.0cm\" (Distal gastrectomy specimen: measuring 11.5cm in length along the lesser curvature, 17.0cm in length along the greater curvature; 6.0cm from the proximal resection margin, and 8.0cm from the distal resection margin) and query UTF8gkai“上切缘距离\"(proximal resection margin), the answer should be 6.0cm which is located in original text from index 32 to 37. With such definition, it unifies the output format of CTS tasks and therefore make the training data shareable, in order to reduce the training data quantity requirement.", "Since BERT BIBREF26 has already demonstrated the usefulness of shared model, we suppose extracting commonality of this problem and unifying the output format will make the model more powerful than dedicated model and meanwhile, for a specific clinical task, use the data for other tasks to supplement the training data." ], [ "In this section, we present an effective model for the question answering based clinical text structuring (QA-CTS). As shown in Fig. FIGREF8, paragraph text $X$ is first passed to a clinical named entity recognition (CNER) model BIBREF12 to capture named entity information and obtain one-hot CNER output tagging sequence for query text $I_{nq}$ and paragraph text $I_{nt}$ with BIEOS (Begin, Inside, End, Outside, Single) tag scheme. $I_{nq}$ and $I_{nt}$ are then integrated together into $I_n$. Meanwhile, the paragraph text $X$ and query text $Q$ are organized and passed to contextualized representation model which is pre-trained language model BERT BIBREF26 here to obtain the contextualized representation vector $V_s$ of both text and query. Afterwards, $V_s$ and $I_n$ are integrated together and fed into a feed forward network to calculate the start and end index of answer-related text. Here we define this calculation problem as a classification for each word to be the start or end word." ], [ "For any clinical free-text paragraph $X$ and query $Q$, contextualized representation is to generate the encoded vector of both of them. Here we use pre-trained language model BERT-base BIBREF26 model to capture contextual information.", "The text input is constructed as `[CLS] $Q$ [SEP] $X$ [SEP]'. For Chinese sentence, each word in this input will be mapped to a pre-trained embedding $e_i$. To tell the model $Q$ and $X$ is two different sentence, a sentence type input is generated which is a binary label sequence to denote what sentence each character in the input belongs to. Positional encoding and mask matrix is also constructed automatically to bring in absolute position information and eliminate the impact of zero padding respectively. Then a hidden vector $V_s$ which contains both query and text information is generated through BERT-base model." ], [ "Since BERT is trained on general corpus, its performance on biomedical domain can be improved by introducing biomedical domain-specific features. In this paper, we introduce clinical named entity information into the model.", "The CNER task aims to identify and classify important clinical terms such as diseases, symptoms, treatments, exams, and body parts from Chinese EHRs. It can be regarded as a sequence labeling task. A CNER model typically outputs a sequence of tags. Each character of the original sentence will be tagged a label following a tag scheme. In this paper we recognize the entities by the model of our previous work BIBREF12 but trained on another corpus which has 44 entity types including operations, numbers, unit words, examinations, symptoms, negative words, etc. An illustrative example of named entity information sequence is demonstrated in Table TABREF2. In Table TABREF2, UTF8gkai“远端胃切除\" is tagged as an operation, `11.5' is a number word and `cm' is an unit word. The named entity tag sequence is organized in one-hot type. We denote the sequence for clinical sentence and query term as $I_{nt}$ and $I_{nq}$, respectively." ], [ "There are two ways to integrate two named entity information vectors $I_{nt}$ and $I_{nq}$ or hidden contextualized representation $V_s$ and named entity information $I_n$, where $I_n = [I_{nt}; I_{nq}]$. The first one is to concatenate them together because they have sequence output with a common dimension. The second one is to transform them into a new hidden representation. For the concatenation method, the integrated representation is described as follows.", "While for the transformation method, we use multi-head attention BIBREF30 to encode the two vectors. It can be defined as follows where $h$ is the number of heads and $W_o$ is used to projects back the dimension of concatenated matrix.", "$Attention$ denotes the traditional attention and it can be defined as follows.", "where $d_k$ is the length of hidden vector." ], [ "The final step is to use integrated representation $H_i$ to predict the start and end index of answer-related text. Here we define this calculation problem as a classification for each word to be the start or end word. We use a feed forward network (FFN) to compress and calculate the score of each word $H_f$ which makes the dimension to $\\left\\langle l_s, 2\\right\\rangle $ where $l_s$ denotes the length of sequence.", "Then we permute the two dimensions for softmax calculation. The calculation process of loss function can be defined as followed.", "where $O_s = softmax(permute(H_f)_0)$ denotes the probability score of each word to be the start word and similarly $O_e = softmax(permute(H_f)_1)$ denotes the end. $y_s$ and $y_e$ denotes the true answer of the output for start word and end word respectively." ], [ "Two-stage training mechanism is previously applied on bilinear model in fine-grained visual recognition BIBREF31, BIBREF32, BIBREF33. Two CNNs are deployed in the model. One is trained at first for coarse-graind features while freezing the parameter of the other. Then unfreeze the other one and train the entire model in a low learning rate for fetching fine-grained features.", "Inspired by this and due to the large amount of parameters in BERT model, to speed up the training process, we fine tune the BERT model with new prediction layer first to achieve a better contextualized representation performance. Then we deploy the proposed model and load the fine tuned BERT weights, attach named entity information layers and retrain the model." ], [ "In this section, we devote to experimentally evaluating our proposed task and approach. The best results in tables are in bold." ], [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs. All question-answer pairs are annotated and reviewed by four clinicians with three types of questions, namely tumor size, proximal resection margin and distal resection margin. These annotated instances have been partitioned into 1,899 training instances (12,412 sentences) and 815 test instances (5,421 sentences). Each instance has one or several sentences. Detailed statistics of different types of entities are listed in Table TABREF20.", "In the following experiments, two widely-used performance measures (i.e., EM-score BIBREF34 and (macro-averaged) F$_1$-score BIBREF35) are used to evaluate the methods. The Exact Match (EM-score) metric measures the percentage of predictions that match any one of the ground truth answers exactly. The F$_1$-score metric is a looser metric measures the average overlap between the prediction and ground truth answer." ], [ "To implement deep neural network models, we utilize the Keras library BIBREF36 with TensorFlow BIBREF37 backend. Each model is run on a single NVIDIA GeForce GTX 1080 Ti GPU. The models are trained by Adam optimization algorithm BIBREF38 whose parameters are the same as the default settings except for learning rate set to $5\\times 10^{-5}$. Batch size is set to 3 or 4 due to the lack of graphical memory. We select BERT-base as the pre-trained language model in this paper. Due to the high cost of pre-training BERT language model, we directly adopt parameters pre-trained by Google in Chinese general corpus. The named entity recognition is applied on both pathology report texts and query texts." ], [ "Since BERT has already achieved the state-of-the-art performance of question-answering, in this section we compare our proposed model with state-of-the-art question answering models (i.e. QANet BIBREF39) and BERT-Base BIBREF26. As BERT has two versions: BERT-Base and BERT-Large, due to the lack of computational resource, we can only compare with BERT-Base model instead of BERT-Large. Prediction layer is attached at the end of the original BERT-Base model and we fine tune it on our dataset. In this section, the named entity integration method is chosen to pure concatenation (Concatenate the named entity information on pathology report text and query text first and then concatenate contextualized representation and concatenated named entity information). Comparative results are summarized in Table TABREF23.", "Table TABREF23 indicates that our proposed model achieved the best performance both in EM-score and F$_1$-score with EM-score of 91.84% and F$_1$-score of 93.75%. QANet outperformed BERT-Base with 3.56% score in F$_1$-score but underperformed it with 0.75% score in EM-score. Compared with BERT-Base, our model led to a 5.64% performance improvement in EM-score and 3.69% in F$_1$-score. Although our model didn't outperform much with QANet in F$_1$-score (only 0.13%), our model significantly outperformed it with 6.39% score in EM-score." ], [ "To further investigate the effects of named entity information and two-stage training mechanism for our model, we apply ablation analysis to see the improvement brought by each of them, where $\\times $ refers to removing that part from our model.", "As demonstrated in Table TABREF25, with named entity information enabled, two-stage training mechanism improved the result by 4.36% in EM-score and 3.8% in F$_1$-score. Without two-stage training mechanism, named entity information led to an improvement by 1.28% in EM-score but it also led to a weak deterioration by 0.12% in F$_1$-score. With both of them enabled, our proposed model achieved a 5.64% score improvement in EM-score and a 3.69% score improvement in F$_1$-score. The experimental results show that both named entity information and two-stage training mechanism are helpful to our model." ], [ "There are two methods to integrate named entity information into existing model, we experimentally compare these two integration methods. As named entity recognition has been applied on both pathology report text and query text, there will be two integration here. One is for two named entity information and the other is for contextualized representation and integrated named entity information. For multi-head attention BIBREF30, we set heads number $h = 16$ with 256-dimension hidden vector size for each head.", "From Table TABREF27, we can observe that applying concatenation on both periods achieved the best performance on both EM-score and F$_1$-score. Unfortunately, applying multi-head attention on both period one and period two can not reach convergence in our experiments. This probably because it makes the model too complex to train. The difference on other two methods are the order of concatenation and multi-head attention. Applying multi-head attention on two named entity information $I_{nt}$ and $I_{nq}$ first achieved a better performance with 89.87% in EM-score and 92.88% in F$_1$-score. Applying Concatenation first can only achieve 80.74% in EM-score and 84.42% in F$_1$-score. This is probably due to the processing depth of hidden vectors and dataset size. BERT's output has been modified after many layers but named entity information representation is very close to input. With big amount of parameters in multi-head attention, it requires massive training to find out the optimal parameters. However, our dataset is significantly smaller than what pre-trained BERT uses. This probably can also explain why applying multi-head attention method on both periods can not converge.", "Although Table TABREF27 shows the best integration method is concatenation, multi-head attention still has great potential. Due to the lack of computational resources, our experiment fixed the head number and hidden vector size. However, tuning these hyper parameters may have impact on the result. Tuning integration method and try to utilize larger datasets may give help to improving the performance." ], [ "To investigate how shared task and shared model can benefit, we split our dataset by query types, train our proposed model with different datasets and demonstrate their performance on different datasets. Firstly, we investigate the performance on model without two-stage training and named entity information.", "As indicated in Table TABREF30, The model trained by mixed data outperforms 2 of the 3 original tasks in EM-score with 81.55% for proximal resection margin and 86.85% for distal resection margin. The performance on tumor size declined by 1.57% score in EM-score and 3.14% score in F$_1$-score but they were still above 90%. 0.69% and 0.37% score improvement in EM-score was brought by shared model for proximal and distal resection margin prediction. Meanwhile F$_1$-score for those two tasks declined 3.11% and 0.77% score.", "Then we investigate the performance on model with two-stage training and named entity information. In this experiment, pre-training process only use the specific dataset not the mixed data. From Table TABREF31 we can observe that the performance on proximal and distal resection margin achieved the best performance on both EM-score and F$_1$-score. Compared with Table TABREF30, the best performance on proximal resection margin improved by 6.9% in EM-score and 7.94% in F$_1$-score. Meanwhile, the best performance on distal resection margin improved by 5.56% in EM-score and 6.32% in F$_1$-score. Other performances also usually improved a lot. This proves the usefulness of two-stage training and named entity information as well.", "Lastly, we fine tune the model for each task with a pre-trained parameter. Table TABREF32 summarizes the result. (Add some explanations for the Table TABREF32). Comparing Table TABREF32 with Table TABREF31, using mixed-data pre-trained parameters can significantly improve the model performance than task-specific data trained model. Except tumor size, the result was improved by 0.52% score in EM-score, 1.39% score in F$_1$-score for proximal resection margin and 2.6% score in EM-score, 2.96% score in F$_1$-score for distal resection margin. This proves mixed-data pre-trained parameters can lead to a great benefit for specific task. Meanwhile, the model performance on other tasks which are not trained in the final stage was also improved from around 0 to 60 or 70 percent. This proves that there is commonality between different tasks and our proposed QA-CTS task make this learnable. In conclusion, to achieve the best performance for a specific dataset, pre-training the model in multiple datasets and then fine tuning the model on the specific dataset is the best way." ], [ "In this paper, we present a question answering based clinical text structuring (QA-CTS) task, which unifies different clinical text structuring tasks and utilize different datasets. A novel model is also proposed to integrate named entity information into a pre-trained language model and adapt it to QA-CTS task. Initially, sequential results of named entity recognition on both paragraph and query texts are integrated together. Contextualized representation on both paragraph and query texts are transformed by a pre-trained language model. Then, the integrated named entity information and contextualized representation are integrated together and fed into a feed forward network for final prediction. Experimental results on real-world dataset demonstrate that our proposed model competes favorably with strong baseline models in all three specific tasks. The shared task and shared model introduced by QA-CTS task has also been proved to be useful for improving the performance on most of the task-specific datasets. In conclusion, the best way to achieve the best performance for a specific dataset is to pre-train the model in multiple datasets and then fine tune it on the specific dataset." ], [ "We would like to thank Ting Li and Xizhou Hong (Ruijin Hospital) who have helped us very much in data fetching and data cleansing. This work is supported by the National Key R&D Program of China for “Precision Medical Research\" (No. 2018YFC0910500)." ] ] }
{ "question": [ "What data is the language model pretrained on?", "What baselines is the proposed model compared against?", "How is the clinical text structuring task defined?", "What are the specific tasks being unified?", "Is all text in this dataset a question, or are there unrelated sentences in between questions?", "How many questions are in the dataset?", "What is the perWhat are the tasks evaluated?", "Are there privacy concerns with clinical data?", "How they introduce domain-specific features into pre-trained language model?", "How big is QA-CTS task dataset?", "How big is dataset of pathology reports collected from Ruijing Hospital?", "What are strong baseline models in specific tasks?" ], "question_id": [ "71a7153e12879defa186bfb6dbafe79c74265e10", "85d1831c28d3c19c84472589a252e28e9884500f", "1959e0ebc21fafdf1dd20c6ea054161ba7446f61", "77cf4379106463b6ebcb5eb8fa5bb25450fa5fb8", "06095a4dee77e9a570837b35fc38e77228664f91", "19c9cfbc4f29104200393e848b7b9be41913a7ac", "6743c1dd7764fc652cfe2ea29097ea09b5544bc3", "14323046220b2aea8f15fba86819cbccc389ed8b", "08a5f8d36298b57f6a4fcb4b6ae5796dc5d944a4", "975a4ac9773a4af551142c324b64a0858670d06e", "326e08a0f5753b90622902bd4a9c94849a24b773", "bd78483a746fda4805a7678286f82d9621bc45cf" ], "nlp_background": [ "infinity", "infinity", "infinity", "infinity", "five", "five", "five", "five", "zero", "zero", "zero", "zero" ], "topic_background": [ "research", "research", "research", "research", "familiar", "familiar", "familiar", "familiar", "unfamiliar", "unfamiliar", "unfamiliar", "unfamiliar" ], "paper_read": [ "no", "no", "no", "no", "no", "no", "no", "no", "no", "no", "no", "no" ], "search_query": [ "question answering", "question answering", "question answering", "question answering", "Question Answering", "Question Answering", "Question Answering", "Question Answering", "", "", "", "" ], "question_writer": [ "ecca0cede84b7af8a918852311d36346b07f0668", "ecca0cede84b7af8a918852311d36346b07f0668", "ecca0cede84b7af8a918852311d36346b07f0668", "ecca0cede84b7af8a918852311d36346b07f0668", "2a18a3656984d04249f100633e4c1003417a2255", "2a18a3656984d04249f100633e4c1003417a2255", "2a18a3656984d04249f100633e4c1003417a2255", "2a18a3656984d04249f100633e4c1003417a2255", "258ee4069f740c400c0049a2580945a1cc7f044c", "258ee4069f740c400c0049a2580945a1cc7f044c", "258ee4069f740c400c0049a2580945a1cc7f044c", "258ee4069f740c400c0049a2580945a1cc7f044c" ], "answers": [ { "answer": [ { "unanswerable": false, "extractive_spans": [ "Chinese general corpus" ], "yes_no": null, "free_form_answer": "", "evidence": [ "To implement deep neural network models, we utilize the Keras library BIBREF36 with TensorFlow BIBREF37 backend. Each model is run on a single NVIDIA GeForce GTX 1080 Ti GPU. The models are trained by Adam optimization algorithm BIBREF38 whose parameters are the same as the default settings except for learning rate set to $5\\times 10^{-5}$. Batch size is set to 3 or 4 due to the lack of graphical memory. We select BERT-base as the pre-trained language model in this paper. Due to the high cost of pre-training BERT language model, we directly adopt parameters pre-trained by Google in Chinese general corpus. The named entity recognition is applied on both pathology report texts and query texts." ], "highlighted_evidence": [ "Due to the high cost of pre-training BERT language model, we directly adopt parameters pre-trained by Google in Chinese general corpus. The named entity recognition is applied on both pathology report texts and query texts." ] }, { "unanswerable": true, "extractive_spans": [], "yes_no": null, "free_form_answer": "", "evidence": [], "highlighted_evidence": [] } ], "annotation_id": [ "0ab604dbe114dba174da645cc06a713e12a1fd9d", "1f1495d06d0abe86ee52124ec9f2f0b25a536147" ], "worker_id": [ "a0b403873302db7cada39008f04d01155ef68f4f", "258ee4069f740c400c0049a2580945a1cc7f044c" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "BERT-Base", "QANet" ], "yes_no": null, "free_form_answer": "", "evidence": [ "Experimental Studies ::: Comparison with State-of-the-art Methods", "Since BERT has already achieved the state-of-the-art performance of question-answering, in this section we compare our proposed model with state-of-the-art question answering models (i.e. QANet BIBREF39) and BERT-Base BIBREF26. As BERT has two versions: BERT-Base and BERT-Large, due to the lack of computational resource, we can only compare with BERT-Base model instead of BERT-Large. Prediction layer is attached at the end of the original BERT-Base model and we fine tune it on our dataset. In this section, the named entity integration method is chosen to pure concatenation (Concatenate the named entity information on pathology report text and query text first and then concatenate contextualized representation and concatenated named entity information). Comparative results are summarized in Table TABREF23." ], "highlighted_evidence": [ "Experimental Studies ::: Comparison with State-of-the-art Methods\nSince BERT has already achieved the state-of-the-art performance of question-answering, in this section we compare our proposed model with state-of-the-art question answering models (i.e. QANet BIBREF39) and BERT-Base BIBREF26. As BERT has two versions: BERT-Base and BERT-Large, due to the lack of computational resource, we can only compare with BERT-Base model instead of BERT-Large." ] }, { "unanswerable": false, "extractive_spans": [ "QANet BIBREF39", "BERT-Base BIBREF26" ], "yes_no": null, "free_form_answer": "", "evidence": [ "Since BERT has already achieved the state-of-the-art performance of question-answering, in this section we compare our proposed model with state-of-the-art question answering models (i.e. QANet BIBREF39) and BERT-Base BIBREF26. As BERT has two versions: BERT-Base and BERT-Large, due to the lack of computational resource, we can only compare with BERT-Base model instead of BERT-Large. Prediction layer is attached at the end of the original BERT-Base model and we fine tune it on our dataset. In this section, the named entity integration method is chosen to pure concatenation (Concatenate the named entity information on pathology report text and query text first and then concatenate contextualized representation and concatenated named entity information). Comparative results are summarized in Table TABREF23.", "FLOAT SELECTED: TABLE III COMPARATIVE RESULTS BETWEEN BERT AND OUR PROPOSED MODEL" ], "highlighted_evidence": [ "Since BERT has already achieved the state-of-the-art performance of question-answering, in this section we compare our proposed model with state-of-the-art question answering models (i.e. QANet BIBREF39) and BERT-Base BIBREF26. As BERT has two versions: BERT-Base and BERT-Large, due to the lack of computational resource, we can only compare with BERT-Base model instead of BERT-Large. ", "FLOAT SELECTED: TABLE III COMPARATIVE RESULTS BETWEEN BERT AND OUR PROPOSED MODEL" ] } ], "annotation_id": [ "0de2087bf0e46b14042de2a6e707bbf544a04556", "c14d9acff1d3e6f47901e7104a7f01a10a727050" ], "worker_id": [ "258ee4069f740c400c0049a2580945a1cc7f044c", "a0b403873302db7cada39008f04d01155ef68f4f" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "Clinical text structuring (CTS) is a critical task for fetching medical research data from electronic health records (EHRs), where structural patient medical data, such as whether the patient has specific symptoms, diseases, or what the tumor size is, how far from the tumor is cut at during the surgery, or what the specific laboratory test result is, are obtained.", "Unlike the traditional CTS task, our QA-CTS task aims to discover the most related text from original paragraph text. " ], "yes_no": null, "free_form_answer": "", "evidence": [ "FLOAT SELECTED: Fig. 1. An illustrative example of QA-CTS task.", "Clinical text structuring (CTS) is a critical task for fetching medical research data from electronic health records (EHRs), where structural patient medical data, such as whether the patient has specific symptoms, diseases, or what the tumor size is, how far from the tumor is cut at during the surgery, or what the specific laboratory test result is, are obtained. It is important to extract structured data from clinical text because bio-medical systems or bio-medical researches greatly rely on structured data but they cannot obtain them directly. In addition, clinical text often contains abundant healthcare information. CTS is able to provide large-scale extracted structured data for enormous down-stream clinical researches.", "To reduce the pipeline depth and break the barrier of non-uniform output formats, we present a question answering based clinical text structuring (QA-CTS) task (see Fig. FIGREF1). Unlike the traditional CTS task, our QA-CTS task aims to discover the most related text from original paragraph text. For some cases, it is already the final answer in deed (e.g., extracting sub-string). While for other cases, it needs several steps to obtain the final answer, such as entity names conversion and negative words recognition. Our presented QA-CTS task unifies the output format of the traditional CTS task and make the training data shareable, thus enriching the training data. The main contributions of this work can be summarized as follows." ], "highlighted_evidence": [ "FLOAT SELECTED: Fig. 1. An illustrative example of QA-CTS task.", "Clinical text structuring (CTS) is a critical task for fetching medical research data from electronic health records (EHRs), where structural patient medical data, such as whether the patient has specific symptoms, diseases, or what the tumor size is, how far from the tumor is cut at during the surgery, or what the specific laboratory test result is, are obtained. It is important to extract structured data from clinical text because bio-medical systems or bio-medical researches greatly rely on structured data but they cannot obtain them directly. In addition, clinical text often contains abundant healthcare information. CTS is able to provide large-scale extracted structured data for enormous down-stream clinical researches.", "Unlike the traditional CTS task, our QA-CTS task aims to discover the most related text from original paragraph text. For some cases, it is already the final answer in deed (e.g., extracting sub-string). While for other cases, it needs several steps to obtain the final answer, such as entity names conversion and negative words recognition. Our presented QA-CTS task unifies the output format of the traditional CTS task and make the training data shareable, thus enriching the training data. " ] }, { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "CTS is extracting structural data from medical research data (unstructured). Authors define QA-CTS task that aims to discover most related text from original text.", "evidence": [ "Clinical text structuring (CTS) is a critical task for fetching medical research data from electronic health records (EHRs), where structural patient medical data, such as whether the patient has specific symptoms, diseases, or what the tumor size is, how far from the tumor is cut at during the surgery, or what the specific laboratory test result is, are obtained. It is important to extract structured data from clinical text because bio-medical systems or bio-medical researches greatly rely on structured data but they cannot obtain them directly. In addition, clinical text often contains abundant healthcare information. CTS is able to provide large-scale extracted structured data for enormous down-stream clinical researches.", "However, end-to-end CTS is a very challenging task. Different CTS tasks often have non-uniform output formats, such as specific-class classifications (e.g. tumor stage), strings in the original text (e.g. result for a laboratory test) and inferred values from part of the original text (e.g. calculated tumor size). Researchers have to construct different models for it, which is already costly, and hence it calls for a lot of labeled data for each model. Moreover, labeling necessary amount of data for training neural network requires expensive labor cost. To handle it, researchers turn to some rule-based structuring methods which often have lower labor cost.", "To reduce the pipeline depth and break the barrier of non-uniform output formats, we present a question answering based clinical text structuring (QA-CTS) task (see Fig. FIGREF1). Unlike the traditional CTS task, our QA-CTS task aims to discover the most related text from original paragraph text. For some cases, it is already the final answer in deed (e.g., extracting sub-string). While for other cases, it needs several steps to obtain the final answer, such as entity names conversion and negative words recognition. Our presented QA-CTS task unifies the output format of the traditional CTS task and make the training data shareable, thus enriching the training data. The main contributions of this work can be summarized as follows." ], "highlighted_evidence": [ "Clinical text structuring (CTS) is a critical task for fetching medical research data from electronic health records (EHRs), where structural patient medical data, such as whether the patient has specific symptoms, diseases, or what the tumor size is, how far from the tumor is cut at during the surgery, or what the specific laboratory test result is, are obtained. It is important to extract structured data from clinical text because bio-medical systems or bio-medical researches greatly rely on structured data but they cannot obtain them directly.", "However, end-to-end CTS is a very challenging task. Different CTS tasks often have non-uniform output formats, such as specific-class classifications (e.g. tumor stage), strings in the original text (e.g. result for a laboratory test) and inferred values from part of the original text (e.g. calculated tumor size).", "To reduce the pipeline depth and break the barrier of non-uniform output formats, we present a question answering based clinical text structuring (QA-CTS) task (see Fig. FIGREF1). Unlike the traditional CTS task, our QA-CTS task aims to discover the most related text from original paragraph text." ] } ], "annotation_id": [ "6d56080358bb7f22dd764934ffcd6d4e93fef0b2", "da233cce57e642941da2446d3e053349c2ab1a15" ], "worker_id": [ "a0b403873302db7cada39008f04d01155ef68f4f", "258ee4069f740c400c0049a2580945a1cc7f044c" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ " three types of questions, namely tumor size, proximal resection margin and distal resection margin" ], "yes_no": null, "free_form_answer": "", "evidence": [ "To reduce the pipeline depth and break the barrier of non-uniform output formats, we present a question answering based clinical text structuring (QA-CTS) task (see Fig. FIGREF1). Unlike the traditional CTS task, our QA-CTS task aims to discover the most related text from original paragraph text. For some cases, it is already the final answer in deed (e.g., extracting sub-string). While for other cases, it needs several steps to obtain the final answer, such as entity names conversion and negative words recognition. Our presented QA-CTS task unifies the output format of the traditional CTS task and make the training data shareable, thus enriching the training data. The main contributions of this work can be summarized as follows.", "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs. All question-answer pairs are annotated and reviewed by four clinicians with three types of questions, namely tumor size, proximal resection margin and distal resection margin. These annotated instances have been partitioned into 1,899 training instances (12,412 sentences) and 815 test instances (5,421 sentences). Each instance has one or several sentences. Detailed statistics of different types of entities are listed in Table TABREF20.", "In this paper, we present a question answering based clinical text structuring (QA-CTS) task, which unifies different clinical text structuring tasks and utilize different datasets. A novel model is also proposed to integrate named entity information into a pre-trained language model and adapt it to QA-CTS task. Initially, sequential results of named entity recognition on both paragraph and query texts are integrated together. Contextualized representation on both paragraph and query texts are transformed by a pre-trained language model. Then, the integrated named entity information and contextualized representation are integrated together and fed into a feed forward network for final prediction. Experimental results on real-world dataset demonstrate that our proposed model competes favorably with strong baseline models in all three specific tasks. The shared task and shared model introduced by QA-CTS task has also been proved to be useful for improving the performance on most of the task-specific datasets. In conclusion, the best way to achieve the best performance for a specific dataset is to pre-train the model in multiple datasets and then fine tune it on the specific dataset." ], "highlighted_evidence": [ "Our presented QA-CTS task unifies the output format of the traditional CTS task and make the training data shareable, thus enriching the training data.", "All question-answer pairs are annotated and reviewed by four clinicians with three types of questions, namely tumor size, proximal resection margin and distal resection margin. ", "Experimental results on real-world dataset demonstrate that our proposed model competes favorably with strong baseline models in all three specific tasks." ] }, { "unanswerable": true, "extractive_spans": [], "yes_no": null, "free_form_answer": "", "evidence": [], "highlighted_evidence": [] } ], "annotation_id": [ "7138d812ea70084e7610e5a2422039da1404afd7", "b732d5561babcf37393ebf6cbb051d04b0b66bd5" ], "worker_id": [ "a0b403873302db7cada39008f04d01155ef68f4f", "258ee4069f740c400c0049a2580945a1cc7f044c" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "the dataset consists of pathology reports including sentences and questions and answers about tumor size and resection margins so it does include additional sentences ", "evidence": [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs. All question-answer pairs are annotated and reviewed by four clinicians with three types of questions, namely tumor size, proximal resection margin and distal resection margin. These annotated instances have been partitioned into 1,899 training instances (12,412 sentences) and 815 test instances (5,421 sentences). Each instance has one or several sentences. Detailed statistics of different types of entities are listed in Table TABREF20." ], "highlighted_evidence": [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs. All question-answer pairs are annotated and reviewed by four clinicians with three types of questions, namely tumor size, proximal resection margin and distal resection margin. " ] } ], "annotation_id": [ "1d4d4965fd44fefbfed0b3267ef5875572994b66" ], "worker_id": [ "fa716cd87ce6fd6905e2f23f09b262e90413167f" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "2,714 ", "evidence": [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs. All question-answer pairs are annotated and reviewed by four clinicians with three types of questions, namely tumor size, proximal resection margin and distal resection margin. These annotated instances have been partitioned into 1,899 training instances (12,412 sentences) and 815 test instances (5,421 sentences). Each instance has one or several sentences. Detailed statistics of different types of entities are listed in Table TABREF20." ], "highlighted_evidence": [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs." ] } ], "annotation_id": [ "229cc59d1545c9e8f47d43053465e2dfd1b763cc" ], "worker_id": [ "fa716cd87ce6fd6905e2f23f09b262e90413167f" ] }, { "answer": [ { "unanswerable": true, "extractive_spans": [], "yes_no": null, "free_form_answer": "", "evidence": [], "highlighted_evidence": [] } ], "annotation_id": [ "e2fe2a3438f28758724d992502a44615051eda90" ], "worker_id": [ "fa716cd87ce6fd6905e2f23f09b262e90413167f" ] }, { "answer": [ { "unanswerable": true, "extractive_spans": [], "yes_no": null, "free_form_answer": "", "evidence": [], "highlighted_evidence": [] } ], "annotation_id": [ "2a73264b743b6dd183c200f7dcd04aed4029f015" ], "worker_id": [ "fa716cd87ce6fd6905e2f23f09b262e90413167f" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "integrate clinical named entity information into pre-trained language model" ], "yes_no": null, "free_form_answer": "", "evidence": [ "We first present a question answering based clinical text structuring (QA-CTS) task, which unifies different specific tasks and make dataset shareable. We also propose an effective model to integrate clinical named entity information into pre-trained language model.", "In this section, we present an effective model for the question answering based clinical text structuring (QA-CTS). As shown in Fig. FIGREF8, paragraph text $X$ is first passed to a clinical named entity recognition (CNER) model BIBREF12 to capture named entity information and obtain one-hot CNER output tagging sequence for query text $I_{nq}$ and paragraph text $I_{nt}$ with BIEOS (Begin, Inside, End, Outside, Single) tag scheme. $I_{nq}$ and $I_{nt}$ are then integrated together into $I_n$. Meanwhile, the paragraph text $X$ and query text $Q$ are organized and passed to contextualized representation model which is pre-trained language model BERT BIBREF26 here to obtain the contextualized representation vector $V_s$ of both text and query. Afterwards, $V_s$ and $I_n$ are integrated together and fed into a feed forward network to calculate the start and end index of answer-related text. Here we define this calculation problem as a classification for each word to be the start or end word." ], "highlighted_evidence": [ "We also propose an effective model to integrate clinical named entity information into pre-trained language model.", "In this section, we present an effective model for the question answering based clinical text structuring (QA-CTS). As shown in Fig. FIGREF8, paragraph text $X$ is first passed to a clinical named entity recognition (CNER) model BIBREF12 to capture named entity information and obtain one-hot CNER output tagging sequence for query text $I_{nq}$ and paragraph text $I_{nt}$ with BIEOS (Begin, Inside, End, Outside, Single) tag scheme. $I_{nq}$ and $I_{nt}$ are then integrated together into $I_n$. Meanwhile, the paragraph text $X$ and query text $Q$ are organized and passed to contextualized representation model which is pre-trained language model BERT BIBREF26 here to obtain the contextualized representation vector $V_s$ of both text and query. Afterwards, $V_s$ and $I_n$ are integrated together and fed into a feed forward network to calculate the start and end index of answer-related text." ] } ], "annotation_id": [ "5f125408e657282669f90a1866d8227c0f94332e" ], "worker_id": [ "e70d8110563d53282f1a26e823d27e6f235772db" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "17,833 sentences, 826,987 characters and 2,714 question-answer pairs" ], "yes_no": null, "free_form_answer": "", "evidence": [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs. All question-answer pairs are annotated and reviewed by four clinicians with three types of questions, namely tumor size, proximal resection margin and distal resection margin. These annotated instances have been partitioned into 1,899 training instances (12,412 sentences) and 815 test instances (5,421 sentences). Each instance has one or several sentences. Detailed statistics of different types of entities are listed in Table TABREF20." ], "highlighted_evidence": [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs. " ] } ], "annotation_id": [ "24c7023a5221b509d34dd6703d6e0607b2777e78" ], "worker_id": [ "e70d8110563d53282f1a26e823d27e6f235772db" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "17,833 sentences, 826,987 characters and 2,714 question-answer pairs" ], "yes_no": null, "free_form_answer": "", "evidence": [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs. All question-answer pairs are annotated and reviewed by four clinicians with three types of questions, namely tumor size, proximal resection margin and distal resection margin. These annotated instances have been partitioned into 1,899 training instances (12,412 sentences) and 815 test instances (5,421 sentences). Each instance has one or several sentences. Detailed statistics of different types of entities are listed in Table TABREF20." ], "highlighted_evidence": [ "Our dataset is annotated based on Chinese pathology reports provided by the Department of Gastrointestinal Surgery, Ruijin Hospital. It contains 17,833 sentences, 826,987 characters and 2,714 question-answer pairs." ] } ], "annotation_id": [ "d046d9ea83c5ffe607465e2fbc8817131c11e037" ], "worker_id": [ "e70d8110563d53282f1a26e823d27e6f235772db" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "state-of-the-art question answering models (i.e. QANet BIBREF39) and BERT-Base BIBREF26" ], "yes_no": null, "free_form_answer": "", "evidence": [ "Since BERT has already achieved the state-of-the-art performance of question-answering, in this section we compare our proposed model with state-of-the-art question answering models (i.e. QANet BIBREF39) and BERT-Base BIBREF26. As BERT has two versions: BERT-Base and BERT-Large, due to the lack of computational resource, we can only compare with BERT-Base model instead of BERT-Large. Prediction layer is attached at the end of the original BERT-Base model and we fine tune it on our dataset. In this section, the named entity integration method is chosen to pure concatenation (Concatenate the named entity information on pathology report text and query text first and then concatenate contextualized representation and concatenated named entity information). Comparative results are summarized in Table TABREF23." ], "highlighted_evidence": [ "Since BERT has already achieved the state-of-the-art performance of question-answering, in this section we compare our proposed model with state-of-the-art question answering models (i.e. QANet BIBREF39) and BERT-Base BIBREF26. As BERT has two versions: BERT-Base and BERT-Large, due to the lack of computational resource, we can only compare with BERT-Base model instead of BERT-Large." ] } ], "annotation_id": [ "b3a3d6e707a67bab827053b40e446f30e416887f" ], "worker_id": [ "e70d8110563d53282f1a26e823d27e6f235772db" ] } ] }
{ "caption": [ "Fig. 1. An illustrative example of QA-CTS task.", "TABLE I AN ILLUSTRATIVE EXAMPLE OF NAMED ENTITY FEATURE TAGS", "Fig. 2. The architecture of our proposed model for QA-CTS task", "TABLE II STATISTICS OF DIFFERENT TYPES OF QUESTION ANSWER INSTANCES", "TABLE V COMPARATIVE RESULTS FOR DIFFERENT INTEGRATION METHOD OF OUR PROPOSED MODEL", "TABLE III COMPARATIVE RESULTS BETWEEN BERT AND OUR PROPOSED MODEL", "TABLE VI COMPARATIVE RESULTS FOR DATA INTEGRATION ANALYSIS (WITHOUT TWO-STAGE TRAINING AND NAMED ENTITY INFORMATION)", "TABLE VII COMPARATIVE RESULTS FOR DATA INTEGRATION ANALYSIS (WITH TWO-STAGE TRAINING AND NAMED ENTITY INFORMATION)", "TABLE VIII COMPARATIVE RESULTS FOR DATA INTEGRATION ANALYSIS (USING MIXED-DATA PRE-TRAINED PARAMETERS)" ], "file": [ "1-Figure1-1.png", "2-TableI-1.png", "3-Figure2-1.png", "4-TableII-1.png", "5-TableV-1.png", "5-TableIII-1.png", "6-TableVI-1.png", "6-TableVII-1.png", "6-TableVIII-1.png" ] }
1811.00942
Progress and Tradeoffs in Neural Language Models
In recent years, we have witnessed a dramatic shift towards techniques driven by neural networks for a variety of NLP tasks. Undoubtedly, neural language models (NLMs) have reduced perplexity by impressive amounts. This progress, however, comes at a substantial cost in performance, in terms of inference latency and energy consumption, which is particularly of concern in deployments on mobile devices. This paper, which examines the quality-performance tradeoff of various language modeling techniques, represents to our knowledge the first to make this observation. We compare state-of-the-art NLMs with"classic"Kneser-Ney (KN) LMs in terms of energy usage, latency, perplexity, and prediction accuracy using two standard benchmarks. On a Raspberry Pi, we find that orders of increase in latency and energy usage correspond to less change in perplexity, while the difference is much less pronounced on a desktop.
{ "section_name": [ "Introduction", "Background and Related Work", "Experimental Setup", "Hyperparameters and Training", "Infrastructure", "Results and Discussion", "Conclusion" ], "paragraphs": [ [ "Deep learning has unquestionably advanced the state of the art in many natural language processing tasks, from syntactic dependency parsing BIBREF0 to named-entity recognition BIBREF1 to machine translation BIBREF2 . The same certainly applies to language modeling, where recent advances in neural language models (NLMs) have led to dramatically better approaches as measured using standard metrics such as perplexity BIBREF3 , BIBREF4 .", "Specifically focused on language modeling, this paper examines an issue that to our knowledge has not been explored: advances in neural language models have come at a significant cost in terms of increased computational complexity. Computing the probability of a token sequence using non-neural techniques requires a number of phrase lookups and perhaps a few arithmetic operations, whereas model inference with NLMs require large matrix multiplications consuming perhaps millions of floating point operations (FLOPs). These performance tradeoffs are worth discussing.", "In truth, language models exist in a quality–performance tradeoff space. As model quality increases (e.g., lower perplexity), performance as measured in terms of energy consumption, query latency, etc. tends to decrease. For applications primarily running in the cloud—say, machine translation—practitioners often solely optimize for the lowest perplexity. This is because such applications are embarrassingly parallel and hence trivial to scale in a data center environment.", "There are, however, applications of NLMs that require less one-sided optimizations. On mobile devices such as smartphones and tablets, for example, NLMs may be integrated into software keyboards for next-word prediction, allowing much faster text entry. Popular Android apps that enthusiastically tout this technology include SwiftKey and Swype. The greater computational costs of NLMs lead to higher energy usage in model inference, translating into shorter battery life.", "In this paper, we examine the quality–performance tradeoff in the shift from non-neural to neural language models. In particular, we compare Kneser–Ney smoothing, widely accepted as the state of the art prior to NLMs, to the best NLMs today. The decrease in perplexity on standard datasets has been well documented BIBREF3 , but to our knowledge no one has examined the performances tradeoffs. With deployment on a mobile device in mind, we evaluate energy usage and inference latency on a Raspberry Pi (which shares the same ARM architecture as nearly all smartphones today). We find that a 2.5 $\\times $ reduction in perplexity on PTB comes at a staggering cost in terms of performance: inference with NLMs takes 49 $\\times $ longer and requires 32 $\\times $ more energy. Furthermore, we find that impressive reductions in perplexity translate into at best modest improvements in next-word prediction, which is arguable a better metric for evaluating software keyboards on a smartphone. The contribution of this paper is the first known elucidation of this quality–performance tradeoff. Note that we refrain from prescriptive recommendations: whether or not a tradeoff is worthwhile depends on the application. Nevertheless, NLP engineers should arguably keep these tradeoffs in mind when selecting a particular operating point." ], [ " BIBREF3 evaluate recent neural language models; however, their focus is not on the computational footprint of each model, but rather the perplexity. To further reduce perplexity, many neural language model extensions exist, such as continuous cache pointer BIBREF5 and mixture of softmaxes BIBREF6 . Since our focus is on comparing “core” neural and non-neural approaches, we disregard these extra optimizations techniques in all of our models.", "Other work focus on designing lightweight models for resource-efficient inference on mobile devices. BIBREF7 explore LSTMs BIBREF8 with binary weights for language modeling; BIBREF9 examine shallow feedforward neural networks for natural language processing.", "AWD-LSTM. BIBREF4 show that a simple three-layer LSTM, with proper regularization and optimization techniques, can achieve state of the art on various language modeling datasets, surpassing more complex models. Specifically, BIBREF4 apply randomized backpropagation through time, variational dropout, activation regularization, embedding dropout, and temporal activation regularization. A novel scheduler for optimization, non-monotonically triggered ASGD (NT-ASGD) is also introduced. BIBREF4 name their three-layer LSTM model trained with such tricks, “AWD-LSTM.”", "Quasi-Recurrent Neural Networks. Quasi-recurrent neural networks (QRNNs; BIBREF10 ) achieve current state of the art in word-level language modeling BIBREF11 . A quasi-recurrent layer comprises two separate parts: a convolution layer with three weights, and a recurrent pooling layer. Given an input $\\mathbf {X} \\in \\mathbb {R}^{k \\times n}$ , the convolution layer is $\n\\mathbf {Z} = \\tanh (\\mathbf {W}_z \\cdot \\mathbf {X})\\\\\n\\mathbf {F} = \\sigma (\\mathbf {W}_f \\cdot \\mathbf {X})\\\\\n\\mathbf {O} = \\sigma (\\mathbf {W}_o \\cdot \\mathbf {X})\n$ ", "where $\\sigma $ denotes the sigmoid function, $\\cdot $ represents masked convolution across time, and $\\mathbf {W}_{\\lbrace z, f, o\\rbrace } \\in \\mathbb {R}^{m \\times k \\times r}$ are convolution weights with $k$ input channels, $m$ output channels, and a window size of $r$ . In the recurrent pooling layer, the convolution outputs are combined sequentially: $\n\\mathbf {c}_t &= \\mathbf {f}_t \\odot \\mathbf {c}_{t-1} + (1 -\n\\mathbf {f}_t) \\odot \\mathbf {z}_t\\\\\n\\mathbf {h}_t &= \\mathbf {o}_t \\odot \\mathbf {c}_t\n$ ", "Multiple QRNN layers can be stacked for deeper hierarchical representation, with the output $\\mathbf {h}_{1:t}$ being fed as the input into the subsequent layer: In language modeling, a four-layer QRNN is a standard architecture BIBREF11 .", "Perplexity–Recall Scale. Word-level perplexity does not have a strictly monotonic relationship with recall-at- $k$ , the fraction of top $k$ predictions that contain the correct word. A given R@ $k$ imposes a weak minimum perplexity constraint—there are many free parameters that allow for large variability in the perplexity given a certain R@ $k$ . Consider the corpus, “choo choo train,” with an associated unigram model $P(\\text{``choo''}) = 0.1$ , $P(\\text{``train''}) = 0.9$ , resulting in an R@1 of $1/3$ and perplexity of $4.8$ . Clearly, R@1 $ =1/3$ for all $P(\\text{``choo''}) \\le 0.5$ ; thus, perplexity can drop as low as 2 without affecting recall." ], [ "We conducted our experiments on Penn Treebank (PTB; BIBREF12 ) and WikiText-103 (WT103; BIBREF13 ). Preprocessed by BIBREF14 , PTB contains 887K tokens for training, 70K for validation, and 78K for test, with a vocabulary size of 10,000. On the other hand, WT103 comprises 103 million tokens for training, 217K for validation, and 245K for test, spanning a vocabulary of 267K unique tokens.", "For the neural language model, we used a four-layer QRNN BIBREF10 , which achieves state-of-the-art results on a variety of datasets, such as WT103 BIBREF11 and PTB. To compare against more common LSTM architectures, we also evaluated AWD-LSTM BIBREF4 on PTB. For the non-neural approach, we used a standard five-gram model with modified Kneser-Ney smoothing BIBREF15 , as explored in BIBREF16 on PTB. We denote the QRNN models for PTB and WT103 as ptb-qrnn and wt103-qrnn, respectively.", "For each model, we examined word-level perplexity, R@3 in next-word prediction, latency (ms/q), and energy usage (mJ/q). To explore the perplexity–recall relationship, we collected individual perplexity and recall statistics for each sentence in the test set." ], [ "The QRNN models followed the exact training procedure and architecture delineated in the official codebase from BIBREF11 . For ptb-qrnn, we trained the model for 550 epochs using NT-ASGD BIBREF4 , then finetuned for 300 epochs using ASGD BIBREF17 , all with a learning rate of 30 throughout. For wt103-qrnn, we followed BIBREF11 and trained the QRNN for 14 epochs, using the Adam optimizer with a learning rate of $10^{-3}$ . We also applied regularization techniques from BIBREF4 ; all the specific hyperparameters are the same as those in the repository. Our model architecture consists of 400-dimensional tied embedding weights BIBREF18 and four QRNN layers, with 1550 hidden units per layer on PTB and 2500 per layer on WT103. Both QRNN models have window sizes of $r=2$ for the first layer and $r=1$ for the rest.", "For the KN-5 model, we trained an off-the-shelf five-gram model using the popular SRILM toolkit BIBREF19 . We did not specify any special hyperparameters." ], [ "We trained the QRNNs with PyTorch (0.4.0; commit 1807bac) on a Titan V GPU. To evaluate the models under a resource-constrained environment, we deployed them on a Raspberry Pi 3 (Model B) running Raspbian Stretch (4.9.41-v7+). The Raspberry Pi (RPi) is not only a standard platform, but also a close surrogate to mobile phones, using the same Cortex-A7 in many phones. We then transferred the trained models to the RPi, using the same frameworks for evaluation. We plugged the RPi into a Watts Up Pro meter, a power meter that can be read programatically over USB at a frequency of 1 Hz. For the QRNNs, we used the first 350 words of the test set, and averaged the ms/query and mJ/query. For KN-5, we used the entire test set for evaluation, since the latency was much lower. To adjust for the base power load, we subtracted idle power draw from energy usage.", "For a different perspective, we further evaluated all the models under a desktop environment, using an i7-4790k CPU and Titan V GPU. Because the base power load for powering a desktop is much higher than running neural language models, we collected only latency statistics. We used the entire test set, since the QRNN runs quickly.", "In addition to energy and latency, another consideration for the NLP developer selecting an operating point is the cost of underlying hardware. For our setup, the RPi costs $35 USD, the CPU costs $350 USD, and the GPU costs $3000 USD." ], [ "To demonstrate the effectiveness of the QRNN models, we present the results of past and current state-of-the-art neural language models in Table 1 ; we report the Skip- and AWD-LSTM results as seen in the original papers, while we report our QRNN results. Skip LSTM denotes the four-layer Skip LSTM in BIBREF3 . BIBREF20 focus on Hebbian softmax, a model extension technique—Rae-LSTM refers to their base LSTM model without any extensions. In our results, KN-5 refers to the traditional five-gram model with modified Kneser-Ney smoothing, and AWD is shorthand for AWD-LSTM.", "Perplexity–recall scale. In Figure 1 , using KN-5 as the model, we plot the log perplexity (cross entropy) and R@3 error ( $1 - \\text{R@3}$ ) for every sentence in PTB and WT103. The horizontal clusters arise from multiple perplexity points representing the same R@3 value, as explained in Section \"Infrastructure\" . We also observe that the perplexity–recall scale is non-linear—instead, log perplexity appears to have a moderate linear relationship with R@3 error on PTB ( $r=0.85$ ), and an even stronger relationship on WT103 ( $r=0.94$ ). This is partially explained by WT103 having much longer sentences, and thus less noisy statistics.", "From Figure 1 , we find that QRNN models yield strongly linear log perplexity–recall plots as well, where $r=0.88$ and $r=0.93$ for PTB and WT103, respectively. Note that, due to the improved model quality over KN-5, the point clouds are shifted downward compared to Figure 1 . We conclude that log perplexity, or cross entropy, provides a more human-understandable indicator of R@3 than perplexity does. Overall, these findings agree with those from BIBREF21 , which explores the log perplexity–word error rate scale in language modeling for speech recognition.", "Quality–performance tradeoff. In Table 2 , from left to right, we report perplexity results on the validation and test sets, R@3 on test, and finally per-query latency and energy usage. On the RPi, KN-5 is both fast and power-efficient to run, using only about 7 ms/query and 6 mJ/query for PTB (Table 2 , row 1), and 264 ms/q and 229 mJ/q on WT103 (row 5). Taking 220 ms/query and consuming 300 mJ/query, AWD-LSTM and ptb-qrnn are still viable for mobile phones: The modern smartphone holds upwards of 10,000 joules BIBREF22 , and the latency is within usability standards BIBREF23 . Nevertheless, the models are still 49 $\\times $ slower and 32 $\\times $ more power-hungry than KN-5. The wt103-qrnn model is completely unusable on phones, taking over 1.2 seconds per next-word prediction. Neural models achieve perplexity drops of 60–80% and R@3 increases of 22–34%, but these improvements come at a much higher cost in latency and energy usage.", "In Table 2 (last two columns), the desktop yields very different results: the neural models on PTB (rows 2–3) are 9 $\\times $ slower than KN-5, but the absolute latency is only 8 ms/q, which is still much faster than what humans perceive as instantaneous BIBREF23 . If a high-end commodity GPU is available, then the models are only twice as slow as KN-5 is. From row 5, even better results are noted with wt103-qrnn: On the CPU, the QRNN is only 60% slower than KN-5 is, while the model is faster by 11 $\\times $ on a GPU. These results suggest that, if only latency is considered under a commodity desktop environment, the QRNN model is humanly indistinguishable from the KN-5 model, even without using GPU acceleration." ], [ "In the present work, we describe and examine the tradeoff space between quality and performance for the task of language modeling. Specifically, we explore the quality–performance tradeoffs between KN-5, a non-neural approach, and AWD-LSTM and QRNN, two neural language models. We find that with decreased perplexity comes vastly increased computational requirements: In one of the NLMs, a perplexity reduction by 2.5 $\\times $ results in a 49 $\\times $ rise in latency and 32 $\\times $ increase in energy usage, when compared to KN-5." ] ] }
{ "question": [ "What aspects have been compared between various language models?", "what classic language models are mentioned in the paper?", "What is a commonly used evaluation metric for language models?" ], "question_id": [ "dd155f01f6f4a14f9d25afc97504aefdc6d29c13", "a9d530d68fb45b52d9bad9da2cd139db5a4b2f7c", "e07df8f613dbd567a35318cd6f6f4cb959f5c82d" ], "nlp_background": [ "two", "two", "two" ], "topic_background": [ "familiar", "familiar", "familiar" ], "paper_read": [ "no", "no", "no" ], "search_query": [ "", "", "" ], "question_writer": [ "50d8b4a941c26b89482c94ab324b5a274f9ced66", "50d8b4a941c26b89482c94ab324b5a274f9ced66", "50d8b4a941c26b89482c94ab324b5a274f9ced66" ], "answers": [ { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "Quality measures using perplexity and recall, and performance measured using latency and energy usage. ", "evidence": [ "For each model, we examined word-level perplexity, R@3 in next-word prediction, latency (ms/q), and energy usage (mJ/q). To explore the perplexity–recall relationship, we collected individual perplexity and recall statistics for each sentence in the test set." ], "highlighted_evidence": [ "For each model, we examined word-level perplexity, R@3 in next-word prediction, latency (ms/q), and energy usage (mJ/q). To explore the perplexity–recall relationship, we collected individual perplexity and recall statistics for each sentence in the test set." ] } ], "annotation_id": [ "c17796e0bd3bfcc64d5a8e844d23d8d39274af6b" ], "worker_id": [ "1ba1b5b562aef9cd264cace5b7bdd46a7c065c0a" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "Kneser–Ney smoothing" ], "yes_no": null, "free_form_answer": "", "evidence": [ "In this paper, we examine the quality–performance tradeoff in the shift from non-neural to neural language models. In particular, we compare Kneser–Ney smoothing, widely accepted as the state of the art prior to NLMs, to the best NLMs today. The decrease in perplexity on standard datasets has been well documented BIBREF3 , but to our knowledge no one has examined the performances tradeoffs. With deployment on a mobile device in mind, we evaluate energy usage and inference latency on a Raspberry Pi (which shares the same ARM architecture as nearly all smartphones today). We find that a 2.5 $\\times $ reduction in perplexity on PTB comes at a staggering cost in terms of performance: inference with NLMs takes 49 $\\times $ longer and requires 32 $\\times $ more energy. Furthermore, we find that impressive reductions in perplexity translate into at best modest improvements in next-word prediction, which is arguable a better metric for evaluating software keyboards on a smartphone. The contribution of this paper is the first known elucidation of this quality–performance tradeoff. Note that we refrain from prescriptive recommendations: whether or not a tradeoff is worthwhile depends on the application. Nevertheless, NLP engineers should arguably keep these tradeoffs in mind when selecting a particular operating point." ], "highlighted_evidence": [ "Kneser–Ney smoothing", "In particular, we compare Kneser–Ney smoothing, widely accepted as the state of the art prior to NLMs, to the best NLMs today." ] } ], "annotation_id": [ "715840b32a89c33e0a1de1ab913664eb9694bd34" ], "worker_id": [ "1ba1b5b562aef9cd264cace5b7bdd46a7c065c0a" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "perplexity" ], "yes_no": null, "free_form_answer": "", "evidence": [ "Deep learning has unquestionably advanced the state of the art in many natural language processing tasks, from syntactic dependency parsing BIBREF0 to named-entity recognition BIBREF1 to machine translation BIBREF2 . The same certainly applies to language modeling, where recent advances in neural language models (NLMs) have led to dramatically better approaches as measured using standard metrics such as perplexity BIBREF3 , BIBREF4 ." ], "highlighted_evidence": [ "Deep learning has unquestionably advanced the state of the art in many natural language processing tasks, from syntactic dependency parsing BIBREF0 to named-entity recognition BIBREF1 to machine translation BIBREF2 . The same certainly applies to language modeling, where recent advances in neural language models (NLMs) have led to dramatically better approaches as measured using standard metrics such as perplexity BIBREF3 , BIBREF4 ." ] }, { "unanswerable": false, "extractive_spans": [ "perplexity" ], "yes_no": null, "free_form_answer": "", "evidence": [ "Deep learning has unquestionably advanced the state of the art in many natural language processing tasks, from syntactic dependency parsing BIBREF0 to named-entity recognition BIBREF1 to machine translation BIBREF2 . The same certainly applies to language modeling, where recent advances in neural language models (NLMs) have led to dramatically better approaches as measured using standard metrics such as perplexity BIBREF3 , BIBREF4 ." ], "highlighted_evidence": [ "recent advances in neural language models (NLMs) have led to dramatically better approaches as measured using standard metrics such as perplexity BIBREF3 , BIBREF4 ." ] } ], "annotation_id": [ "062dcccfdfb5af1c6ee886885703f9437d91a9dc", "1cc952fc047d0bb1a961c3ce65bada2e983150d1" ], "worker_id": [ "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4", "1ba1b5b562aef9cd264cace5b7bdd46a7c065c0a" ] } ] }
{ "caption": [ "Table 1: Comparison of neural language models on Penn Treebank and WikiText-103.", "Figure 1: Log perplexity–recall error with KN-5.", "Figure 2: Log perplexity–recall error with QRNN.", "Table 2: Language modeling results on performance and model quality." ], "file": [ "3-Table1-1.png", "4-Figure1-1.png", "4-Figure2-1.png", "4-Table2-1.png" ] }
1907.05664
Saliency Maps Generation for Automatic Text Summarization
Saliency map generation techniques are at the forefront of explainable AI literature for a broad range of machine learning applications. Our goal is to question the limits of these approaches on more complex tasks. In this paper we apply Layer-Wise Relevance Propagation (LRP) to a sequence-to-sequence attention model trained on a text summarization dataset. We obtain unexpected saliency maps and discuss the rightfulness of these"explanations". We argue that we need a quantitative way of testing the counterfactual case to judge the truthfulness of the saliency maps. We suggest a protocol to check the validity of the importance attributed to the input and show that the saliency maps obtained sometimes capture the real use of the input features by the network, and sometimes do not. We use this example to discuss how careful we need to be when accepting them as explanation.
{ "section_name": [ "Introduction", "The Task and the Model", "Dataset and Training Task", "The Model", "Obtained Summaries", "Layer-Wise Relevance Propagation", "Mathematical Description", "Generation of the Saliency Maps", "Experimental results", "First Observations", "Validating the Attributions", "Conclusion" ], "paragraphs": [ [ "Ever since the LIME algorithm BIBREF0 , \"explanation\" techniques focusing on finding the importance of input features in regard of a specific prediction have soared and we now have many ways of finding saliency maps (also called heat-maps because of the way we like to visualize them). We are interested in this paper by the use of such a technique in an extreme task that highlights questions about the validity and evaluation of the approach. We would like to first set the vocabulary we will use. We agree that saliency maps are not explanations in themselves and that they are more similar to attribution, which is only one part of the human explanation process BIBREF1 . We will prefer to call this importance mapping of the input an attribution rather than an explanation. We will talk about the importance of the input relevance score in regard to the model's computation and not make allusion to any human understanding of the model as a result.", "There exist multiple ways to generate saliency maps over the input for non-linear classifiers BIBREF2 , BIBREF3 , BIBREF4 . We refer the reader to BIBREF5 for a survey of explainable AI in general. We use in this paper Layer-Wise Relevance Propagation (LRP) BIBREF2 which aims at redistributing the value of the classifying function on the input to obtain the importance attribution. It was first created to “explain\" the classification of neural networks on image recognition tasks. It was later successfully applied to text using convolutional neural networks (CNN) BIBREF6 and then Long-Short Term Memory (LSTM) networks for sentiment analysis BIBREF7 .", "Our goal in this paper is to test the limits of the use of such a technique for more complex tasks, where the notion of input importance might not be as simple as in topic classification or sentiment analysis. We changed from a classification task to a generative task and chose a more complex one than text translation (in which we can easily find a word to word correspondence/importance between input and output). We chose text summarization. We consider abstractive and informative text summarization, meaning that we write a summary “in our own words\" and retain the important information of the original text. We refer the reader to BIBREF8 for more details on the task and the different variants that exist. Since the success of deep sequence-to-sequence models for text translation BIBREF9 , the same approaches have been applied to text summarization tasks BIBREF10 , BIBREF11 , BIBREF12 which use architectures on which we can apply LRP.", "We obtain one saliency map for each word in the generated summaries, supposed to represent the use of the input features for each element of the output sequence. We observe that all the saliency maps for a text are nearly identical and decorrelated with the attention distribution. We propose a way to check their validity by creating what could be seen as a counterfactual experiment from a synthesis of the saliency maps, using the same technique as in Arras et al. Arras2017. We show that in some but not all cases they help identify the important input features and that we need to rigorously check importance attributions before trusting them, regardless of whether or not the mapping “makes sense\" to us. We finally argue that in the process of identifying the important input features, verifying the saliency maps is as important as the generation step, if not more." ], [ "We present in this section the baseline model from See et al. See2017 trained on the CNN/Daily Mail dataset. We reproduce the results from See et al. See2017 to then apply LRP on it." ], [ "The CNN/Daily mail dataset BIBREF12 is a text summarization dataset adapted from the Deepmind question-answering dataset BIBREF13 . It contains around three hundred thousand news articles coupled with summaries of about three sentences. These summaries are in fact “highlights\" of the articles provided by the media themselves. Articles have an average length of 780 words and the summaries of 50 words. We had 287 000 training pairs and 11 500 test pairs. Similarly to See et al. See2017, we limit during training and prediction the input text to 400 words and generate summaries of 200 words. We pad the shorter texts using an UNKNOWN token and truncate the longer texts. We embed the texts and summaries using a vocabulary of size 50 000, thus recreating the same parameters as See et al. See2017." ], [ "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." ], [ "We train the 21 350 992 parameters of the network for about 60 epochs until we achieve results that are qualitatively equivalent to the results of See et al. See2017. We obtain summaries that are broadly relevant to the text but do not match the target summaries very well. We observe the same problems such as wrong reproduction of factual details, replacing rare words with more common alternatives or repeating non-sense after the third sentence. We can see in Figure 1 an example of summary obtained compared to the target one.", "The “summaries\" we generate are far from being valid summaries of the information in the texts but are sufficient to look at the attribution that LRP will give us. They pick up the general subject of the original text." ], [ "We present in this section the Layer-Wise Relevance Propagation (LRP) BIBREF2 technique that we used to attribute importance to the input features, together with how we adapted it to our model and how we generated the saliency maps. LRP redistributes the output of the model from the output layer to the input by transmitting information backwards through the layers. We call this propagated backwards importance the relevance. LRP has the particularity to attribute negative and positive relevance: a positive relevance is supposed to represent evidence that led to the classifier's result while negative relevance represents evidence that participated negatively in the prediction." ], [ "We initialize the relevance of the output layer to the value of the predicted class before softmax and we then describe locally the propagation backwards of the relevance from layer to layer. For normal neural network layers we use the form of LRP with epsilon stabilizer BIBREF2 . We write down $R_{i\\leftarrow j}^{(l, l+1)}$ the relevance received by the neuron $i$ of layer $l$ from the neuron $j$ of layer $l+1$ : ", "$$\\begin{split}\n\nR_{i\\leftarrow j}^{(l, l+1)} &= \\dfrac{w_{i\\rightarrow j}^{l,l+1}\\textbf {z}^l_i + \\dfrac{\\epsilon \\textrm { sign}(\\textbf {z}^{l+1}_j) + \\textbf {b}^{l+1}_j}{D_l}}{\\textbf {z}^{l+1}_j + \\epsilon * \\textrm { sign}(\\textbf {z}^{l+1}_j)} * R_j^{l+1} \\\\\n\\end{split}$$ (Eq. 7) ", "where $w_{i\\rightarrow j}^{l,l+1}$ is the network's weight parameter set during training, $\\textbf {b}^{l+1}_j$ is the bias for neuron $j$ of layer $l+1$ , $\\textbf {z}^{l}_i$ is the activation of neuron $i$ on layer $l$ , $\\epsilon $ is the stabilizing term set to 0.00001 and $D_l$ is the dimension of the $l$ -th layer.", "The relevance of a neuron is then computed as the sum of the relevance he received from the above layer(s).", "For LSTM cells we use the method from Arras et al.Arras2017 to solve the problem posed by the element-wise multiplications of vectors. Arras et al. noted that when such computation happened inside an LSTM cell, it always involved a “gate\" vector and another vector containing information. The gate vector containing only value between 0 and 1 is essentially filtering the second vector to allow the passing of “relevant\" information. Considering this, when we propagate relevance through an element-wise multiplication operation, we give all the upper-layer's relevance to the “information\" vector and none to the “gate\" vector." ], [ "We use the same method to transmit relevance through the attention mechanism back to the encoder because Bahdanau's attention BIBREF9 uses element-wise multiplications as well. We depict in Figure 2 the transmission end-to-end from the output layer to the input through the decoder, attention mechanism and then the bidirectional encoder. We then sum up the relevance on the word embedding to get the token's relevance as Arras et al. Arras2017.", "The way we generate saliency maps differs a bit from the usual context in which LRP is used as we essentially don't have one classification, but 200 (one for each word in the summary). We generate a relevance attribution for the 50 first words of the generated summary as after this point they often repeat themselves.", "This means that for each text we obtain 50 different saliency maps, each one supposed to represent the relevance of the input for a specific generated word in the summary." ], [ "In this section, we present our results from extracting attributions from the sequence-to-sequence model trained for abstractive text summarization. We first have to discuss the difference between the 50 different saliency maps we obtain and then we propose a protocol to validate the mappings." ], [ "The first observation that is made is that for one text, the 50 saliency maps are almost identical. Indeed each mapping highlights mainly the same input words with only slight variations of importance. We can see in Figure 3 an example of two nearly identical attributions for two distant and unrelated words of the summary. The saliency map generated using LRP is also uncorrelated with the attention distribution that participated in the generation of the output word. The attention distribution changes drastically between the words in the generated summary while not impacting significantly the attribution over the input text. We deleted in an experiment the relevance propagated through the attention mechanism to the encoder and didn't observe much changes in the saliency map.", "It can be seen as evidence that using the attention distribution as an “explanation\" of the prediction can be misleading. It is not the only information received by the decoder and the importance it “allocates\" to this attention state might be very low. What seems to happen in this application is that most of the information used is transmitted from the encoder to the decoder and the attention mechanism at each decoding step just changes marginally how it is used. Quantifying the difference between attention distribution and saliency map across multiple tasks is a possible future work.", "The second observation we can make is that the saliency map doesn't seem to highlight the right things in the input for the summary it generates. The saliency maps on Figure 3 correspond to the summary from Figure 1 , and we don't see the word “video\" highlighted in the input text, which seems to be important for the output.", "This allows us to question how good the saliency maps are in the sense that we question how well they actually represent the network's use of the input features. We will call that truthfulness of the attribution in regard to the computation, meaning that an attribution is truthful in regard to the computation if it actually highlights the important input features that the network attended to during prediction. We proceed to measure the truthfulness of the attributions by validating them quantitatively." ], [ "We propose to validate the saliency maps in a similar way as Arras et al. Arras2017 by incrementally deleting “important\" words from the input text and observe the change in the resulting generated summaries.", "We first define what “important\" (and “unimportant\") input words mean across the 50 saliency maps per texts. Relevance transmitted by LRP being positive or negative, we average the absolute value of the relevance across the saliency maps to obtain one ranking of the most “relevant\" words. The idea is that input words with negative relevance have an impact on the resulting generated word, even if it is not participating positively, while a word with a relevance close to zero should not be important at all. We did however also try with different methods, like averaging the raw relevance or averaging a scaled absolute value where negative relevance is scaled down by a constant factor. The absolute value average seemed to deliver the best results.", "We delete incrementally the important words (words with the highest average) in the input and compared it to the control experiment that consists of deleting the least important word and compare the degradation of the resulting summaries. We obtain mitigated results: for some texts, we observe a quick degradation when deleting important words which are not observed when deleting unimportant words (see Figure 4 ), but for other test examples we don't observe a significant difference between the two settings (see Figure 5 ).", "One might argue that the second summary in Figure 5 is better than the first one as it makes better sentences but as the model generates inaccurate summaries, we do not wish to make such a statement.", "This however allows us to say that the attribution generated for the text at the origin of the summaries in Figure 4 are truthful in regard to the network's computation and we may use it for further studies of the example, whereas for the text at the origin of Figure 5 we shouldn't draw any further conclusions from the attribution generated.", "One interesting point is that one saliency map didn't look “better\" than the other, meaning that there is no apparent way of determining their truthfulness in regard of the computation without doing a quantitative validation. This brings us to believe that even in simpler tasks, the saliency maps might make sense to us (for example highlighting the animal in an image classification task), without actually representing what the network really attended too, or in what way.", "We defined without saying it the counterfactual case in our experiment: “Would the important words in the input be deleted, we would have a different summary\". Such counterfactuals are however more difficult to define for image classification for example, where it could be applying a mask over an image, or just filtering a colour or a pattern. We believe that defining a counterfactual and testing it allows us to measure and evaluate the truthfulness of the attributions and thus weight how much we can trust them." ], [ "In this work, we have implemented and applied LRP to a sequence-to-sequence model trained on a more complex task than usual: text summarization. We used previous work to solve the difficulties posed by LRP in LSTM cells and adapted the same technique for Bahdanau et al. Bahdanau2014 attention mechanism.", "We observed a peculiar behaviour of the saliency maps for the words in the output summary: they are almost all identical and seem uncorrelated with the attention distribution. We then proceeded to validate our attributions by averaging the absolute value of the relevance across the saliency maps. We obtain a ranking of the word from the most important to the least important and proceeded to delete one or another.", "We showed that in some cases the saliency maps are truthful to the network's computation, meaning that they do highlight the input features that the network focused on. But we also showed that in some cases the saliency maps seem to not capture the important input features. This brought us to discuss the fact that these attributions are not sufficient by themselves, and that we need to define the counter-factual case and test it to measure how truthful the saliency maps are.", "Future work would look into the saliency maps generated by applying LRP to pointer-generator networks and compare to our current results as well as mathematically justifying the average that we did when validating our saliency maps. Some additional work is also needed on the validation of the saliency maps with counterfactual tests. The exploitation and evaluation of saliency map are a very important step and should not be overlooked." ] ] }
{ "question": [ "Which baselines did they compare?", "How many attention layers are there in their model?", "Is the explanation from saliency map correct?" ], "question_id": [ "6e2ad9ad88cceabb6977222f5e090ece36aa84ea", "aacb0b97aed6fc6a8b471b8c2e5c4ddb60988bf5", "710c1f8d4c137c8dad9972f5ceacdbf8004db208" ], "nlp_background": [ "five", "five", "five" ], "topic_background": [ "familiar", "familiar", "familiar" ], "paper_read": [ "no", "no", "no" ], "search_query": [ "saliency", "saliency", "saliency" ], "question_writer": [ "50d8b4a941c26b89482c94ab324b5a274f9ced66", "50d8b4a941c26b89482c94ab324b5a274f9ced66", "50d8b4a941c26b89482c94ab324b5a274f9ced66" ], "answers": [ { "answer": [ { "unanswerable": false, "extractive_spans": [ "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." ], "yes_no": null, "free_form_answer": "", "evidence": [ "We present in this section the baseline model from See et al. See2017 trained on the CNN/Daily Mail dataset. We reproduce the results from See et al. See2017 to then apply LRP on it.", "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." ], "highlighted_evidence": [ "We present in this section the baseline model from See et al. See2017 trained on the CNN/Daily Mail dataset.", "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." ] }, { "unanswerable": false, "extractive_spans": [ "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." ], "yes_no": null, "free_form_answer": "", "evidence": [ "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." ], "highlighted_evidence": [ "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." ] } ], "annotation_id": [ "0850b7c0555801d057062480de6bb88adb81cae3", "93216bca45711b73083372495d9a2667736fbac9" ], "worker_id": [ "c1fbdd7a261021041f75fbe00a55b4c386ebbbb4", "101dbdd2108b3e676061cb693826f0959b47891b" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "one", "evidence": [ "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." ], "highlighted_evidence": [ "The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. " ] } ], "annotation_id": [ "e0ca6b95c1c051723007955ce6804bd29f325379" ], "worker_id": [ "101dbdd2108b3e676061cb693826f0959b47891b" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": false, "free_form_answer": "", "evidence": [ "We showed that in some cases the saliency maps are truthful to the network's computation, meaning that they do highlight the input features that the network focused on. But we also showed that in some cases the saliency maps seem to not capture the important input features. This brought us to discuss the fact that these attributions are not sufficient by themselves, and that we need to define the counter-factual case and test it to measure how truthful the saliency maps are.", "The second observation we can make is that the saliency map doesn't seem to highlight the right things in the input for the summary it generates. The saliency maps on Figure 3 correspond to the summary from Figure 1 , and we don't see the word “video\" highlighted in the input text, which seems to be important for the output." ], "highlighted_evidence": [ "But we also showed that in some cases the saliency maps seem to not capture the important input features. ", "The second observation we can make is that the saliency map doesn't seem to highlight the right things in the input for the summary it generates" ] } ], "annotation_id": [ "79e54a7b9ba9cde5813c3434e64a02d722f13b23" ], "worker_id": [ "101dbdd2108b3e676061cb693826f0959b47891b" ] } ] }
{ "caption": [ "Figure 2: Representation of the propagation of the relevance from the output to the input. It passes through the decoder and attention mechanism for each previous decoding time-step, then is passed onto the encoder which takes into account the relevance transiting in both direction due to the bidirectional nature of the encoding LSTM cell.", "Figure 3: Left : Saliency map over the truncated input text for the second generated word “the”. Right : Saliency map over the truncated input text for the 25th generated word “investigation”. We see that the difference between the mappings is marginal.", "Figure 4: Summary from Figure 1 generated after deleting important and unimportant words from the input text. We observe a significant difference in summary degradation between the two experiments, where the decoder just repeats the UNKNOWN token over and over." ], "file": [ "3-Figure2-1.png", "3-Figure3-1.png", "4-Figure4-1.png" ] }
1910.14497
Probabilistic Bias Mitigation in Word Embeddings
It has been shown that word embeddings derived from large corpora tend to incorporate biases present in their training data. Various methods for mitigating these biases have been proposed, but recent work has demonstrated that these methods hide but fail to truly remove the biases, which can still be observed in word nearest-neighbor statistics. In this work we propose a probabilistic view of word embedding bias. We leverage this framework to present a novel method for mitigating bias which relies on probabilistic observations to yield a more robust bias mitigation algorithm. We demonstrate that this method effectively reduces bias according to three separate measures of bias while maintaining embedding quality across various popular benchmark semantic tasks
{ "section_name": [ "Introduction", "Background ::: Geometric Bias Mitigation", "Background ::: Geometric Bias Mitigation ::: WEAT", "Background ::: Geometric Bias Mitigation ::: RIPA", "Background ::: Geometric Bias Mitigation ::: Neighborhood Metric", "A Probabilistic Framework for Bias Mitigation", "A Probabilistic Framework for Bias Mitigation ::: Probabilistic Bias Mitigation", "A Probabilistic Framework for Bias Mitigation ::: Nearest Neighbor Bias Mitigation", "Experiments", "Discussion", "Discussion ::: Acknowledgements", "Experiment Notes", "Professions", "WEAT Word Sets" ], "paragraphs": [ [ "Word embeddings, or vector representations of words, are an important component of Natural Language Processing (NLP) models and necessary for many downstream tasks. However, word embeddings, including embeddings commonly deployed for public use, have been shown to exhibit unwanted societal stereotypes and biases, raising concerns about disparate impact on axes of gender, race, ethnicity, and religion BIBREF0, BIBREF1. The impact of this bias has manifested in a range of downstream tasks, ranging from autocomplete suggestions BIBREF2 to advertisement delivery BIBREF3, increasing the likelihood of amplifying harmful biases through the use of these models.", "The most well-established method thus far for mitigating bias relies on projecting target words onto a bias subspace (such as a gender subspace) and subtracting out the difference between the resulting distances BIBREF0. On the other hand, the most popular metric for measuring bias is the WEAT statistic BIBREF1, which compares the cosine similarities between groups of words. However, WEAT has been recently shown to overestimate bias as a result of implicitly relying on similar frequencies for the target words BIBREF4, and BIBREF5 demonstrated that evidence of bias can still be recovered after geometric bias mitigation by examining the neighborhood of a target word among socially-biased words.", "In response to this, we propose an alternative framework for bias mitigation in word embeddings that approaches this problem from a probabilistic perspective. The motivation for this approach is two-fold. First, most popular word embedding algorithms are probabilistic at their core – i.e., they are trained (explicitly or implicitly BIBREF6) to minimize some form of word co-occurrence probabilities. Thus, we argue that a framework for measuring and treating bias in these embeddings should take into account, in addition to their geometric aspect, their probabilistic nature too. On the other hand, the issue of bias has also been approached (albeit in different contexts) in the fairness literature, where various intuitive notions of equity such as equalized odds have been formalized through probabilistic criteria. By considering analogous criteria for the word embedding setting, we seek to draw connections between these two bodies of work.", "We present experiments on various bias mitigation benchmarks and show that our framework is comparable to state-of-the-art alternatives according to measures of geometric bias mitigation and that it performs far better according to measures of neighborhood bias. For fair comparison, we focus on mitigating a binary gender bias in pre-trained word embeddings using SGNS (skip-gram with negative-sampling), though we note that this framework and methods could be extended to other types of bias and word embedding algorithms." ], [ "Geometric bias mitigation uses the cosine distances between words to both measure and remove gender bias BIBREF0. This method implicitly defines bias as a geometric asymmetry between words when projected onto a subspace, such as the gender subspace constructed from a set of gender pairs such as $\\mathcal {P} = \\lbrace (he,she),(man,woman),(king,queen)...\\rbrace $. The projection of a vector $v$ onto $B$ (the subspace) is defined by $v_B = \\sum _{j=1}^{k} (v \\cdot b_j) b_j$ where a subspace $B$ is defined by k orthogonal unit vectors $B = {b_1,...,b_k}$." ], [ "The WEAT statistic BIBREF1 demonstrates the presence of biases in word embeddings with an effect size defined as the mean test statistic across the two word sets:", "Where $s$, the test statistic, is defined as: $s(w,A,B) = mean_{a \\in A} cos(w,a) - mean_{b \\in B} cos(w,a)$, and $X$,$Y$,$A$, and $B$ are groups of words for which the association is measured. Possible values range from $-2$ to 2 depending on the association of the words groups, and a value of zero indicates $X$ and $Y$ are equally associated with $A$ and $B$. See BIBREF4 for further details on WEAT." ], [ "The RIPA (relational inner product association) metric was developed as an alternative to WEAT, with the critique that WEAT is likely to overestimate the bias of a target attribute BIBREF4. The RIPA metric formalizes the measure of bias used in geometric bias mitigation as the inner product association of a word vector $v$ with respect to a relation vector $b$. The relation vector is constructed from the first principal component of the differences between gender word pairs. We report the absolute value of the RIPA metric as the value can be positive or negative according to the direction of the bias. A value of zero indicates a lack of bias, and the value is bound by $[-||w||,||w||]$." ], [ "The neighborhood bias metric proposed by BIBREF5 quantifies bias as the proportion of male socially-biased words among the $k$ nearest socially-biased male and female neighboring words, whereby biased words are obtained by projecting neutral words onto a gender relation vector. As we only examine the target word among the 1000 most socially-biased words in the vocabulary (500 male and 500 female), a word’s bias is measured as the ratio of its neighborhood of socially-biased male and socially-biased female words, so that a value of 0.5 in this metric would indicate a perfectly unbiased word, and values closer to 0 and 1 indicate stronger bias." ], [ "Our objective here is to extend and complement the geometric notions of word embedding bias described in the previous section with an alternative, probabilistic, approach. Intuitively, we seek a notion of equality akin to that of demographic parity in the fairness literature, which requires that a decision or outcome be independent of a protected attribute such as gender. BIBREF7. Similarly, when considering a probabilistic definition of unbiased in word embeddings, we can consider the conditional probabilities of word pairs, ensuring for example that $p(doctor|man) \\approx p(doctor|woman)$, and can extend this probabilistic framework to include the neighborhood of a target word, addressing the potential pitfalls of geometric bias mitigation.", "Conveniently, most word embedding frameworks allow for immediate computation of the conditional probabilities $P(w|c)$. Here, we focus our attention on the Skip-Gram method with Negative Sampling (SGNS) of BIBREF8, although our framework can be equivalently instantiated for most other popular embedding methods, owing to their core similarities BIBREF6, BIBREF9. Leveraging this probabilistic nature, we construct a bias mitigation method in two steps, and examine each step as an independent method as well as the resulting composite method." ], [ "This component of our bias mitigation framework seeks to enforce that the probability of prediction or outcome cannot depend on a protected class such as gender. We can formalize this intuitive goal through a loss function that penalizes the discrepancy between the conditional probabilities of a target word (i.e., one that should not be affected by the protected attribute) conditioned on two words describing the protected attribute (e.g., man and woman in the case of gender). That is, for every target word we seek to minimize:", "where $\\mathcal {P} = \\lbrace (he,she),(man,woman),(king,queen), \\dots \\rbrace $ is a set of word pairs characterizing the protected attribute, akin to that used in previous work BIBREF0.", "At this point, the specific form of the objective will depend on the type of word embeddings used. For our expample of SGNS, recall that this algorithm models the conditional probability of a target word given a context word as a function of the inner product of their representations. Though an exact method for calculating the conditional probability includes summing over conditional probability of all the words in the vocabulary, we can use the estimation of log conditional probability proposed by BIBREF8, i.e., $ \\log p(w_O|w_I) \\approx \\log \\sigma ({v^{\\prime }_{wo}}^T v_{wI}) + \\sum _{i=1}^{k} [\\log {\\sigma ({{-v^{\\prime }_{wi}}^T v_{wI}})}] $." ], [ "Based on observations by BIBREF5, we extend our method to consider the composition of the neighborhood of socially-gendered words of a target word. We note that bias in a word embedding depends not only on the relationship between a target word and explicitly gendered words like man and woman, but also between a target word and socially-biased male or female words. Bolukbasi et al BIBREF0 proposed a method for eliminating this kind of indirect bias through geometric bias mitigation, but it is shown to be ineffective by the neighborhood metric BIBREF5.", "Instead, we extend our method of bias mitigation to account for this neighborhood effect. Specifically, we examine the conditional probabilities of a target word given the $k/2$ nearest neighbors from the male socially-biased words as well as given the $k/2$ female socially-biased words (in sorted order, from smallest to largest). The groups of socially-biased words are constructed as described in the neighborhood metric. If the word is unbiased according to the neighborhood metric, these probabilities should be comparable. We then use the following as our loss function:", "", "where $m$ and $f$ represent the male and female neighbors sorted by distance to the target word $t$ (we use $L1$ distance)." ], [ "We evaluate our framework on fastText embeddings trained on Wikipedia (2017), UMBC webbase corpus and statmt.org news dataset (16B tokens) BIBREF11. For simplicity, only the first 22000 words are used in all embeddings, though preliminary results indicate the findings extend to the full corpus. For our novel methods of mitigating bias, a shallow neural network is used to adjust the embedding. The single layer of the model is an embedding layer with weights initialized to those of the original embedding. For the composite method, these weights are initialized to those of the embedding after probabilistic bias mitigation. A batch of word indices is fed into the model, which are then embedded and for which a loss value is calculated, allowing back-propagation to adjust the embeddings. For each of the models, a fixed number of iterations is used to prevent overfitting, which can eventually hurt performance on the embedding benchmarks (See Figure FIGREF12). We evaluated the embedding after 1000 iterations, and stopped training if performance on a benchmark decreased significantly.", "We construct a list of candidate words to debias, taken from the words used in the WEAT gender bias statistics. Words in this list should be gender neutral, and are related to the topics of career, arts, science, math, family and professions (see appendix). We note that this list can easily be expanded to include a greater proportion of words in the corpus. For example, BIBREF4 suggested a method for identifying inappropriately gendered words using unsupervised learning.", "We compare this method of bias mitigation with the no bias mitigation (\"Orig\"), geometric bias mitigation (\"Geo\"), the two pieces of our method alone (\"Prob\" and \"KNN\") and the composite method (\"KNN+Prob\"). We note that the composite method performs reasonably well according the the RIPA metric, and much better than traditional geometric bias mitigation according to the neighborhood metric, without significant performance loss according to the accepted benchmarks. To our knowledge this is the first bias mitigation method to perform reasonably both on both metrics." ], [ "We proposed a simple method of bias mitigation based on this probabilistic notions of fairness, and showed that it leads to promising results in various benchmark bias mitigation tasks. Future work should include considering a more rigorous definition and non-binary of bias and experimenting with various embedding algorithms and network architectures." ], [ "The authors would like to thank Tommi Jaakkola for stimulating discussions during the initial stages of this work." ], [ "For Equation 4, as described in the original work, in regards to the k sample words $w_i$ is drawn from the corpus using the Unigram distribution raised to the 3/4 power.", "For reference, the most male socially-biased words include words such as:’john’, ’jr’, ’mlb’, ’dick’, ’nfl’, ’cfl’, ’sgt’, ’abbot’, ’halfback’, ’jock’, ’mike’, ’joseph’,while the most female socially-biased words include words such as:’feminine’, ’marital’, ’tatiana’, ’pregnancy’, ’eva’, ’pageant’, ’distress’, ’cristina’, ’ida’, ’beauty’, ’sexuality’,’fertility’" ], [ "'accountant', 'acquaintance', 'actor', 'actress', 'administrator', 'adventurer', 'advocate', 'aide', 'alderman', 'ambassador', 'analyst', 'anthropologist', 'archaeologist', 'archbishop', 'architect', 'artist', 'assassin', 'astronaut', 'astronomer', 'athlete', 'attorney', 'author', 'baker', 'banker', 'barber', 'baron', 'barrister', 'bartender', 'biologist', 'bishop', 'bodyguard', 'boss', 'boxer', 'broadcaster', 'broker', 'businessman', 'butcher', 'butler', 'captain', 'caretaker', 'carpenter', 'cartoonist', 'cellist', 'chancellor', 'chaplain', 'character', 'chef', 'chemist', 'choreographer', 'cinematographer', 'citizen', 'cleric', 'clerk', 'coach', 'collector', 'colonel', 'columnist', 'comedian', 'comic', 'commander', 'commentator', 'commissioner', 'composer', 'conductor', 'confesses', 'congressman', 'constable', 'consultant', 'cop', 'correspondent', 'counselor', 'critic', 'crusader', 'curator', 'dad', 'dancer', 'dean', 'dentist', 'deputy', 'detective', 'diplomat', 'director', 'doctor', 'drummer', 'economist', 'editor', 'educator', 'employee', 'entertainer', 'entrepreneur', 'envoy', 'evangelist', 'farmer', 'filmmaker', 'financier', 'fisherman', 'footballer', 'foreman', 'gangster', 'gardener', 'geologist', 'goalkeeper', 'guitarist', 'headmaster', 'historian', 'hooker', 'illustrator', 'industrialist', 'inspector', 'instructor', 'inventor', 'investigator', 'journalist', 'judge', 'jurist', 'landlord', 'lawyer', 'lecturer', 'legislator', 'librarian', 'lieutenant', 'lyricist', 'maestro', 'magician', 'magistrate', 'maid', 'manager', 'marshal', 'mathematician', 'mechanic', 'midfielder', 'minister', 'missionary', 'monk', 'musician', 'nanny', 'narrator', 'naturalist', 'novelist', 'nun', 'nurse', 'observer', 'officer', 'organist', 'painter', 'pastor', 'performer', 'philanthropist', 'philosopher', 'photographer', 'physician', 'physicist', 'pianist', 'planner', 'playwright', 'poet', 'policeman', 'politician', 'preacher', 'president', 'priest', 'principal', 'prisoner', 'professor', 'programmer', 'promoter', 'proprietor', 'prosecutor', 'protagonist', 'provost', 'psychiatrist', 'psychologist', 'rabbi', 'ranger', 'researcher', 'sailor', 'saint', 'salesman', 'saxophonist', 'scholar', 'scientist', 'screenwriter', 'sculptor', 'secretary', 'senator', 'sergeant', 'servant', 'singer', 'skipper', 'sociologist', 'soldier', 'solicitor', 'soloist', 'sportsman', 'statesman', 'steward', 'student', 'substitute', 'superintendent', 'surgeon', 'surveyor', 'swimmer', 'teacher', 'technician', 'teenager', 'therapist', 'trader', 'treasurer', 'trooper', 'trumpeter', 'tutor', 'tycoon', 'violinist', 'vocalist', 'waiter', 'waitress', 'warden', 'warrior', 'worker', 'wrestler', 'writer'" ], [ "Words used for WEAT statistic, consisting of baseline bias tests and gender bias tests in the format X vs Y / A vs B", "Flowers vs Insects / Pleasant vs Unpleasant", "X: \"aster\", \"clover\", \"hyacinth\", \"marigold\", \"poppy\", \"azalea\", \"crocus\", \"iris\", \"orchid\", \"rose\", \"bluebell\", \"daffodil\", \"lilac\", \"pansy\", \"tulip\", \"buttercup\", \"daisy\", \"lily\", \"peony\", \"violet\", \"carnation\", \"gladiola\", \"magnolia\", \"petunia\", \"zinnia\"", "Y: \"ant\", \"caterpillar\", \"flea\", \"locust\", \"spider\", \"bedbug\", \"centipede\", \"fly\", \"maggot\", \"tarantula\", \"bee\", \"cockroach\", \"gnat\", \"mosquito\", \"termite\", \"beetle\", \"cricket\", \"hornet\", \"moth\", \"wasp\", \"blackfly\", \"dragonfly\", \"horsefly\", \"roach\", \"weevil\"", "A: \"caress\", \"freedom\", \"health\", \"love\", \"peace\", \"cheer\", \"friend\", \"heaven\", \"loyal\", \"pleasure\", \"diamond\", \"gentle\", \"honest\", \"lucky\", \"rainbow\", \"diploma\", \"gift\", \"honor\", \"miracle\", \"sunrise\", \"family\", \"happy\", \"laughter\", \"paradise\", \"vacation\"", "B: \"abuse\", \"crash\", \"filth\", \"murder\", \"sickness\", \"accident\", \"death\", \"grief\", \"poison\", \"stink\", \"assault\", \"disaster\", \"hatred\", \"pollute\", \"tragedy\", \"divorce\", \"jail\", \"poverty\", \"ugly\", \"cancer\", \"kill\", \"rotten\", \"vomit\", \"agony\", \"prison\"", "Instruments vs Weapons / Pleasant vs Unpleasant:", "X: \"bagpipe\", \"cello\", \"guitar\", \"lute\", \"trombone\", \"banjo\", \"clarinet\", \"harmonica\", \"mandolin\", \"trumpet\", \"bassoon\", \"drum\", \"harp\", \"oboe\", \"tuba\", \"bell\", \"fiddle\", \"harpsichord\", \"piano\", \"viola\", \"bongo\", \"flute\", \"horn\", \"saxophone\", \"violin\"", "Y: \"arrow\", \"club\", \"gun\", \"missile\", \"spear\", \"ax\", \"dagger\", \"harpoon\", \"pistol\", \"sword\", \"blade\", \"dynamite\", \"hatchet\", \"rifle\", \"tank\", \"bomb\", \"firearm\", \"knife\", \"shotgun\", \"teargas\", \"cannon\", \"grenade\", \"mace\", \"slingshot\", \"whip\"", "A: \"caress\", \"freedom\", \"health\", \"love\", \"peace\", \"cheer\", \"friend\", \"heaven\", \"loyal\", \"pleasure\", \"diamond\", \"gentle\", \"honest\", \"lucky\", \"rainbow\", \"diploma\", \"gift\", \"honor\", \"miracle\", \"sunrise\", \"family\", \"happy\", \"laughter\", \"paradise\", \"vacation\"", "B: \"abuse\", \"crash\", \"filth\", \"murder\", \"sickness\", \"accident\", \"death\", \"grief\", \"poison\", \"stink\", \"assault\", \"disaster\", \"hatred\", \"pollute\", \"tragedy\", \"divorce\", \"jail\", \"poverty\", \"ugly\", \"cancer\", \"kill\", \"rotten\", \"vomit\", \"agony\", \"prison\"", "Male vs Female / Career vs Family:", "X: \"brother\", \"father\", \"uncle\", \"grandfather\", \"son\", \"he\", \"his\", \"him\", \"man\", \"himself\", \"men\", \"husband\", \"boy\", \"uncle\", \"nephew\", \"boyfriend\", \"king\", \"actor\"", "Y: \"sister\", \"mother\", \"aunt\", \"grandmother\", \"daughter\", \"she\", \"hers\", \"her\", \"woman\", \"herself\", \"women\", \"wife\", \"aunt\", \"niece\", \"girlfriend\", \"queen\", \"actress\"", "A: \"executive\", \"management\", \"professional\", \"corporation\", \"salary\", \"office\", \"business\", \"career\", \"industry\", \"company\", \"promotion\", \"profession\", \"CEO\", \"manager\", \"coworker\", \"entrepreneur\"", "B: \"home\", \"parents\", \"children\", \"family\", \"cousins\", \"marriage\", \"wedding\", \"relatives\", \"grandparents\", \"grandchildren\", \"nurture\", \"child\", \"toddler\", \"infant\", \"teenager\"", "Math vs Art / Male vs Female:", "X: \"math\", \"algebra\", \"geometry\", \"calculus\", \"equations\", \"computation\", \"numbers\", \"addition\", \"trigonometry\", \"arithmetic\", \"logic\", \"proofs\", \"multiplication\", \"mathematics\"", "Y: \"poetry\", \"art\", \"Shakespeare\", \"dance\", \"literature\", \"novel\", \"symphony\", \"drama\", \"orchestra\", \"music\", \"ballet\", \"arts\", \"creative\", \"sculpture\"", "A: \"brother\", \"father\", \"uncle\", \"grandfather\", \"son\", \"he\", \"his\", \"him\", \"man\", \"himself\", \"men\", \"husband\", \"boy\", \"uncle\", \"nephew\", \"boyfriend\", \"king\", \"actor\"", "B: \"sister\", \"mother\", \"aunt\", \"grandmother\", \"daughter\", \"she\", \"hers\", \"her\", \"woman\", \"herself\", \"women\", \"wife\", \"aunt\", \"niece\", \"girlfriend\", \"queen\", \"actress\"", "Science vs Art / Male8 vs Female8:", "X:\"science\", \"technology\", \"physics\", \"chemistry\", \"Einstein\", \"NASA\", \"experiment\", \"astronomy\", \"biology\", \"aeronautics\", \"mechanics\", \"thermodynamics\"", "Y: \"poetry\", \"art\", \"Shakespeare\", \"dance\", \"literature\", \"novel\", \"symphony\", \"drama\", \"orchestra\", \"music\", \"ballet\", \"arts\", \"creative\", \"sculpture\"", "A: \"brother\", \"father\", \"uncle\", \"grandfather\", \"son\", \"he\", \"his\", \"him\", \"man\", \"himself\", \"men\", \"husband\", \"boy\", \"uncle\", \"nephew\", \"boyfriend\"", "B: \"sister\", \"mother\", \"aunt\", \"grandmother\", \"daughter\", \"she\", \"hers\", \"her\", \"woman\", \"herself\", \"women\", \"wife\", \"aunt\", \"niece\", \"girlfriend\"" ] ] }
{ "question": [ "How is embedding quality assessed?", "What are the three measures of bias which are reduced in experiments?", "What are the probabilistic observations which contribute to the more robust algorithm?" ], "question_id": [ "47726be8641e1b864f17f85db9644ce676861576", "8958465d1eaf81c8b781ba4d764a4f5329f026aa", "31b6544346e9a31d656e197ad01756813ee89422" ], "nlp_background": [ "five", "five", "five" ], "topic_background": [ "familiar", "familiar", "familiar" ], "paper_read": [ "no", "no", "no" ], "search_query": [ "bias", "bias", "bias" ], "question_writer": [ "e8b24c3133e0bec0a6465e1f13acd3a5ed816b37", "e8b24c3133e0bec0a6465e1f13acd3a5ed816b37", "e8b24c3133e0bec0a6465e1f13acd3a5ed816b37" ], "answers": [ { "answer": [ { "unanswerable": false, "extractive_spans": [ "We compare this method of bias mitigation with the no bias mitigation (\"Orig\"), geometric bias mitigation (\"Geo\"), the two pieces of our method alone (\"Prob\" and \"KNN\") and the composite method (\"KNN+Prob\"). We note that the composite method performs reasonably well according the the RIPA metric, and much better than traditional geometric bias mitigation according to the neighborhood metric, without significant performance loss according to the accepted benchmarks. To our knowledge this is the first bias mitigation method to perform reasonably both on both metrics." ], "yes_no": null, "free_form_answer": "", "evidence": [ "We evaluate our framework on fastText embeddings trained on Wikipedia (2017), UMBC webbase corpus and statmt.org news dataset (16B tokens) BIBREF11. For simplicity, only the first 22000 words are used in all embeddings, though preliminary results indicate the findings extend to the full corpus. For our novel methods of mitigating bias, a shallow neural network is used to adjust the embedding. The single layer of the model is an embedding layer with weights initialized to those of the original embedding. For the composite method, these weights are initialized to those of the embedding after probabilistic bias mitigation. A batch of word indices is fed into the model, which are then embedded and for which a loss value is calculated, allowing back-propagation to adjust the embeddings. For each of the models, a fixed number of iterations is used to prevent overfitting, which can eventually hurt performance on the embedding benchmarks (See Figure FIGREF12). We evaluated the embedding after 1000 iterations, and stopped training if performance on a benchmark decreased significantly.", "We construct a list of candidate words to debias, taken from the words used in the WEAT gender bias statistics. Words in this list should be gender neutral, and are related to the topics of career, arts, science, math, family and professions (see appendix). We note that this list can easily be expanded to include a greater proportion of words in the corpus. For example, BIBREF4 suggested a method for identifying inappropriately gendered words using unsupervised learning.", "We compare this method of bias mitigation with the no bias mitigation (\"Orig\"), geometric bias mitigation (\"Geo\"), the two pieces of our method alone (\"Prob\" and \"KNN\") and the composite method (\"KNN+Prob\"). We note that the composite method performs reasonably well according the the RIPA metric, and much better than traditional geometric bias mitigation according to the neighborhood metric, without significant performance loss according to the accepted benchmarks. To our knowledge this is the first bias mitigation method to perform reasonably both on both metrics." ], "highlighted_evidence": [ "We evaluate our framework on fastText embeddings trained on Wikipedia (2017), UMBC webbase corpus and statmt.org news dataset (16B tokens) BIBREF11. For simplicity, only the first 22000 words are used in all embeddings, though preliminary results indicate the findings extend to the full corpus. For our novel methods of mitigating bias, a shallow neural network is used to adjust the embedding. The single layer of the model is an embedding layer with weights initialized to those of the original embedding. For the composite method, these weights are initialized to those of the embedding after probabilistic bias mitigation. A batch of word indices is fed into the model, which are then embedded and for which a loss value is calculated, allowing back-propagation to adjust the embeddings. For each of the models, a fixed number of iterations is used to prevent overfitting, which can eventually hurt performance on the embedding benchmarks (See Figure FIGREF12). We evaluated the embedding after 1000 iterations, and stopped training if performance on a benchmark decreased significantly.", "We construct a list of candidate words to debias, taken from the words used in the WEAT gender bias statistics. Words in this list should be gender neutral, and are related to the topics of career, arts, science, math, family and professions (see appendix). We note that this list can easily be expanded to include a greater proportion of words in the corpus. For example, BIBREF4 suggested a method for identifying inappropriately gendered words using unsupervised learning.", "We compare this method of bias mitigation with the no bias mitigation (\"Orig\"), geometric bias mitigation (\"Geo\"), the two pieces of our method alone (\"Prob\" and \"KNN\") and the composite method (\"KNN+Prob\"). We note that the composite method performs reasonably well according the the RIPA metric, and much better than traditional geometric bias mitigation according to the neighborhood metric, without significant performance loss according to the accepted benchmarks. To our knowledge this is the first bias mitigation method to perform reasonably both on both metrics." ] }, { "unanswerable": true, "extractive_spans": [], "yes_no": null, "free_form_answer": "", "evidence": [], "highlighted_evidence": [] } ], "annotation_id": [ "50e0354ccb4d7d6fda33c34e69133daaa8978a2f", "eb66f1f7e89eca5dcf2ae6ef450b1693a43f4e69" ], "worker_id": [ "a0b403873302db7cada39008f04d01155ef68f4f", "c1018a31c3272ce74964a3280069f62f314a1a58" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [], "yes_no": null, "free_form_answer": "RIPA, Neighborhood Metric, WEAT", "evidence": [ "Geometric bias mitigation uses the cosine distances between words to both measure and remove gender bias BIBREF0. This method implicitly defines bias as a geometric asymmetry between words when projected onto a subspace, such as the gender subspace constructed from a set of gender pairs such as $\\mathcal {P} = \\lbrace (he,she),(man,woman),(king,queen)...\\rbrace $. The projection of a vector $v$ onto $B$ (the subspace) is defined by $v_B = \\sum _{j=1}^{k} (v \\cdot b_j) b_j$ where a subspace $B$ is defined by k orthogonal unit vectors $B = {b_1,...,b_k}$.", "The WEAT statistic BIBREF1 demonstrates the presence of biases in word embeddings with an effect size defined as the mean test statistic across the two word sets:", "Where $s$, the test statistic, is defined as: $s(w,A,B) = mean_{a \\in A} cos(w,a) - mean_{b \\in B} cos(w,a)$, and $X$,$Y$,$A$, and $B$ are groups of words for which the association is measured. Possible values range from $-2$ to 2 depending on the association of the words groups, and a value of zero indicates $X$ and $Y$ are equally associated with $A$ and $B$. See BIBREF4 for further details on WEAT.", "The RIPA (relational inner product association) metric was developed as an alternative to WEAT, with the critique that WEAT is likely to overestimate the bias of a target attribute BIBREF4. The RIPA metric formalizes the measure of bias used in geometric bias mitigation as the inner product association of a word vector $v$ with respect to a relation vector $b$. The relation vector is constructed from the first principal component of the differences between gender word pairs. We report the absolute value of the RIPA metric as the value can be positive or negative according to the direction of the bias. A value of zero indicates a lack of bias, and the value is bound by $[-||w||,||w||]$.", "The neighborhood bias metric proposed by BIBREF5 quantifies bias as the proportion of male socially-biased words among the $k$ nearest socially-biased male and female neighboring words, whereby biased words are obtained by projecting neutral words onto a gender relation vector. As we only examine the target word among the 1000 most socially-biased words in the vocabulary (500 male and 500 female), a word’s bias is measured as the ratio of its neighborhood of socially-biased male and socially-biased female words, so that a value of 0.5 in this metric would indicate a perfectly unbiased word, and values closer to 0 and 1 indicate stronger bias.", "FLOAT SELECTED: Table 1: Remaining Bias (as measured by RIPA and Neighborhood metrics) in fastText embeddings for baseline (top two rows) and our (bottom three) methods. Figure 2: Remaining Bias (WEAT score)" ], "highlighted_evidence": [ "Geometric bias mitigation uses the cosine distances between words to both measure and remove gender bias BIBREF0.", "The WEAT statistic BIBREF1 demonstrates the presence of biases in word embeddings with an effect size defined as the mean test statistic across the two word sets:\n\nWhere $s$, the test statistic, is defined as: $s(w,A,B) = mean_{a \\in A} cos(w,a) - mean_{b \\in B} cos(w,a)$, and $X$,$Y$,$A$, and $B$ are groups of words for which the association is measured.", "The RIPA (relational inner product association) metric was developed as an alternative to WEAT, with the critique that WEAT is likely to overestimate the bias of a target attribute BIBREF4. ", "The neighborhood bias metric proposed by BIBREF5 quantifies bias as the proportion of male socially-biased words among the $k$ nearest socially-biased male and female neighboring words, whereby biased words are obtained by projecting neutral words onto a gender relation vector.", "FLOAT SELECTED: Table 1: Remaining Bias (as measured by RIPA and Neighborhood metrics) in fastText embeddings for baseline (top two rows) and our (bottom three) methods. Figure 2: Remaining Bias (WEAT score)" ] } ], "annotation_id": [ "08a22700ab88c5fb568745e6f7c1b5da25782626" ], "worker_id": [ "c1018a31c3272ce74964a3280069f62f314a1a58" ] }, { "answer": [ { "unanswerable": true, "extractive_spans": [], "yes_no": null, "free_form_answer": "", "evidence": [], "highlighted_evidence": [] } ], "annotation_id": [ "9b4792d66cec53f8ea37bccd5cf7cb9c22290d82" ], "worker_id": [ "c1018a31c3272ce74964a3280069f62f314a1a58" ] } ] }
{ "caption": [ "Figure 1: Word embedding semantic quality benchmarks for each bias mitigation method (higher is better). See Jastrzkebski et al. [11] for details of each metric.", "Table 1: Remaining Bias (as measured by RIPA and Neighborhood metrics) in fastText embeddings for baseline (top two rows) and our (bottom three) methods. Figure 2: Remaining Bias (WEAT score)" ], "file": [ "4-Figure1-1.png", "4-Table1-1.png" ] }
1810.04528
Is there Gender bias and stereotype in Portuguese Word Embeddings?
In this work, we propose an analysis of the presence of gender bias associated with professions in Portuguese word embeddings. The objective of this work is to study gender implications related to stereotyped professions for women and men in the context of the Portuguese language.
{ "section_name": [ "Introduction", "Related Work", "Portuguese Embedding", "Proposed Approach", "Experiments", "Final Remarks" ], "paragraphs": [ [ "Recently, the transformative potential of machine learning (ML) has propelled ML into the forefront of mainstream media. In Brazil, the use of such technique has been widely diffused gaining more space. Thus, it is used to search for patterns, regularities or even concepts expressed in data sets BIBREF0 , and can be applied as a form of aid in several areas of everyday life.", "Among the different definitions, ML can be seen as the ability to improve performance in accomplishing a task through the experience BIBREF1 . Thus, BIBREF2 presents this as a method of inferences of functions or hypotheses capable of solving a problem algorithmically from data representing instances of the problem. This is an important way to solve different types of problems that permeate computer science and other areas.", "One of the main uses of ML is in text processing, where the analysis of the content the entry point for various learning algorithms. However, the use of this content can represent the insertion of different types of bias in training and may vary with the context worked. This work aims to analyze and remove gender stereotypes from word embedding in Portuguese, analogous to what was done in BIBREF3 for the English language. Hence, we propose to employ a public word2vec model pre-trained to analyze gender bias in the Portuguese language, quantifying biases present in the model so that it is possible to reduce the spreading of sexism of such models. There is also a stage of bias reducing over the results obtained in the model, where it is sought to analyze the effects of the application of gender distinction reduction techniques.", "This paper is organized as follows: Section SECREF2 discusses related works. Section SECREF3 presents the Portuguese word2vec embeddings model used in this paper and Section SECREF4 proposes our method. Section SECREF5 presents experimental results, whose purpose is to verify results of a de-bias algorithm application in Portuguese embeddings word2vec model and a short discussion about it. Section SECREF6 brings our concluding remarks." ], [ "There is a wide range of techniques that provide interesting results in the context of ML algorithms geared to the classification of data without discrimination; these techniques range from the pre-processing of data BIBREF4 to the use of bias removal techniques BIBREF5 in fact. Approaches linked to the data pre-processing step usually consist of methods based on improving the quality of the dataset after which the usual classification tools can be used to train a classifier. So, it starts from a baseline already stipulated by the execution of itself. On the other side of the spectrum, there are Unsupervised and semi-supervised learning techniques, that are attractive because they do not imply the cost of corpus annotation BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 .", "The bias reduction is studied as a way to reduce discrimination through classification through different approaches BIBREF10 BIBREF11 . In BIBREF12 the authors propose to specify, implement, and evaluate the “fairness-aware\" ML interface called themis-ml. In this interface, the main idea is to pick up a data set from a modified dataset. Themis-ml implements two methods for training fairness-aware models. The tool relies on two methods to make agnostic model type predictions: Reject Option Classification and Discrimination-Aware Ensemble Classification, these procedures being used to post-process predictions in a way that reduces potentially discriminatory predictions. According to the authors, it is possible to perceive the potential use of the method as a means of reducing bias in the use of ML algorithms.", "In BIBREF3 , the authors propose a method to hardly reduce bias in English word embeddings collected from Google News. Using word2vec, they performed a geometric analysis of gender direction of the bias contained in the data. Using this property with the generation of gender-neutral analogies, a methodology was provided for modifying an embedding to remove gender stereotypes. Some metrics were defined to quantify both direct and indirect gender biases in embeddings and to develop algorithms to reduce bias in some embedding. Hence, the authors show that embeddings can be used in applications without amplifying gender bias." ], [ "In BIBREF13 , the quality of the representation of words through vectors in several models is discussed. According to the authors, the ability to train high-quality models using simplified architectures is useful in models composed of predictive methods that try to predict neighboring words with one or more context words, such as Word2Vec. Word embeddings have been used to provide meaningful representations for words in an efficient way.", "In BIBREF14 , several word embedding models trained in a large Portuguese corpus are evaluated. Within the Word2Vec model, two training strategies were used. In the first, namely Skip-Gram, the model is given the word and attempts to predict its neighboring words. The second, Continuous Bag-of-Words (CBOW), the model is given the sequence of words without the middle one and attempts to predict this omitted word. The latter was chosen for application in the present proposal.", "The authors of BIBREF14 claim to have collected a large corpus from several sources to obtain a multi-genre corpus representative of the Portuguese language. Hence, it comprehensively covers different expressions of the language, making it possible to analyze gender bias and stereotype in Portuguese word embeddings. The dataset used was tokenized and normalized by the authors to reduce the corpus vocabulary size, under the premise that vocabulary reduction provides more representative vectors." ], [ "Some linguists point out that the female gender is, in Portuguese, a particularization of the masculine. In this way the only gender mark is the feminine, the others being considered without gender (including names considered masculine). In BIBREF15 the gender representation in Portuguese is associated with a set of phenomena, not only from a linguistic perspective but also from a socio-cultural perspective. Since most of the termination of words (e.g., advogada and advogado) are used to indicate to whom the expression refers, stereotypes can be explained through communication. This implies the presence of biases when dealing with terms such as those referring to professions.", "Figure FIGREF1 illustrates the approach proposed in this work. First, using a list of professions relating the identification of female and male who perform it as a parameter, we evaluate the accuracy of similarity generated by the embeddings. Then, getting the biased results, we apply the De-bias algorithm BIBREF3 aiming to reduce sexist analogies previous generated. Thus, all the results are analyzed by comparing the accuracies.", "Using the word2vec model available in a public repository BIBREF14 , the proposal involves the analysis of the most similar analogies generated before and after the application of the BIBREF3 . The work is focused on the analysis of gender bias associated with professions in word embeddings. So therefore into the evaluation of the accuracy of the associations generated, aiming at achieving results as good as possible without prejudicing the evaluation metrics.", "Algorithm SECREF4 describes the method performed during the evaluation of the gender bias presence. In this method we try to evaluate the accuracy of the analogies generated through the model, that is, to verify the cases of association matching generated between the words.", "[!htb] Model Evaluation [1]", "w2v_evaluate INLINEFORM0 open_model( INLINEFORM1 ) count = 0 INLINEFORM2 in INLINEFORM3 read list of tuples x = model.most_similar(positive=[`ela', male], negative=[`ele'])", "x = female count += 1 accuracy = count/size(profession_pairs) return accuracy" ], [ "The purpose of this section is to perform different analysis concerning bias in word2vec models with Portuguese embeddings. The Continuous Bag-of-Words model used was provided by BIBREF14 (described in Section SECREF3 ). For these experiments, we use a model containing 934966 words of dimension 300 per vector representation. To realize the experiments, a list containing fifty professions labels for female and male was used as the parameter of similarity comparison.", "Using the python library gensim, we evaluate the extreme analogies generated when comparing vectors like: INLINEFORM0 , where INLINEFORM1 represents the item from professions list and INLINEFORM2 the expected association. The most similarity function finds the top-N most similar entities, computing cosine similarity between a simple mean of the projection weight vectors of the given docs. Figure FIGREF4 presents the most extreme analogies results obtained from the model using these comparisons.", "Applying the Algorithm SECREF4 , we check the accuracy obtained with the similarity function before and after the application of the de-bias method. Table TABREF3 presents the corresponding results. In cases like the analogy of `garçonete' to `stripper' (Figure FIGREF4 , line 8), it is possible to observe that the relationship stipulated between terms with sexual connotation and females is closer than between females and professions. While in the male model, even in cases of non-compliance, the closest analogy remains in the professional environment.", "Using a confidence factor of 99%, when comparing the correctness levels of the model with and without the reduction of bias, the prediction of the model with bias is significantly better. Different authors BIBREF16 BIBREF17 show that the removal of bias in models produces a negative impact on the quality of the model. On the other hand, it is observed that even with a better hit rate the correctness rate in the prediction of related terms is still low." ], [ "This paper presents an analysis of the presence of gender bias in Portuguese word embeddings. Even though it is a work in progress, the proposal showed promising results in analyzing predicting models.", "A possible extension of the work involves deepening the analysis of the results obtained, seeking to achieve higher accuracy rates and fairer models to be used in machine learning techniques. Thus, these studies can involve tests with different methods of pre-processing the data to the use of different models, as well as other factors that may influence the results generated. This deepening is necessary since the model's accuracy is not high.", "To conclude, we believe that the presence of gender bias and stereotypes in the Portuguese language is found in different spheres of language, and it is important to study ways of mitigating different types of discrimination. As such, it can be easily applied to analyze racists bias into the language, such as different types of preconceptions." ] ] }
{ "question": [ "Does this paper target European or Brazilian Portuguese?", "What were the word embeddings trained on?", "Which word embeddings are analysed?" ], "question_id": [ "519db0922376ce1e87fcdedaa626d665d9f3e8ce", "99a10823623f78dbff9ccecb210f187105a196e9", "09f0dce416a1e40cc6a24a8b42a802747d2c9363" ], "nlp_background": [ "five", "five", "five" ], "topic_background": [ "familiar", "familiar", "familiar" ], "paper_read": [ "no", "no", "no" ], "search_query": [ "bias", "bias", "bias" ], "question_writer": [ "e8b24c3133e0bec0a6465e1f13acd3a5ed816b37", "e8b24c3133e0bec0a6465e1f13acd3a5ed816b37", "e8b24c3133e0bec0a6465e1f13acd3a5ed816b37" ], "answers": [ { "answer": [ { "unanswerable": true, "extractive_spans": [], "yes_no": null, "free_form_answer": "", "evidence": [], "highlighted_evidence": [] }, { "unanswerable": true, "extractive_spans": [], "yes_no": null, "free_form_answer": "", "evidence": [], "highlighted_evidence": [] } ], "annotation_id": [ "0a93ba2daf6764079c983e70ca8609d6d1d8fa5c", "c6686e4e6090f985be4cc72a08ca2d4948b355bb" ], "worker_id": [ "71f73551e7aabf873649e8fe97aefc54e6dd14f8", "34c35a1877e453ecaebcf625df3ef788e1953cc4" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "large Portuguese corpus" ], "yes_no": null, "free_form_answer": "", "evidence": [ "In BIBREF14 , several word embedding models trained in a large Portuguese corpus are evaluated. Within the Word2Vec model, two training strategies were used. In the first, namely Skip-Gram, the model is given the word and attempts to predict its neighboring words. The second, Continuous Bag-of-Words (CBOW), the model is given the sequence of words without the middle one and attempts to predict this omitted word. The latter was chosen for application in the present proposal.", "Using the word2vec model available in a public repository BIBREF14 , the proposal involves the analysis of the most similar analogies generated before and after the application of the BIBREF3 . The work is focused on the analysis of gender bias associated with professions in word embeddings. So therefore into the evaluation of the accuracy of the associations generated, aiming at achieving results as good as possible without prejudicing the evaluation metrics." ], "highlighted_evidence": [ "In BIBREF14 , several word embedding models trained in a large Portuguese corpus are evaluated. ", "Using the word2vec model available in a public repository BIBREF14 , the proposal involves the analysis of the most similar analogies generated before and after the application of the BIBREF3 . " ] } ], "annotation_id": [ "e0cd186397ec9543e48d25f5944fc9318542f1d5" ], "worker_id": [ "34c35a1877e453ecaebcf625df3ef788e1953cc4" ] }, { "answer": [ { "unanswerable": false, "extractive_spans": [ "Continuous Bag-of-Words (CBOW)" ], "yes_no": null, "free_form_answer": "", "evidence": [ "In BIBREF14 , several word embedding models trained in a large Portuguese corpus are evaluated. Within the Word2Vec model, two training strategies were used. In the first, namely Skip-Gram, the model is given the word and attempts to predict its neighboring words. The second, Continuous Bag-of-Words (CBOW), the model is given the sequence of words without the middle one and attempts to predict this omitted word. The latter was chosen for application in the present proposal." ], "highlighted_evidence": [ "The second, Continuous Bag-of-Words (CBOW), the model is given the sequence of words without the middle one and attempts to predict this omitted word. " ] } ], "annotation_id": [ "8b5278bfc35cf0a1b43ceb3418c2c5d20f213a31" ], "worker_id": [ "34c35a1877e453ecaebcf625df3ef788e1953cc4" ] } ] }
{ "caption": [ "Fig. 1. Proposal", "Fig. 2. Extreme Analogies" ], "file": [ "3-Figure1-1.png", "5-Figure2-1.png" ] }
2002.02224
Citation Data of Czech Apex Courts
"In this paper, we introduce the citation data of the Czech apex courts (Supreme Court, Supreme Admi(...TRUNCATED)
{"section_name":["Introduction","Related work ::: Legal Citation Analysis","Related work ::: Referen(...TRUNCATED)
{"question":["Did they experiment on this dataset?","How is quality of the citation measured?","How (...TRUNCATED)
{"caption":["Figure 1: NLP pipeline including the text segmentation, reference recognition and parsi(...TRUNCATED)
2003.07433
"LAXARY: A Trustworthy Explainable Twitter Analysis Model for Post-Traumatic Stress Disorder Assessm(...TRUNCATED)
"Veteran mental health is a significant national problem as large number of veterans are returning f(...TRUNCATED)
{"section_name":["Introduction","Overview","Related Works","Demographics of Clinically Validated PTS(...TRUNCATED)
{"question":["Do they evaluate only on English datasets?","Do the authors mention any possible confo(...TRUNCATED)
{"caption":["Fig. 1. Overview of our framework","Fig. 2. WordStat dictionary sample","TABLE I DRYHOO(...TRUNCATED)
2003.12218
Comprehensive Named Entity Recognition on CORD-19 with Distant or Weak Supervision
"We created this CORD-19-NER dataset with comprehensive named entity recognition (NER) on the COVID-(...TRUNCATED)
{"section_name":["Introduction","CORD-19-NER Dataset ::: Corpus","CORD-19-NER Dataset ::: NER Method(...TRUNCATED)
{"question":["Did they experiment with the dataset?","What is the size of this dataset?","Do they li(...TRUNCATED)
{"caption":["Table 1: Performance comparison on three major biomedical entity types in COVID-19 corp(...TRUNCATED)
1904.09678
UniSent: Universal Adaptable Sentiment Lexica for 1000+ Languages
"In this paper, we introduce UniSent a universal sentiment lexica for 1000 languages created using a(...TRUNCATED)
{"section_name":["Introduction","Method","Experimental Setup","Results","Conclusion"],"paragraphs":[(...TRUNCATED)
{"question":["how is quality measured?","how many languages exactly is the sentiment lexica for?","w(...TRUNCATED)
{"caption":["Figure 1: Neighbors of word ’sensual’ in Spanish, in bible embedding graph (a) and (...TRUNCATED)
End of preview. Expand in Data Studio

Dataset Card for "emozillaqasper-pruned-llama-gptneox-8k"

More Information needed

Downloads last month
4