Datasets:
Dataset Viewer
ID
large_stringlengths 10
67
| year
int64 1.96k
2.03k
| title
large_stringlengths 4
560
| abstract
large_stringlengths 0
12.8k
|
|---|---|---|---|
buhnila-etal-2025-chain | 2,025 | Chain-of-MetaWriting: Linguistic and Textual Analysis of How Small Language Models Write Young Students Texts | Large Language Models (LLMs) have been used to generate texts in response to different writing tasks: reports, essays, story telling. However, language models do not have a metarepresentation of the text writing process, nor inherent communication learning needs, comparable to those of young human students. This paper introduces a fine-grained linguistic and textual analysis of multilingual Small Language Models' (SLMs) writing. With our method, Chain-of-MetaWriting, SLMs can imitate some steps of the human writing process, such as planning and evaluation. We mainly focused on short story and essay writing tasks in French for schoolchildren and undergraduate students respectively. Our results show that SLMs encounter difficulties in assisting young students on sensitive topics such as violence in the schoolyard, and they sometimes use words too complex for the target audience. In particular, the output is quite different from the human produced texts in term of text cohesion and coherence regarding temporal connectors, topic progression, reference. |
shi-penn-2025-semantic | 2,025 | Semantic Masking in a Needle-in-a-haystack Test for Evaluating Large Language Model Long-Text Capabilities | In this paper, we introduce the concept of Semantic Masking, where semantically coherent surrounding text (the haystack) interferes with the retrieval and comprehension of specific information (the needle) embedded within it. We propose the Needle-in-a-Haystack-QA Test, an evaluation pipeline that assesses LLMs' long-text capabilities through question answering, explicitly accounting for the Semantic Masking effect. We conduct experiments to demonstrate that Semantic Masking significantly impacts LLM performance more than text length does. By accounting for Semantic Masking, we provide a more accurate assessment of LLMs' true proficiency in utilizing extended contexts, paving the way for future research to develop models that are not only capable of handling longer inputs but are also adept at navigating complex semantic landscapes. |
khallaf-etal-2025-reading | 2,025 | Reading Between the Lines: A dataset and a study on why some texts are tougher than others | Our research aims at better understanding what makes a text difficult to read for specific audiences with intellectual disabilities, more specifically, people who have limitations in cognitive functioning, such as reading and understanding skills, an IQ below 70, and challenges in conceptual domains. We introduce a scheme for the annotation of difficulties which is based on empirical research in psychology as well as on research in translation studies. The paper describes the annotated dataset, primarily derived from the parallel texts (standard English and Easy to Read English translations) made available online. we fine-tuned four different pre-trained transformer models to perform the task of multiclass classification to predict the strategies required for simplification. We also investigate the possibility to interpret the decisions of this language model when it is aimed at predicting the difficulty of sentences in this dataset. |
jourdan-etal-2025-pararev | 2,025 | ParaRev : Building a dataset for Scientific Paragraph Revision annotated with revision instruction | Revision is a crucial step in scientific writing, where authors refine their work to improve clarity, structure, and academic quality. Existing approaches to automated writing assistance often focus on sentence-level revisions, which fail to capture the broader context needed for effective modification. In this paper, we explore the impact of shifting from sentence-level to paragraph-level scope for the task of scientific text revision. The paragraph level definition of the task allows for more meaningful changes, and is guided by detailed revision instructions rather than general ones. To support this task, we introduce ParaRev, the first dataset of revised scientific paragraphs with an evaluation subset manually annotated with revision instructions. Our experiments demonstrate that using detailed instructions significantly improves the quality of automated revisions compared to general approaches, no matter the model or the metric considered. |
maggi-vitaletti-2025-towards | 2,025 | Towards an operative definition of creative writing: a preliminary assessment of creativeness in AI and human texts | Nowadays, AI is present in all our activities. This pervasive presence is perceived as a threat by many category of users that might be substituted by their AI counterpart. While the potential of AI in handling repetitive tasks is clear, the potentials of its creativeness is still misunderstood. We believe that understanding this aspects of AI can transform a threat into an opportunity. This paper is a first attempt to provide a measurable definition of creativeness. We applied our definition to AI and human generated texts, proving the viability of the proposed approach. Our preliminary experiments show that human texts are more creative. |
sato-kobayashi-2025-decoding | 2,025 | Decoding Semantic Representations in the Brain Under Language Stimuli with Large Language Models | Brain decoding technology is paving the way for breakthroughs in the interpretation of neural activity to recreate thoughts, emotions, and movements. Tang et al. (2023) introduced a novel approach that uses language models as generative models for brain decoding based on functional magnetic resonance imaging (fMRI) data. Building on their work, this study explored the use of three additional language models along with the GPT model used in previous research to improve decoding accuracy. Furthermore, we added an evaluation metric using an embedding model, providing higher-level semantic similarity than the BERTScore. By comparing the decoding performance and identifying the factors contributing to good performance, we found that high decoding accuracy does not solely depend on the ability to accurately predict brain activity. Instead, the type of text (e.g., web text, blogs, news articles, and books) that the model tends to generate plays a more significant role in achieving more precise sentence reconstruction. |
drakesmith-etal-2025-towards | 2,025 | Towards a Social Media-based Disease Surveillance System for Early Detection of Influenza-like Illnesses: A Twitter Case Study in Wales | Social media offers the potential to provide detection of outbreaks or public health incidents faster than traditional reporting mechanisms. In this paper, we developed and tested a pipeline to produce alerts of influenza-like illness (ILI) using Twitter data. Data was collected from the Twitter API, querying keywords referring to ILI symptoms and geolocated to Wales. Tweets that described first-hand descriptions of symptoms (as opposed to non-personal descriptions) were classified using transformer-based language models specialised on social media (BERTweet and TimeLMs), which were trained on a manually labelled dataset matching the above criteria. After gathering this data, weekly tweet counts were applied to the regression-based Noufaily algorithm to identify exceedances throughout 2022. The algorithm was also applied to counts of ILI-related GP consultations for comparison. Exceedance detection applied to the classified tweet counts produced alerts starting four weeks earlier than by using GP consultation data. These results demonstrate the potential to facilitate advanced preparedness for unexpected increases in healthcare burdens. |
du-etal-2025-sentiment | 2,025 | Sentiment Analysis on Video Transcripts: Comparing the Value of Textual and Multimodal Annotations | This study explores the differences between textual and multimodal sentiment annotations on videos and their impact on transcript-based sentiment modelling. Using the UniC and CH-SIMS datasets which are annotated at both the unimodal and multimodal level, we conducted a statistical analysis and sentiment modelling experiments. Results reveal significant differences between the two annotation types, with textual annotations yielding better performance in sentiment modelling and demonstrating superior generalization ability. These findings highlight the challenges of cross-modality generalization and provide insights for advancing sentiment analysis. |
shmidman-shmidman-2025-restoring | 2,025 | Restoring Missing Spaces in Scraped Hebrew Social Media | A formidable challenge regarding scraped corpora of social media is the omission of whitespaces, causing pairs of words to be conflated together as one. In order for the text to be properly parsed and analyzed, these missing spaces must be detected and restored. However, it is particularly hard to restore whitespace in languages such as Hebrew which are written without vowels, because a conflated form can often be split into multiple different pairs of valid words. Thus, a simple dictionary lookup is not feasible. In this paper, we present and evaluate a series of neural approaches to restore missing spaces in scraped Hebrew social media. Our best all-around method involved pretraining a new character-based BERT model for Hebrew, and then fine-tuning a space restoration model on top of this new BERT model. This method is blazing fast, high-performing, and open for unrestricted use, providing a practical solution to process huge Hebrew social media corpora with a consumer-grade GPU. We release the new BERT model and the fine-tuned space-restoration model to the NLP community. |
juffs-naismith-2025-identifying | 2,025 | Identifying and analyzing \textquoteleftnoisy' spelling errors in a second language corpus | This paper addresses the problem of identifying and analyzing {\textquoteleft}noisy' spelling errors in texts written by second language (L2) learners' texts in a written corpus. Using Python, spelling errors were identified in 5774 texts greater than or equal to 66 words (total=1,814,209 words), selected from a corpus of 4.2 million words (Authors-1). The statistical analysis used hurdle() models in R, which are appropriate for non-normal, count data, with many zeros. |
maarouf-tanguy-2025-automatic | 2,025 | Automatic normalization of noisy technical reports with an LLM: What effects on a downstream task? | This study explores the automatic normalization of noisy and highly technical anomaly reports by an LLM. Different prompts are tested to instruct the LLM to clean the text without changing the structure, vocabulary or specialized lexicon. The evaluation of this task is made in two steps. First, the Character Error Rate (CER) is calculated to assess the changes made compared to a gold standard on a small sample. Second, an automatic sequence labeling task is performed on the original and on the corrected datasets with a transformer-based classifier. If some configurations of LLM and prompts can reach satisfying CER scores, the sequence labeling task shows that the normalization has a small negative impact on performance. |
srivastava-chiang-2025-calling | 2,025 | We`re Calling an Intervention: Exploring Fundamental Hurdles in Adapting Language Models to Nonstandard Text | We present a suite of experiments that allow us to understand the underlying challenges of language model adaptation to nonstandard text. We do so by designing interventions that approximate core features of user-generated text and their interactions with existing biases of language models. Applying our interventions during language model adaptation to nonstandard text variations, we gain important insights into when such adaptation is successful, as well as the aspects of text variation and noise that are particularly difficult for language models to handle. For instance, on text with character-level variation, out-of-the-box performance improves even with a few additional training examples but approaches a plateau, suggesting that more data is not the solution. In contrast, on text with variation involving new words or meanings, far more data is needed, but it leads to a massive breakthrough in performance. Our findings reveal that existing models lack the necessary infrastructure to handle diverse forms of nonstandard text, guiding the development of more resilient language modeling techniques. We make the code for our interventions, which can be applied to any English text data, publicly available. |
birkmose-etal-2025-device | 2,025 | On-Device LLMs for Home Assistant: Dual Role in Intent Detection and Response Generation | This paper investigates whether Large Language Models (LLMs), fine-tuned on synthetic but domain-representative data, can perform the twofold task of (i) slot and intent detection and (ii) natural language response generation for a smart home assistant, while running solely on resource-limited, CPU-only edge hardware. We fine-tune LLMs to produce both JSON action calls and text responses. Our experiments show that 16-bit and 8-bit quantized variants preserve high accuracy on slot and intent detection and maintain strong semantic coherence in generated text, while the 4-bit model, while retaining generative fluency, suffers a noticeable drop in device-service classification accuracy. Further evaluations on noisy human (non-synthetic) prompts and out-of-domain intents confirm the models' generalization ability, obtaining around 80{--}86{\%} accuracy. While the average inference time is 5{--}6 seconds per query{---}acceptable for one-shot commands but suboptimal for multi-turn dialogue{---}our results affirm that an on-device LLM can effectively unify command interpretation and flexible response generation for home automation without relying on specialized hardware. |
gonzalez-lopez-etal-2025-applying | 2,025 | Applying Transformer Architectures to Detect Cynical Comments in Spanish Social Media | Detecting cynical comments in online communication poses a significant challenge in human-computer interaction, especially given the massive proliferation of discussions on platforms like YouTube. These comments often include offensive or disruptive patterns, such as sarcasm, negative feelings, specific reasons, and an attitude of being right. To address this problem, we present a web platform for the Spanish language that has been developed and leverages natural language processing and machine learning techniques. The platform detects comments and provides valuable information to users by focusing on analyzing comments. The core models are based on pre-trained architectures, including BETO, SpanBERTa, Multilingual BERT, RoBERTuito, and BERT, enabling robust detection of cynical comments. Our platform was trained and tested with Spanish comments from car analysis channels on YouTube. The results show that models achieve performance above 0.8 F1 for all types of cynical comments in the text classification task but achieve lower performance (around 0.6-0.7 F1) for the more arduous token classification task. |
mirbeygi-beigy-2025-prompt | 2,025 | Prompt Guided Diffusion for Controllable Text Generation | Controlled text generation, originally a task to generate coherent, contextually relevant text with specified attributes such as sentiment, topic, or style, has seen a lot of development with methods that use PPLM, FUDGE, and diffusion-based models. However, most state-of-the-art methods balance control precision with fluency. Classifier-guided approaches, like PPLM, are well-known for unstable updates of gradients, yielding incoherent outputs, while autoregressive models, like FUDGE, depend on rigid templates that limit creativity. While recent diffusion models show promises in iterative refinement and diversity, they often lack mechanisms to explicitly incorporate task-specific knowledge and hence require various complicated auxiliary classifiers for training and inference.We now propose a prompt-guided diffusion framework that integrates structured prompts seamlessly into the process of diffusion for precise and flexible control of generated texts.Each prompt combines a target condition (e.g., sentiment label), an in-class example (e.g., a positive movie review), and a placeholder for the generated sentence. Explicit, human-readable guidance is thereby given, spanning high-level intent to low-level text generation.Our approach encodes prompts using large pre-trained language models, e.g., BART, fusing these in a cross-attention manner with the diffusion dynamics, achieves new state-of-the-art results for all benchmarks, including IMDB for sentiment, AG News for topic, and E2E for structured data-to-text generation. |
masumi-etal-2025-fabert | 2,025 | FaBERT: Pre-training BERT on Persian Blogs | We introduce FaBERT, a Persian BERT-base model pre-trained on the HmBlogs corpus, encompassing both informal and formal Persian texts. FaBERT is designed to excel in traditional Natural Language Understanding (NLU) tasks, addressing the intricacies of diverse sentence structures and linguistic styles prevalent in the Persian language. In our comprehensive evaluation of FaBERT on 12 datasets in various downstream tasks, encompassing Sentiment Analysis (SA), Named Entity Recognition (NER), Natural Language Inference (NLI), Question Answering (QA), and Question Paraphrasing (QP), it consistently demonstrated improved performance, all achieved within a compact model size. The findings highlight the importance of utilizing diverse corpora, such as HmBlogs, to enhance the performance of language models like BERT in Persian Natural Language Processing (NLP) applications. |
qian-etal-2025-automatically | 2,025 | Automatically Generating Chinese Homophone Words to Probe Machine Translation Estimation Systems | Evaluating machine translation (MT) of user-generated content (UGC) involves unique challenges such as checking whether the nuance of emotions from the source are preserved in the target text. Recent studies have proposed emotion-related datasets, frameworks and models to automatically evaluate MT quality of Chinese UGC, without relying on reference translations. However, whether these models are robust to the challenge of preserving emotional nuances has been left largely unexplored. To this end, we introduce a novel method inspired by information theory which generates challenging Chinese homophone words related to emotions, by leveraging the concept of *self-information*. Our approach generates homophones that were observed to cause translation errors in emotion preservation, and exposes vulnerabilities in MT models struggling to preserve relevant emotions. We evaluate the efficacy of our method using human evaluation and compare it with an existing one, showing that our method achieves higher correlation with human judgments. The generated Chinese homophones, along with their manual translations, are utilized to generate perturbations and to probe the robustness of existing quality evaluation models, including models trained using multi-task learning, fine-tuned variants of multilingual language models, as well as large language models (LLMs). Our results indicate that LLMs with larger size exhibit higher stability and robustness to such perturbations. We release our data and code for reproducibility and further research. |
abed-azad-beigy-2025-multi | 2,025 | Multi-BERT: Leveraging Adapters for Low-Resource Multi-Domain Adaptation | Multi-domain text analysis presents significant challenges, particularly in Persian name entity recognition (NER). Using a single model for multiple domains often fails to capture the specific features of different domains. That is why many scientists have focused on prompting chatbots for this issue. However, studies show that these models do not achieve remarkable results in NER tasks without proper fine-tuning while training and storing a chatbot is extremely costly. This paper presents a new approach using one core model with various sets of domain-specific parameters. By using techniques like LoRAs and pre-fix tuning, along with extra layers, we train each set of trainable parameters for a specific domain. This allows the model to perform as well as individual models for each domain. Tests on various formal and informal datasets show that by using these added parameters, the proposed model performs much better than existing practical models. The model needs only one instance for storage but achieves excellent results across all domains. This paper also examines each adaptation strategy, outlining its strengths, weaknesses, and the best settings and hyperparameters for Persian NER. Lastly, this study introduces a new document-based domain detection system for situations where text domains are unknown. This novel pipeline enhances the adaptability and practicality of the proposed approach for real-world applications. |
ehsan-solorio-2025-enhancing | 2,025 | Enhancing NER Performance in Low-Resource Pakistani Languages using Cross-Lingual Data Augmentation | Named Entity Recognition (NER), a fundamental task in Natural Language Processing (NLP), has shown significant advancements for high-resource languages. However, due to a lack of annotated datasets and limited representation in Pre-trained Language Models (PLMs), it remains understudied and challenging for low-resource languages. To address these challenges, in this paper, we propose a data augmentation technique that generates culturally plausible sentences and experiments on four low-resource Pakistani languages; Urdu, Shahmukhi, Sindhi, and Pashto. By fine-tuning multilingual masked Large Language Models (LLMs), our approach demonstrates significant improvements in NER performance for Shahmukhi and Pashto. We further explore the capability of generative LLMs for NER and data augmentation using few-shot learning. |
borkakoty-espinosa-anke-2025-wikipedia | 2,025 | Wikipedia is Not a Dictionary, Delete! Text Classification as a Proxy for Analysing Wiki Deletion Discussions | Automated content moderation for collaborative knowledge hubs like Wikipedia or Wikidata is an important yet challenging task due to multiple factors. In this paper, we construct a database of discussions happening around articles marked for deletion in several Wikis and in three languages, which we then use to evaluate a range of LMs on different tasks (from predicting the outcome of the discussion to identifying the implicit policy an individual comment might be pointing to). Our results reveal, among others, that discussions leading to deletion are easier to predict, and that, surprisingly, self-produced tags (keep, delete or redirect) don`t always help guiding the classifiers, presumably because of users' hesitation or deliberation within comments |
znotins-gruzitis-2025-conversational | 2,025 | From Conversational Speech to Readable Text: Post-Processing Noisy Transcripts in a Low-Resource Setting | We present ongoing research on automatic post-processing approaches to enhance the readability of noisy speech transcripts in low-resource languages, with a focus on conversational speech in Latvian. We compare transformer-based sequence-labeling models and large language models (LLMs) for the standard punctuation and capitalization restoration task, while also considering automatic correction of mispronounced words and disfluency, and partial inverse text normalization. Our results show that very small LLMs (approx. 2B parameters), fine-tuned on a modest text corpus, can achieve near state-of-the-art performance, rivaling orders of magnitude larger LLMs. Additionally, we demonstrate that a fine-tuned Whisper model, leveraging acoustic cues, outperforms text-only systems on challenging conversational data, even for a low-resource language. Error analysis reveals recurring pitfalls in sentence boundary determination and disfluency handling, emphasizing the importance of consistent annotation and domain adaptation for robust post-processing. Our findings highlight the feasibility of developing efficient post-processing solutions that significantly refine ASR output in low-resource settings, while opening new possibilities for editing and formatting speech transcripts beyond mere restoration of punctuation and capitalization. |
kondo-etal-2025-text | 2,025 | Text Normalization for Japanese Sentiment Analysis | We manually normalize noisy Japanese expressions on social networking services (SNS) to improve the performance of sentiment polarity classification.Despite advances in pre-trained language models, informal expressions found in social media still plague natural language processing.In this study, we analyzed 6,000 posts from a sentiment analysis corpus for Japanese SNS text, and constructed a text normalization taxonomy consisting of 33 types of editing operations.Text normalization according to our taxonomy significantly improved the performance of BERT-based sentiment analysis in Japanese.Detailed analysis reveals that most types of editing operations each contribute to improve the performance of sentiment analysis. |
piper-bagga-2025-narradetect | 2,025 | NarraDetect: An annotated dataset for the task of narrative detection | Narrative detection is an important task across diverse research domains where storytelling serves as a key mechanism for explaining human beliefs and behavior. However, the task faces three significant challenges: (1) inter-narrative heterogeneity, or the variation in narrative communication across social contexts; (2) intra-narrative heterogeneity, or the dynamic variation of narrative features within a single text over time; and (3) the lack of theoretical consensus regarding the concept of narrative. This paper introduces the NarraDetect dataset, a comprehensive resource comprising over 13,000 passages from 18 distinct narrative and non-narrative genres. Through a manually annotated subset of {\textasciitilde}400 passages, we also introduce a novel theoretical framework for annotating for a scalar concept of {\textquotedblleft}narrativity.{\textquotedblright} Our findings indicate that while supervised models outperform large language models (LLMs) on this dataset, LLMs exhibit stronger generalization and alignment with the scalar concept of narrativity. |
dey-lal-2025-transferability | 2,025 | On the Transferability of Causal Knowledge for Language Models | Language understanding includes identifying logical connections between events in a discourse, such as news and instructional text. We study the transferability of causal knowledge across these two domains by analyzing the extent to which understanding preconditions in narratives such as news articles can help models reason about cooking recipes, and vice-versa. Our experiments show that using instructions to pretrain small models on one domain before similarly finetuning it on the other shows a slight improvement over just finetuning it. We also find that finetuning the models on a mix of both types of data is better ({\textasciitilde}3-7{\%}) for understanding causal relations in instructional text. While we find that the improvements do not translate to larger or already instruction tuned models, our analysis highlights the aspects of a plan that are better captured through the interoperability of causal knowledge. |
shokri-etal-2025-finding | 2,025 | Finding Common Patterns in Domestic Violence Stories Posted on Reddit | Domestic violence survivors often share their experiences in online spaces, offering valuable insights into common abuse patterns. This study analyzes a dataset of personal narratives about domestic violence from Reddit, focusing on event extraction and topic modeling to uncover recurring themes. We evaluate GPT-4 and LLaMA-3.1 for extracting key sentences, finding that GPT-4 exhibits higher precision, while LLaMA-3.1 achieves better recall. Using LLM-based topic assignment, we identify dominant themes such as psychological aggression, financial abuse, and physical assault which align with previously published psychology findings. A co-occurrence and PMI analysis further reveals the interdependencies among different abuse types, emphasizing the multifaceted nature of domestic violence. Our findings provide a structured approach to analyzing survivor narratives, with implications for social support systems and policy interventions. |
bissell-etal-2025-theoretical | 2,025 | A Theoretical Framework for Evaluating Narrative Surprise in Large Language Models | Narrative surprise is a core element of storytelling for engaging audiences, and yet it remains underexplored in the context of large language models (LLMs) and narrative generation. While surprise arises from events that deviate from expectations while maintaining retrospective coherence, current computational approaches lack comprehensive frameworks to evaluate this phenomenon. This paper presents a novel framework for assessing narrative surprise, drawing on psychological theories of narrative comprehension and surprise intensity. We operationalize six criteria{---}initiatoriness, immutability violation, predictability, post-dictability, importance, and valence{---}to measure narrative surprise in story endings. Our study evaluates 120 story endings, generated by both human authors and LLMs, across 30 mystery narratives. Through a ranked-choice voting methodology, we identify significant correlations between reader preferences and four of the six criteria. Results underscore the continuing advantage of human-authored endings in achieving compelling narrative surprise, while also revealing significant progress in LLM-generated narratives. |
baruah-narayanan-2025-chatter | 2,025 | CHATTER: A character-attribution dataset for narrative understanding | Computational narrative understanding studies the identification, description, and interaction of the elements of a narrative: characters, attributes, events, and relations.Narrative research has given considerable attention to defining and classifying character types.However, these character-type taxonomies do not generalize well because they are small, too simple, or specific to a domain.We require robust and reliable benchmarks to test whether narrative models truly understand the nuances of the character`s development in the story.Our work addresses this by curating the CHATTER dataset that labels whether a character portrays some attribute for 88124 character-attribute pairs, encompassing 2998 characters, 12967 attributes and 660 movies.We validate a subset of CHATTER, called CHATTEREVAL, using human annotations to serve as an evaluation benchmark for the character attribution task in movie scripts.CHATTEREVAL also assesses narrative understanding and the long-context modeling capacity of language models. |
sancheti-rudinger-2025-tracking | 2,025 | Tracking Evolving Relationship Between Characters in Books in the Era of Large Language Models | This work aims to assess the zero-shot social reasoning capabilities of LLMs by proposing various strategies based on the granularity of information used to track the fine-grained evolution in the relationship between characters in a book. Without gold annotations, we thoroughly analyze the agreements between predictions from multiple LLMs and manually examine their consensus at a local and global level via the task of trope prediction. Our findings reveal low-to-moderate agreement among LLMs and humans, reflecting the complexity of the task. Analysis shows that LLMs are sensitive to subtle contextual changes and often rely on surface-level cues. Humans, too, may interpret relationships differently, leading to disagreements in annotations. |
ghaffari-hokamp-2025-narrative | 2,025 | Narrative Studio: Visual narrative exploration using LLMs and Monte Carlo Tree Search | Interactive storytelling benefits from planning and exploring multiple {\textquotedblleft}what if{\textquotedblright} scenarios. Modern LLMs are useful tools for ideation and exploration, but current chat-based user interfaces restrict users to a single linear flow. To address this limitation, we propose Narrative Studio {--} a novel in-browser narrative exploration environment featuring a tree-like interface that allows branching exploration from user-defined points in a story. Each branch is extended via iterative LLM inference guided by system and user-defined prompts. Additionally, we employ Monte Carlo Tree Search (MCTS) to automatically expand promising narrative paths based on user-specified criteria, enabling more diverse and robust story development. We also allow users to enhance narrative coherence by grounding the generated text in a graph that represents the actors and environment of the story. |
gobara-etal-2025-speaker | 2,025 | Speaker Identification and Dataset Construction Using LLMs: A Case Study on Japanese Narratives | Speaker identification in narrative analysis is a challenging task due to complex dialogues, diverse utterance patterns, and ambiguous character references. Cosly and time-intensive manual annotation limits the scalability of high-quality dataset creation.This study demonstrates a cost-efficient approach of constructing speaker identification datasets by combining small-scale manual annotation with LLM-based labeling. A subset of data is manually annotated and is used to guide LLM predictions with a few-shot approach followed by refinement through minimal human corrections. Our results show that LLMs achieve approximately 90{\%} accuracy on challenging narratives, such as the {\textquotedblleft}Three Kingdoms{\textquotedblright} dataset, underscoring the importance of targeted human corrections. This approach proves effective for constructing scalable and cost-efficient datasets for Japanese and complex narratives. |
lamsiyah-etal-2025-arabicsense | 2,025 | ArabicSense: A Benchmark for Evaluating Commonsense Reasoning in Arabic with Large Language Models | Recent efforts in natural language processing (NLP) commonsense reasoning research have led to the development of numerous new datasets and benchmarks. However, these resources have predominantly been limited to English, leaving a gap in evaluating commonsense reasoning in other languages. In this paper, we introduce the ArabicSense Benchmark, which is designed to thoroughly evaluate the world-knowledge commonsense reasoning abilities of large language models (LLMs) in Arabic. This benchmark includes three main tasks: first, it tests whether a system can distinguish between natural language statements that make sense and those that do not; second, it requires a system to identify the most crucial reason why a nonsensical statement fails to make sense; and third, it involves generating explanations for why statements do not make sense. We evaluate several Arabic BERT-based models and causal LLMs on these tasks. Experimental results demonstrate improvements after fine-tuning on our dataset. For instance, AraBERT v2 achieved an 87{\%} F1 score on the second task, while Gemma and Mistral-7b achieved F1 scores of 95.5{\%} and 94.8{\%}, respectively. For the generation task, LLaMA-3 achieved the best performance with a BERTScore F1 of 77.3{\%}, closely followed by Mistral-7b at 77.1{\%}. All codes and the benchmark will be made publicly available at https://github.com/. |
hamed-etal-2025-lahjawi | 2,025 | Lahjawi: Arabic Cross-Dialect Translator | In this paper, we explore the rich diversity of Arabic dialects by introducing a suite of pioneering models called Lahjawi. The primary model, Lahjawi-D2D, is the first designed for cross-dialect translation among 15 Arabic dialects. Furthermore, we introduce Lahjawi-D2MSA, a model designed to convert any Arabic dialect into Modern Standard Arabic (MSA). Both models are fine-tuned versions of Kuwain-1.5B an in-house built small language model, tailored for Arabic linguistic characteristics. We provide a detailed overview of Lahjawi`s architecture and training methods, along with a comprehensive evaluation of its performance. The results demonstrate Lahjawi`s success in preserving meaning and style, with BLEU scores of 9.62 for dialect-to-MSA and 9.88 for dialect-to- dialect tasks. Additionally, human evaluation reveals an accuracy score of 58{\%} and a fluency score of 78{\%}, underscoring Lahjawi`s robust handling of diverse dialectal nuances. This research sets a foundation for future advancements in Arabic NLP and cross-dialect communication technologies. |
bezancon-etal-2025-lost | 2,025 | Lost in Variation: An Unsupervised Methodology for Mining Lexico-syntactic Patterns in Middle Arabic Texts | While MSA and some dialects of Arabic have been extensively studied in NLP, Middle Arabic is still very much unknown to the field. However, Middle Arabic holds issues that are still not covered: it is characterized by variation since it mixes standard features, colloquial ones, as well as features that belong to neither of the two. Here, we introduce a methodology to identify, extract and rank variations of 13 manually retrieved formulas. Those formulas come from the nine first booklets of S ̄IRAT AL-MALIK AL-Z. ̄AHIR BAYBAR S., a corpus of Damascene popular literature written in Middle Arabic and composed of 53,843 sentences. In total, we ranked 20, sequences according to their similarity with the original formulas on multiple linguistic layers. We noticed that the variations in these formulas occur in a lexical, morphological and graphical level, but in opposition, the semantic and syntactic levels remain strictly invariable. |
alahmari-2025-sadslyc | 2,025 | SADSLyC: A Corpus for Saudi Arabian Multi-dialect Identification through Song Lyrics | This paper presents the Saudi Arabian Dialects Song Lyrics Corpus (SADSLyC), the first dataset featuring song lyrics from the five major Saudi dialects: Najdi (Central Region), Hijazi (Western Region), Shamali (Northern Region), Janoubi (Southern Region), and Shargawi (Eastern Region). The dataset consists of 31,358 sentences, with each sentence representing a self-contained verse in a song, totaling 151,841 words. Additionally, we present a baseline experiment using the SaudiBERT model to classify the fine-grained dialects in the SADSLyC Corpus. The model achieved an overall accuracy of 73{\%} on the test dataset. |
hossain-etal-2025-enhancing | 2,025 | Enhancing Dialectal Arabic Intent Detection through Cross-Dialect Multilingual Input Augmentation | Addressing the challenges of Arabic intent detection amid extensive dialectal variation, this study presents a crossdialtectal, multilingual approach for classifying intents in banking and migration contexts. By augmenting dialectal inputs with Modern Standard Arabic (MSA) and English translations, our method leverages cross-lingual context to improve classification accuracy. We evaluate single-input (dialect-only), dual-input (dialect + MSA), and triple-input (dialect + MSA + English) models, applying language-specific tokenization for each. Results demonstrate that, in the migration dataset, our model achieved an accuracy gain of over 50{\%} on Tunisian dialect, increasing from 43.3{\%} with dialect-only input to 94{\%} with the full multilingual setup. Similarly, in the PAL (Palestinian dialect) dataset, accuracy improved from 87.7{\%} to 93.5{\%} with translation augmentation, reflecting a gain of 5.8 percentage points. These findings underscore the effectiveness of our approach for intent detection across various Arabic dialects. |
khered-etal-2025-dial2msa | 2,025 | Dial2MSA-Verified: A Multi-Dialect Arabic Social Media Dataset for Neural Machine Translation to Modern Standard Arabic | Social media has become an essential focus for Natural Language Processing (NLP) research due to its widespread use and unique linguistic characteristics. Normalising social media content, especially for morphologically rich languages like Arabic, remains a complex task due to limited parallel corpora. Arabic encompasses Modern Standard Arabic (MSA) and various regional dialects, collectively termed Dialectal Arabic (DA), which complicates NLP efforts due to their informal nature and variability. This paper presents Dial2MSA-Verified, an extension of the Dial2MSA dataset that includes verified translations for Gulf, Egyptian, Levantine, and Maghrebi dialects. We evaluate the performance of Seq2Seq models on this dataset, highlighting the effectiveness of state-of-the-art models in translating local Arabic dialects. We also provide insights through error analysis and outline future directions for enhancing Seq2Seq models and dataset development. The Dial2MSA-Verified dataset is publicly available to support further research. |
el-ghawi-2025-web | 2,025 | Web-Based Corpus Compilation of the Emirati Arabic Dialect | This paper displays some initial efforts conducted in the compilation pursuits of Arabic dialectal corpora in the form of raw text, the end purpose of which is to fine-tune existing Arabic large language models (LLM) to better understand and generate text in the Emirati dialect as instructed. The focus of the paper is on the process of compiling corpora from the web, which includes the exploration of possible methods, tools and techniques specific to web search, as well as examples of genres and domains to explore. The results of these efforts and the importance of native speaker contributions to corpus compilation for low-resource languages are also touched upon. |
al-laith-kebdani-2025-evaluating | 2,025 | Evaluating Calibration of Arabic Pre-trained Language Models on Dialectal Text | While pre-trained language models have made significant progress in different classification tasks, little attention has been given to the reliability of their confidence scores. Calibration, how well model confidence aligns with actual accuracy, is essential for real-world applications where decisions rely on probabilistic outputs. This study addresses this gap in Arabic dialect identification by assessing the calibration of eight pre-trained language models, ensuring their predictions are not only accurate but also reliable for practical applications. We analyze two datasets: one with over 1 million text samples and the Nuanced Arabic Dialect Identification dataset(NADI-2023). Using Expected Calibration Error (ECE) as a metric, we reveal substantial variation in model calibration across dialects in both datasets, showing that prediction confidence can vary significantly depending on regional data. This research has implications for improving the reliability of Arabic dialect models in applications like sentiment analysis and social media monitoring. |
aftiss-etal-2025-empirical | 2,025 | Empirical Evaluation of Pre-trained Language Models for Summarizing Moroccan Darija News Articles | Moroccan Dialect (MD), or {\textquotedblleft}Darija,{\textquotedblright} is a primary spoken variant of Arabic in Morocco, yet remains underrepresented in Natural Language Processing (NLP) research, particularly in tasks like summarization. Despite a growing volume of MD textual data online, there is a lack of robust resources and NLP models tailored to handle the unique linguistic challenges posed by MD. In response, we introduce .MA{\_}v2, an expanded version of the GOUD.MA dataset, containing over 50k articles with their titles across 11 categories. This dataset provides a more comprehensive resource for developing summarization models. We evaluate the application of large language models (LLMs) for MD summarization, utilizing both fine-tuning and zero-shot prompting with encoder-decoder and causal LLMs, respectively. Our findings demonstrate that an expanded dataset improves summarization performance and highlights the capabilities of recent LLMs in handling MD text. We open-source our dataset, fine-tuned models, and all experimental code, establishing a foundation for future advancements in MD NLP. We release the code at https://github.com/AzzedineAftiss/Moroccan-Dialect-Summarization. |
chafik-etal-2025-dialect2sql | 2,025 | Dialect2SQL: A Novel Text-to-SQL Dataset for Arabic Dialects with a Focus on Moroccan Darija | The task of converting natural language questions into executable SQL queries, known as text-to-SQL, has gained significant interest in recent years, as it enables non-technical users to interact with relational databases. Many benchmarks, such as SPIDER and WikiSQL, have contributed to the development of new models and the evaluation of their performance. In addition, other datasets, like SEDE and BIRD, have introduced more challenges and complexities to better map real-world scenarios. However, these datasets primarily focus on high-resource languages such as English and Chinese. In this work, we introduce Dialect2SQL, the first large-scale, cross-domain text-to-SQL dataset in an Arabic dialect. It consists of 9,428 NLQ-SQL pairs across 69 databases in various domains. Along with SQL-related challenges such as long schemas, dirty values, and complex queries, our dataset also incorporates the complexities of the Moroccan dialect, which is known for its diverse source lan-guages, numerous borrowed words, and unique expressions. This demonstrates that our dataset will be a valuable contribution to both the text-to-SQL community and the development of resources for low-resource languages. |
bouomar-abbas-2025-arasim | 2,025 | AraSim: Optimizing Arabic Dialect Translation in Children`s Literature with LLMs and Similarity Scores | The goal of the paper is to address the linguistic gap faced by young Egyptian Arabic speakers through translating children stories from Modern Standard Arabic to the Egyptian Cairo dialect. Claude is used for initial translation, and a fine-tuned AraT5 model is used for backtranslation. The translation quality is assessed using semantic similarity and BLUE scores to compare the original texts and the translations. The resulting corpus contains 130 stories which were revised by native Egyptian speakers who are professional translators. The strengths of this paper are multiple: working on a less-resourced variety, addressing an important social issue, creating a dataset with potential real-life applications, and ensuring the quality of the produced dataset through human validation. |
haj-ahmed-etal-2025-navigating | 2,025 | Navigating Dialectal Bias and Ethical Complexities in Levantine Arabic Hate Speech Detection | Social media platforms have become central to global communication, yet they also facilitate the spread of hate speech. For underrepresented dialects like Levantine Arabic, detecting hate speech presents unique cultural, ethical, and linguistic challenges. This paper explores the complex sociopolitical and linguistic landscape of Levantine Arabic and critically examines the limitations of current datasets used in hate speech detection. We highlight the scarcity of publicly available, diverse datasets and analyze the consequences of dialectal bias within existing resources. By emphasizing the need for culturally and contextually informed natural language processing (NLP) tools, we advocate for a more nuanced and inclusive approach to hate speech detection in the Arab world. |
scherrer-etal-2025-findings | 2,025 | Findings of the VarDial Evaluation Campaign 2025: The NorSID Shared Task on Norwegian Slot, Intent and Dialect Identification | The VarDial Evaluation Campaign 2025 was organized as part of the twelfth workshop on Natural Language Processing for Similar Languages, Varieties and Dialects (VarDial), colocated with COLING 2025. It consisted of one shared task with three subtasks: intent detection, slot filling and dialect identification for Norwegian dialects. This report presents the results of this shared task. Four participating teams have submitted systems with very high performance ({\ensuremath{>}} 97{\%} accuracy) for intent detection, whereas slot detection and dialect identification showed to be much more challenging, with respectively span-F1 scores up to 89{\%}, and weighted dialect F1 scores of 84{\%}. |
alves-2025-information | 2,025 | Information Theory and Linguistic Variation: A Study of Brazilian and European Portuguese | We present a general analysis of the lexical and grammatical differences between Brazilian and European Portuguese by applying entropy measures, including Kullback-Leibler divergence and word order entropy, across various linguistic levels. Using a parallel corpus of BP and EP sentences translated from English, we quantified these differences and identified characteristic phenomena underlying the divergences between the two varieties. The highest divergence was observed at the lexical level due to word pairs unique to each variety but also related to grammatical distinctions. Furthermore, the analysis of parts-of-speech (POS), dependency relations, and POS tri-grams provided information concerning distinctive grammatical constructions. Finally, the word order entropy analysis revealed that while most of the syntactic features analysed showed similar patterns across BP and EP, specific word order preferences were still apparent. |
ng-markov-2025-leveraging | 2,025 | Leveraging Open-Source Large Language Models for Native Language Identification | Native Language Identification (NLI) {--} the task of identifying the native language (L1) of a person based on their writing in the second language (L2) {--} has applications in forensics, marketing, and second language acquisition. Historically, conventional machine learning approaches that heavily rely on extensive feature engineering have outperformed transformer-based language models on this task. Recently, closed-source generative large language models (LLMs), e.g., GPT-4, have demonstrated remarkable performance on NLI in a zero-shot setting, including promising results in open-set classification. However, closed-source LLMs have many disadvantages, such as high costs and undisclosed nature of training data. This study explores the potential of using open-source LLMs for NLI. Our results indicate that open-source LLMs do not reach the accuracy levels of closed-source LLMs when used out-of-the-box. However, when fine-tuned on labeled training data, open-source LLMs can achieve performance comparable to that of commercial LLMs. |
torgbi-etal-2025-adapting | 2,025 | Adapting Whisper for Regional Dialects: Enhancing Public Services for Vulnerable Populations in the United Kingdom | We collect novel data in the public service domain to evaluate the capability of the state-of-the-art automatic speech recognition (ASR) models in capturing regional differences in accents in the United Kingdom (UK), specifically focusing on two accents from Scotland with distinct dialects. This study addresses real-world problems where biased ASR models can lead to miscommunication in public services, disadvantaging individuals with regional accents particularly those in vulnerable populations. We first examine the out-of-the-box performance of the Whisper large-v3 model on a baseline dataset and our data. We then explore the impact of fine-tuning Whisper on the performance in the two UK regions and investigate the effectiveness of existing model evaluation techniques for our real-world application through manual inspection of model errors. We observe that the Whisper model has a higher word error rate (WER) on our test datasets compared to the baseline data and fine-tuning on a given data improves performance on the test dataset with the same domain and accent. The fine-tuned models also appear to show improved performance when applied to the test data outside of the region it was trained on suggesting that fine-tuned models may be transferable within parts of the UK. Our manual analysis of model outputs reveals the benefits and drawbacks of using WER as an evaluation metric and fine-tuning to adapt to regional dialects. |
alam-anastasopoulos-2025-large | 2,025 | Large Language Models as a Normalizer for Transliteration and Dialectal Translation | NLP models trained on standardized language data often struggle with variations. We assess various Large Language Models (LLMs) for transliteration and dialectal normalization. Tuning open-source LLMs with as little as 10,000 parallel examples using LoRA can achieve results comparable to or better than closed-source LLMs. We perform dialectal normalization experiments for twelve South Asian languages and dialectal translation experiments for six language continua worldwide. The dialectal normalization task can also be a preliminary step for the downstream dialectal translation task. Among the six languages used in dialectal translation, our approach enables Italian and Swiss German to surpass the baseline model by 21.5 and 25.8 BLEU points, respectively. |
faisal-anastasopoulos-2025-testing | 2,025 | Testing the Boundaries of LLMs: Dialectal and Language-Variety Tasks | This study evaluates the performance of large language models (LLMs) on benchmark datasets designed for dialect-specific NLP tasks. Dialectal NLP is a low-resource field, yet it is crucial for evaluating the robustness of language models against linguistic diversity. This work is the first to systematically compare state-of-the-art instruction-tuned LLMs{---}both open-weight multilingual and closed-weight generative models{---}with encoder-based models that rely on supervised task-specific fine-tuning for dialectal tasks. We conduct extensive empirical analyses to provide insights into the current LLM landscape for dialect-focused tasks. Our findings indicate that certain tasks, such as dialect identification, are challenging for LLMs to replicate effectively due to the complexity of multi-class setups and the suitability of these tasks for supervised fine-tuning. Additionally, the structure of task labels{---}whether categorical or continuous scoring{---}significantly affects model performance. While LLMs excel in tasks like machine reading comprehension, their instruction-following ability declines in simpler tasks like POS tagging when task instructions are inherently complex. Overall, subtle variations in prompt design can greatly impact performance, underscoring the need for careful prompt engineering in dialectal evaluations. |
plum-etal-2025-text | 2,025 | Text Generation Models for Luxembourgish with Limited Data: A Balanced Multilingual Strategy | This paper addresses the challenges in developing language models for less-represented languages, with a focus on Luxembourgish. Despite its active development, Luxembourgish faces a digital data scarcity, exacerbated by Luxembourg`s multilingual context. We propose a novel text generation model based on the T5 architecture, combining limited Luxembourgish data with equal amounts, in terms of size and type, of German and French data. We hypothesise that a model trained on Luxembourgish, German, and French will improve the model`s cross-lingual transfer learning capabilities and outperform monolingual and large multilingual models. To verify this, the study at hand explores whether multilingual or monolingual training is more beneficial for Luxembourgish language generation. For the evaluation, we introduce LuxGen, a text generation benchmark that is the first of its kind for Luxembourgish. |
lendvai-etal-2025-retrieval | 2,025 | Retrieval of Parallelizable Texts Across Church Slavic Variants | The goal of our study is to identify parallelizable texts for Church Slavic, across chronological and regional variants. Next to using a benchmark text, we utilize a recently digitized, large text collection and compile new resources for the retrieval of similar texts: a ground truth dataset holding a small amount of manually aligned sentences in Old Church Slavic and in Old East Slavic, and a large unaligned dataset that has a subset of ground truth (GT) quality texts but contains noise from handwritten text recognition (HTR) for the majority of the collection. We discuss preprocessing challenges in the data and the impact of sentence segmentation on retrieval performance. We evaluate sentence snippets mapped across these two diachronic variants of Church Slavic, expressed by mean reciprocal rank, using embedding representations from large language models (LLMs) as well as classical string similarity based approaches combined with k-nearest neighbor (kNN) search. Experimental results indicate that in the current setup (short text snippets, off-the-shelf multilingual embeddings), classical string similarity based retrieval can still outperform embedding based retrieval. |
lutgen-etal-2025-neural | 2,025 | Neural Text Normalization for Luxembourgish Using Real-Life Variation Data | Orthographic variation is very common in Luxembourgish texts due to the absence of a fully-fledged standard variety. Additionally, developing NLP tools for Luxembourgish is a difficult task given the lack of annotated and parallel data, which is exacerbated by ongoing standardization. In this paper, we propose the first sequence-to-sequence normalization models using the ByT5 and mT5 architectures with training data obtained from word-level real-life variation data. We perform a fine-grained, linguistically-motivated evaluation to test byte-based, word-based and pipeline-based models for their strengths and weaknesses in text normalization. We show that our sequence model using real-life variation data is an effective approach for tailor-made normalization in Luxembourgish. |
kruckl-etal-2025-improving | 2,025 | Improving Dialectal Slot and Intent Detection with Auxiliary Tasks: A Multi-Dialectal Bavarian Case Study | Reliable slot and intent detection (SID) is crucial in natural language understanding for applications like digital assistants. Encoder-only transformer models fine-tuned on high-resource languages generally perform well on SID. However, they struggle with dialectal data, where no standardized form exists and training data is scarce and costly to produce. We explore zero-shot transfer learning for SID, focusing on multiple Bavarian dialects, for which we release a new dataset for the Munich dialect. We evaluate models trained on auxiliary tasks in Bavarian, and compare joint multi-task learning with intermediate-task training. We also compare three types of auxiliary tasks: token-level syntactic tasks, named entity recognition (NER), and language modelling. We find that the included auxiliary tasks have a more positive effect on slot filling than intent classification (with NER having the most positive effect), and that intermediate-task training yields more consistent performance gains. Our best-performing approach improves intent classification performance on Bavarian dialects by 5.1 and slot filling F1 by 8.4 percentage points. |
coats-etal-2025-regional | 2,025 | Regional Distribution of the /el/-/\ael/ Merger in Australian English | Prelateral merger of /e/ and /{\ae}/ is a salient acoustic feature of speech from Melbourne and the state of Victoria in Australia, but little is known about its presence in other parts of the country. In this study, automated methods of data collection, forced alignment, and formant extraction are used to analyze the regional distribution of the vowel merger within all of Australia, in 4.3 million vowel tokens from naturalistic speech in 252 locations. The extent of the merger is quantified using the difference in Bhattacharyya`s distance scores based on phonetic context, and the regional distribution is assessed using spatial autocorrelation. The principal findings are that the merger is most prominent in Victoria and least prominent in Sydney and New South Wales. We also find preliminary indications that it may be present in other parts of the country. |
khalifa-etal-2025-learning | 2,025 | Learning Cross-Dialectal Morphophonology with Syllable Structure Constraints | We investigate learning surface forms from underlying morphological forms for low-resource language varieties. We concentrate on learning explicit rules with the aid of learned syllable structure constraints, which outperforms neural methods on this small data task and provides interpretable output. Evaluating across one relatively high-resource and two related low-resource Arabic dialects, we find that a model trained only on the high-resource dialect achieves decent performance on the low-resource dialects, useful when no low-resource training data is available. The best results are obtained when our system is trained only on the low-resource dialect data without augmentation from the related higher-resource dialect. We discuss the impact of syllable structure constraints and the strengths and weaknesses of data augmentation and transfer learning from a related dialect. |
lopetegui-etal-2025-common | 2,025 | Common Ground, Diverse Roots: The Difficulty of Classifying Common Examples in Spanish Varieties | Variations in languages across geographic regions or cultures are crucial to address to avoid biases in NLP systems designed for culturally sensitive tasks, such as hate speech detection or dialog with conversational agents. In languages such as Spanish, where varieties can significantly overlap, many examples can be valid across them, which we refer to as common examples. Ignoring these examples may cause misclassifications, reducing model accuracy and fairness. Therefore, accounting for these common examples is essential to improve the robustness and representativeness of NLP systems trained on such data. In this work, we address this problem in the context of Spanish varieties. We use training dynamics to automatically detect common examples or errors in existing Spanish datasets. We demonstrate the efficacy of using predicted label confidence for our Datamaps (CITATION) implementation for the identification of hard-to-classify examples, especially common examples, enhancing model performance in variety identification tasks. Additionally, we introduce a Cuban Spanish Variety Identification dataset with common examples annotations developed to facilitate more accurate detection of Cuban and Caribbean Spanish varieties. To our knowledge, this is the first dataset focused on identifying the Cuban, or any other Caribbean, Spanish variety. |
blaschke-etal-2025-add | 2,025 | Add Noise, Tasks, or Layers? MaiNLP at the VarDial 2025 Shared Task on Norwegian Dialectal Slot and Intent Detection | Slot and intent detection (SID) is a classic natural language understanding task. Despite this, research has only more recently begun focusing on SID for dialectal and colloquial varieties. Many approaches for low-resource scenarios have not yet been applied to dialectal SID data, or compared to each other on the same datasets. We participate in the VarDial 2025 shared task on slot and intent detection in Norwegian varieties, and compare multiple set-ups: varying the training data (English, Norwegian, or dialectal Norwegian), injecting character-level noise, training on auxiliary tasks, and applying Layer Swapping, a technique in which layers of models fine-tuned on different datasets are assembled into a model. We find noise injection to be beneficial while the effects of auxiliary tasks are mixed. Though some experimentation was required to successfully assemble a model from layers, it worked surprisingly well; a combination of models trained on English and small amounts of dialectal data produced the most robust slot predictions. Our best models achieve 97.6{\%} intent accuracy and 85.6{\%} slot F1 in the shared task. |
midtgaard-etal-2025-ltg | 2,025 | LTG at VarDial 2025 NorSID: More and Better Training Data for Slot and Intent Detection | This paper describes the LTG submission to the VarDial 2025 shared task, where we participate in the Norwegian slot and intent detection subtasks. The shared task focuses on Norwegian dialects, which present challenges due to their low-resource nature and variation. We test a variety of neural models and training data configurations, with the focus on improving and extending the available Norwegian training data. This includes automatically re-aligning slot spans in Norwegian Bokm{\r{a}}l, as well as re-translating the original English training data into both Bokm{\r{a}}l and Nynorsk. {\%} to address dialectal diversity. We also re-annotate an external Norwegian dataset to augment the training data. Our best models achieve first place in both subtasks, achieving an span F1 score of 0.893 for slot filling and an accuracy of 0.980 for intent detection. Our results indicate that while translation quality is less critical, improving the slot labels has a notable impact on slot performance. Moreover, adding more standard Norwegian data improves performance, but incorporating even small amounts of dialectal data leads to greater gains. |
bengoetxea-etal-2025-hitz | 2,025 | HiTZ at VarDial 2025 NorSID: Overcoming Data Scarcity with Language Transfer and Automatic Data Annotation | In this paper we present our submission for the NorSID Shared Task as part of the 2025 VarDial Workshop, consisting of three tasks: Intent Detection, Slot Filling and Dialect Identification, evaluated using data in different dialects of the Norwegian language. For Intent Detection and Slot Filling, we have fine-tuned a multitask model in a cross-lingual setting, to leverage the xSID dataset available in 17 languages. In the case of Dialect Identification, our final submission consists of a model fine-tuned on the provided development set, which has obtained the highest scores within our experiments. Our final results on the test set show that our models do not drop in performance compared to the development set, likely due to the domain-specificity of the dataset and the similar distribution of both subsets. Finally, we also report an in-depth analysis of the provided datasets and their artifacts, as well as other sets of experiments that have been carried out but did not yield the best results. Additionally, we present an analysis on the reasons why some methods have been more successful than others; mainly the impact of the combination of languages and domain-specificity of the training data on the results. |
ibrahim-2025-cufe | 2,025 | CUFE@VarDial 2025 NorSID: Multilingual BERT for Norwegian Dialect Identification and Intent Detection | Dialect identification is crucial in enhancing various tasks, including sentiment analysis, as a speaker`s geographical origin can significantly affect their perspective on a topic, also, intent detection has gained significant traction in natural language processing due to its applications in various domains, including virtual assistants, customer service automation, and information retrieval systems. This work describes a system developed for VarDial 2025: Norwegian slot and intent detection and dialect identification shared task (Scherrer et al., 2025), a challenge designed to address the dialect recognition and intent detection problems for a low-resource language like Norwegian. More specifically, this work investigates the performance of different BERT models in solving this problem. Finally, the output of the multilingual version of the BERT model was submitted to this shared task, the developed system achieved a weighted F1 score of 79.64 for dialect identification and an accuracy of 94.38 for intent detection. |
pavlich-etal-2025-beyond | 2,025 | Beyond Text-to-SQL for IoT Defense: A Comprehensive Framework for Querying and Classifying IoT Threats | Recognizing the promise of natural language interfaces to databases, prior studies have emphasized the development of text-to-SQL systems. Existing research has generally focused on generating SQL statements from text queries, and the broader challenge lies in inferring new information about the returned data. Our research makes two major contributions to address this gap. First, we introduce a novel Internet-of-Things (IoT) text-to-SQL dataset comprising 10,985 text-SQL pairs and 239,398 rows of network traffic activity. The dataset contains additional query types limited in prior text-to-SQL datasets, notably, temporal-related queries. Our dataset is sourced from a smart building`s IoT ecosystem exploring sensor read and network traffic data. Second, our dataset allows two-stage processing, where the returned data (network traffic) from a generated SQL can be categorized as malicious or not. Our results show that joint training to query and infer information about the data improves overall text-to-SQL performance, nearly matching that of substantially larger models. We also show that current large language models (e.g., GPT3.5) struggle to infer new information about returned data (i.e., they are bad at tabular data understanding), thus our dataset provides a novel test bed for integrating complex domain-specific reasoning into LLMs. |
cheng-etal-2025-gibberish | 2,025 | Gibberish is All You Need for Membership Inference Detection in Contrastive Language-Audio Pretraining | Audio can disclose PII, particularly when combined with related text data. Therefore, it is essential to develop tools to detect privacy leakage in Contrastive Language-Audio Pretraining(CLAP). Existing MIAs need audio as input, risking exposure of voiceprint and requiring costly shadow models. We first propose PRMID, a membership inference detector based probability ranking given by CLAP, which does not require training shadow models but still requires both audio and text of the individual as input. To address these limitations, we then propose USMID, a textual unimodal speaker-level membership inference detector, querying the target model using only text data. We randomly generate textual gibberish that are clearly not in training dataset. Then we extract feature vectors from these texts using the CLAP model and train a set of anomaly detectors on them. During inference, the feature vector of each test text is input into the anomaly detector to determine if the speaker is in the training set (anomalous) or not (normal). If available, USMID can further enhance detection by integrating real audio of the tested speaker. Extensive experiments on various CLAP model architectures and datasets demonstrate that USMID outperforms baseline methods using only text data. |
cheng-etal-2025-pbi | 2,025 | PBI-Attack: Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for Toxicity Maximization | Understanding the vulnerabilities of Large Vision Language Models (LVLMs) to jailbreak attacks is essential for their responsible real-world deployment. Most previous work requires access to model gradients, or is based on human knowledge (prompt engineering) to complete jailbreak, and they hardly consider the interaction of images and text, resulting in inability to jailbreak in black box scenarios or poor performance. To overcome these limitations, we propose a Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for toxicity maximization, referred to as PBI-Attack. Our method begins by extracting malicious features from a harmful corpus using an alternative LVLM and embedding these features into a benign image as prior information. Subsequently, we enhance these features through bidirectional cross-modal interaction optimization, which iteratively optimizes the bimodal perturbations in an alternating manner through greedy search, aiming to maximize the toxicity of the generated response. The toxicity level is quantified using a well-trained evaluation model.Experiments demonstrate that PBI-Attack outperforms previous state-of-the-art jailbreak methods, achieving an average attack success rate of 92.5{\%} across three open-source LVLMs and around 67.3{\%} on three closed-source LVLMs.redDisclaimer: This paper contains potentially disturbing and offensive content. |
shi-etal-2025-ambiguity | 2,025 | Ambiguity Detection and Uncertainty Calibration for Question Answering with Large Language Models | Large Language Models (LLMs) have demonstrated excellent capabilities in Question Answering (QA) tasks, yet their ability to identify and address ambiguous questions remains underdeveloped. Ambiguities in user queries often lead to inaccurate or misleading answers, undermining user trust in these systems. Despite prior attempts using prompt-based methods, performance has largely been equivalent to random guessing, leaving a significant gap in effective ambiguity detection. To address this, we propose a novel framework for detecting ambiguous questions within LLM-based QA systems. We first prompt an LLM to generate multiple answers to a question, and then analyze them to infer the ambiguity. We propose to use a lightweight Random Forest model, trained on a bootstrapped and shuffled 6-shot examples dataset. Experimental results on ASQA, PACIFIC, and ABG-COQA datasets demonstrate the effectiveness of our approach, with accuracy up to 70.8{\%}. Furthermore, our framework enhances the confidence calibration of LLM outputs, leading to more trustworthy QA systems able to handle complex questions. |
liu-etal-2025-smaller | 2,025 | Smaller Large Language Models Can Do Moral Self-Correction | Self-correction is one of the most amazing emerging capabilities of Large Language Models (LLMs), enabling LLMs to self-modify an inappropriate output given a natural language feedback which describes the problems of that output. Moral self-correction is a post-hoc approach correcting unethical generations without requiring a gradient update, making it both computationally lightweight and capable of preserving the language modeling ability. Previous works have shown that LLMs can self-debias, and it has been reported that small models, i.e., those with less than 22B parameters, are \textit{not} capable of moral self-correction.However, there is no direct proof as to why such smaller models fall short of moral self-correction, though previous research hypothesizes that larger models are skilled in following instructions and understanding abstract social norms.In this paper, we empirically validate this hypothesis in the context of social stereotyping, through meticulous prompting.Our experimental results indicate that \textbf{(i)} surprisingly, 3.8B LLMs with proper safety alignment fine-tuning can achieve very good moral self-correction performance, highlighting the significant effects of safety alignment; and \textbf{(ii)} small LLMs are indeed weaker than larger-scale models in terms of comprehending social norms and self-explanation through CoT, but all scales of LLMs show bad self-correction performance given unethical instructions. |
bonnier-2025-error | 2,025 | Error Detection for Multimodal Classification | Machine learning models have proven to be useful in various key applications such as autonomous driving or diagnosis prediction. When a model is implemented under real-world conditions, it is thus essential to detect potential errors with a trustworthy approach. This monitoring practice will render decision-making safer by avoiding catastrophic failures. In this paper, the focus is on multimodal classification. We introduce a method that addresses error detection based on unlabeled data. It leverages fused representations and computes the probability that a model will fail based on detected fault patterns in validation data. To improve transparency, we employ a sampling-based approximation of Shapley values in multimodal settings in order to explain why a prediction is assessed as erroneous in terms of feature values. Further, as explanation methods can sometimes disagree, we suggest evaluating the consistency of explanations produced by different value functions and algorithms. To show the relevance of our method, we measure it against a selection of 9 baselines from various domains on tabular-text and text-image datasets, and 2 multimodal fusion strategies for the classification models. Lastly, we show the usefulness of our explanation algorithm on misclassified samples. |
kim-cho-2025-break | 2,025 | Break the Breakout: Reinventing LM Defense Against Jailbreak Attacks with Self-Refine | Language models (LMs) are vulnerable to exploitation for adversarial misuse. Training LMs for safety alignment is extensive, making it hard to respond to fast-developing attacks immediately, such as jailbreaks. We propose self-refine with formatting that achieves outstanding safety even in non-safety-aligned LMsand evaluate our method alongside several defense baselines, demonstrating that it is the safest training-free method against jailbreak attacks.Additionally, we proposed a formatting method that improves the efficiency of the self-refine process while reducing attack success rates in fewer iterations. We observed that non-safety-aligned LMs outperform safety-aligned LMs in safety tasks by giving more helpful and safe responses.In conclusion, our findings can achieve less safety risk with fewer computational costs, allowing non-safety LM to be efficiently utilized in real-world service. |
li-etal-2025-minimal | 2,025 | Minimal Evidence Group Identification for Claim Verification | When verifying a claim in real-world settings, e.g. against a large collection of candidate evidence text retrieved from the web, a model is typically expected to identify and aggregate a complete set of evidence pieces that collectively provide full support to a claim.The problem becomes particularly challenging as there might exist different sets of evidence that could be used to verify the claim from different perspectives. In this paper, we formally define and study the problem of identifying such minimal evidence groups (MEGs) for fact verification. We show that MEG identification can be reduced to a Set Cover-like problem, based on an entailment model which estimates whether a given evidence group provides full or partial support to a claim. Our proposed approach achieves 18.4{\%} {\&} 34.8{\%} absolute improvements on WiCE and SciFact datasets over LLM prompting. Finally, we demonstrate the downstream benefit of MEGs in applications such as claim generation. |
wei-etal-2025-cracking | 2,025 | Cracking the Code: Enhancing Implicit Hate Speech Detection through Coding Classification | The internet has become a hotspot for hate speech (HS), threatening societal harmony and individual well-being. While automatic detection methods perform well in identifying explicit hate speech (ex-HS), they struggle with more subtle forms, such as implicit hate speech (im-HS). We tackle this problem by introducing a new taxonomy for im-HS detection, defining six encoding strategies named *codetypes*. We present two methods for integrating codetypes into im-HS detection: 1) prompting large language models (LLMs) directly to classify sentences based on generated responses, and 2) using LLMs as encoders with codetypes embedded during the encoding process. Experiments show that the use of codetypes improves im-HS detection in both Chinese and English datasets, validating the effectiveness of our approach across different languages. |
kale-vrn-2025-line | 2,025 | Line of Duty: Evaluating LLM Self-Knowledge via Consistency in Feasibility Boundaries | As LLMs grow more powerful, their most profound achievement may be recognising when to say {\textquotedblleft}I don`t know{\textquotedblright}. Existing studies on LLM self-knowledge have been largely constrained by human-defined notions of feasibility, often neglecting the reasons behind unanswerability by LLMs and failing to study deficient types of self-knowledge. This study aims to obtain intrinsic insights into different types of LLM self-knowledge with a novel methodology: allowing them the flexibility to set their own feasibility boundaries and then analysing the consistency of these limits. We find that even frontier models like GPT-4o and Mistral Large are not sure of their own capabilities more than 80{\%} of the time, highlighting a significant lack of trustworthiness in responses. Our analysis of confidence balance in LLMs indicates that models swing between overconfidence and conservatism in feasibility boundaries depending on task categories and that the most significant self-knowledge weaknesses lie in temporal awareness and contextual understanding. These difficulties in contextual comprehension additionally lead models to question their operational boundaries, resulting in considerable confusion within the self-knowledge of LLMs. We make our code and results available publicly. |
singhania-etal-2025-multi | 2,025 | Multi-lingual Multi-turn Automated Red Teaming for LLMs | Language Model Models (LLMs) have improved dramatically in the past few years, increasing their adoption and the scope of their capabilities over time. A significant amount of work is dedicated to {\textquotedblleft}model alignment{\textquotedblright}, i.e., preventing LLMs to generate unsafe responses when deployed into customer-facing applications. One popular method to evaluate safety risks is red-teaming, where agents attempt to bypass alignment by crafting elaborate prompts that trigger unsafe responses from a model. Standard human-driven red-teaming is costly, time-consuming and rarely covers all the recent features (e.g., multi-lingual, multi-modal aspects), while proposed automation methods only cover a small subset of LLMs capabilities (i.e., English or single-turn). We present Multi-lingual Multi-turn Automated Red Teaming (MM-ART), a method to fully automate conversational, multi-lingual red-teaming operations and quickly identify prompts leading to unsafe responses. Through extensive experiments on different languages, we show the studied LLMs are on average 71{\%} more vulnerable after a 5-turn conversation in English than after the initial turn. For conversations in non-English languages, models display up to 195{\%} more safety vulnerabilities than the standard single-turn English approach, confirming the need for automated red-teaming methods matching LLMs capabilities. |
krasnodebska-etal-2025-rainbow | 2,025 | Rainbow-Teaming for the Polish Language: A Reproducibility Study | The development of multilingual large language models (LLMs) presents challenges in evaluating their safety across all supported languages. Enhancing safety in one language (e.g., English) may inadvertently introduce vulnerabilities in others. To address this issue, we implement a methodology for the automatic creation of red-teaming datasets for safety evaluation in Polish language. Our approach generates both harmful and non-harmful prompts by sampling different risk categories and attack styles. We test several open-source models, including those trained on Polish data, and evaluate them using metrics such as Attack Success Rate (ASR) and False Reject Rate (FRR). The results reveal clear gaps in safety performance between models and show that better testing across languages is needed. |
xu-etal-2025-biasedit | 2,025 | BiasEdit: Debiasing Stereotyped Language Models via Model Editing | Previous studies have established that language models manifest stereotyped biases. Existing debiasing strategies, such as retraining a model with counterfactual data, representation projection, and prompting often fail to efficiently eliminate bias or directly alter the models' biased internal representations. To address these issues, we propose BiasEdit, an efficient model editing method to remove stereotypical bias from language models through lightweight networks that act as editors to generate parameter updates. BiasEdit employs a *debiasing loss* guiding editor networks to conduct local edits on partial parameters of a language model for debiasing while preserving the language modeling abilities during editing through a *retention loss*. Experiments on StereoSet and Crows-Pairs demonstrate the effectiveness, efficiency, and robustness of BiasEdit in eliminating bias compared to tangental debiasing baselines, and little to no impact on the language models' general capabilities. In addition, we conduct bias tracing to probe bias in various modules and explore bias editing impacts on different components of language models. |
rawte-etal-2025-voters | 2,025 | Do Voters Get the Information They Want? Understanding Authentic Voter FAQs in the US and How to Improve for Informed Electoral Participation | Accurate information is crucial for democracy as it empowers voters to make informed decisions about their representatives and keeping them accountable. In the US, state election commissions (SECs), often required by law, are the primary providers of Frequently Asked Questions (FAQs) to voters, and secondary sources like non-profits such as League of Women Voters (LWV) try to complement their information shortfall. However, surprisingly, to the best of our knowledge, there is neither a single source with comprehensive FAQs nor a study analyzing the data at national level to identify current practices and ways to improve the status quo. This paper addresses it by providing the \textbf{first dataset on Voter FAQs covering all the US states}. Second, we introduce metrics for FAQ information quality (FIQ) with respect to questions, answers, and answers to corresponding questions. Third, we use FIQs to analyze US FAQs to identify leading, mainstream and lagging content practices and corresponding states. Finally, we identify what states across the spectrum can do to improve FAQ quality and thus, the overall information ecosystem. Across all 50 U.S. states, 12{\%} were identified as leaders and 8{\%} as laggards for FIQSvoter, while 14{\%} were leaders and 12{\%} laggards for FIQSdeveloper. The code and sample data are provided at \url{https://anonymous.4open.science/r/election-qa-analysis-BE4E}. |
rawte-etal-2025-vibe | 2,025 | ViBe: A Text-to-Video Benchmark for Evaluating Hallucination in Large Multimodal Models | Recent advances in Large Multimodal Models (LMMs) have expanded their capabilities to video understanding, with Text-to-Video (T2V) models excelling in generating videos from textual prompts. However, they still frequently produce hallucinated content, revealing AI-generated inconsistencies. We introduce ViBe \url{https://huggingface.co/datasets/ViBe-T2V-Bench/ViBe}: a large-scale dataset of hallucinated videos from open-source T2V models. We identify five major hallucination types: Vanishing Subject, Omission Error, Numeric Variability, Subject Dysmorphia, and Visual Incongruity. Using ten T2V models, we generated and manually annotated 3,782 videos from 837 diverse MS COCO captions. Our proposed benchmark includes a dataset of hallucinated videos and a classification framework using video embeddings. ViBe serves as a critical resource for evaluating T2V reliability and advancing hallucination detection. We establish classification as a baseline, with the TimeSFormer + CNN ensemble achieving the best performance (0.345 accuracy, 0.342 F1 score). While initial baselines proposed achieve modest accuracy, this highlights the difficulty of automated hallucination detection and the need for improved methods. Our research aims to drive the development of more robust T2V models and evaluate their outputs based on user preferences. Our code is available at: \url{https://anonymous.4open.science/r/vibe-1840/} |
borszukovszki-etal-2025-know | 2,025 | Know What You do Not Know: Verbalized Uncertainty Estimation Robustness on Corrupted Images in Vision-Language Models | To leverage the full potential of Large Language Models (LLMs) it is crucial to have some information on their answers' uncertainty. This means that the model has to be able to quantify how certain it is in the correctness of a given response. Bad uncertainty estimates can lead to overconfident wrong answers undermining trust in these models. Quite a lot of research has been done on language models that work with text inputs and provide text outputs. Still, since the visual capabilities have been added to these models recently, there has not been much progress on the uncertainty of Visual Language Models (VLMs). We tested three state-of-the-art VLMs on corrupted image data. We found that the severity of the corruption negatively impacted the models' ability to estimate their uncertainty and the models also showed overconfidence in most of the experiments. |
rahman-harris-2025-summary | 2,025 | Summary the Savior: Harmful Keyword and Query-based Summarization for LLM Jailbreak Defense | Large Language Models (LLMs) are widely used for their capabilities, but face threats from jailbreak attacks, which exploit LLMs to generate inappropriate information and bypass their defense system. Existing defenses are often specific to jailbreak attacks and as a result, a robust, attack-independent solution is needed to address both Natural Language Processing (NLP) ambiguities and attack variability. In this study, we have introduced, Summary The Savior, a novel jailbreak detection mechanism leveraging harmful keywords and query-based security-aware summary classification. By analyzing the illegal and improper contents of prompts within the summaries, the proposed method remains robust against attack diversity and NLP ambiguities. Two novel datasets for harmful keyword extraction and security aware summaries utilizing GPT-4 and Llama-3.1 70B respectively have been generated in this regard. Moreover, an {\textquotedblleft}ambiguous harmful{\textquotedblright} class has been introduced to address content and intent ambiguities. Evaluation results demonstrate that, Summary The Savior achieves higher defense performance, outperforming state-of-the-art defense mechanisms namely Perplexity Filtering, SmoothLLM, Erase and Check with lowest attack success rates across various jailbreak attacks namely PAIR, GCG, JBC and Random Search, on Llama-2, Vicuna-13B and GPT-4. Our codes, models, and results are available at: https://github.com/shrestho10/SummaryTheSavior |
yang-etal-2025-bias | 2,025 | Bias A-head? Analyzing Bias in Transformer-Based Language Model Attention Heads | Transformer-based pretrained large language models (PLM) such as BERT and GPT have achieved remarkable success in NLP tasks. However, PLMs are prone to encoding stereotypical biases. Although a burgeoning literature has emerged on stereotypical bias mitigation in PLMs, such as work on debiasing gender and racial stereotyping, how such biases manifest and behave internally within PLMs remains largely unknown. Understanding the internal stereotyping mechanisms may allow better assessment of model fairness and guide the development of effective mitigation strategies. In this work, we focus on attention heads, a major component of the Transformer architecture, and propose a bias analysis framework to explore and identify a small set of biased heads that are found to contribute to a PLM`s stereotypical bias. We conduct extensive experiments to validate the existence of these biased heads and to better understand how they behave. We investigate gender and racial bias in the English language in two types of Transformer-based PLMs: the encoder-based BERT model and the decoder-based autoregressive GPT model, LLaMA-2 (7B), and LLaMA-2-Chat (7B). Overall, the results shed light on understanding the bias behavior in pretrained language models. |
trusca-allein-2025-mimicking | 2,025 | Mimicking How Humans Interpret Out-of-Context Sentences Through Controlled Toxicity Decoding | Interpretations of a single sentence can vary, particularly when its context is lost. This paper aims to simulate how readers perceive content with varying toxicity levels by generating diverse interpretations of out-of-context sentences. By modeling toxicity we can anticipate misunderstandings and reveal hidden toxic meanings. Our proposed decoding strategy explicitly controls toxicity in the set of generated interpretations by (i) aligning interpretation toxicity with the input, (ii) relaxing toxicity constraints for more toxic input sentences, and (iii) promoting diversity in toxicity levels within the set of generated interpretations. Experimental results show that our method improves alignment with human-written interpretations in both syntax and semantics while reducing model prediction uncertainty. |
rabinovich-anaby-tavor-2025-robustness | 2,025 | On the Robustness of Agentic Function Calling | Large Language Models (LLMs) are increasingly acting as autonomous agents, with function calling (FC) capabilities enabling them to invoke specific tools for tasks. While prior research has primarily focused on improving FC accuracy, little attention has been given to the robustness of these agents to perturbations in their input. We introduce a benchmark assessing FC robustness in two key areas: resilience to naturalistic query variations, and stability in function calling when the toolkit expands with semantically related tools. Evaluating best-performing FC models on a carefully expanded subset of the Berkeley function calling leaderboard (BFCL), we identify critical weaknesses in existing evaluation methodologies, and highlight areas for improvement in real-world agentic deployments. |
cecere-etal-2025-monte | 2,025 | Monte Carlo Temperature: a robust sampling strategy for LLM`s uncertainty quantification methods | Uncertainty quantification (UQ) in Large Language Models (LLMs) is essential for their safe and reliable deployment, particularly in critical applications where incorrect outputs can have serious consequences. Current UQ methods typically rely on querying the model multiple times using non-zero temperature sampling to generate diverse outputs for uncertainty estimation. However, the impact of selecting a given temperature parameter is understudied, and our analysis reveals that temperature plays a fundamental role in the quality of uncertainty estimates. The conventional approach of identifying optimal temperature values requires expensive hyperparameter optimization (HPO) that must be repeated for each new model-dataset combination. We propose Monte Carlo Temperature (MCT), a robust sampling strategy that eliminates the need for temperature calibration. Our analysis reveals that: 1) MCT provides more robust uncertainty estimates across a wide range of temperatures, 2) MCT improves the performance of UQ methods by replacing fixed-temperature strategies that do not rely on HPO, and 3) MCT achieves statistical parity with oracle temperatures, which represent the ideal outcome of a well-tuned but computationally expensive HPO process. These findings demonstrate that effective UQ can be achieved without the computational burden of temperature parameter calibration. |
bhat-etal-2025-know | 2,025 | Know Thyself: Validating Knowledge Awareness of LLM-based Persona Agents | Large Language Models (LLMs) have demonstrated remarkable capability in simulating human behaviors, personality, and language. Such synthetic agents with personalities are considered as cost-effective proxies for real users to facilitate crowd-sourcing efforts like annotations, surveys, and A/B testing. Accordingly, it is imperative to validate knowledge awareness of these LLM persona agents when they are customized for further usage. Currently, there is no established way for such evaluation and appropriate mitigation. In this work, we propose a generic evaluation approach to validate LLM based persona agents for correctness, relevance, and diversity in the context of self-awareness and domain knowledge.We evaluate the efficacy of this framework using three LLMs ( Llama, GPT-4o, and Gemma) for domains such as air travel, gaming, and fitness. We also experiment with advanced prompting strategies such as ReAct and Reflexion. We find that though GPT-4o and Llama demonstrate comparable performance, they fail some of basic consistency checks under certain perturbations. |
purpura-etal-2025-building | 2,025 | Building Safe GenAI Applications: An End-to-End Overview of Red Teaming for Large Language Models | The rapid growth of Large Language Models (LLMs) presents significant privacy, security, and ethical concerns. While much research has proposed methods for defending LLM systems against misuse by malicious actors, researchers have recently complemented these efforts with an offensive approach that involves red teaming, i.e., proactively attacking LLMs with the purpose of identifying their vulnerabilities. This paper provides a concise and practical overview of the LLM red teaming literature, structured so as to describe a multi-component system end-to-end. To motivate red teaming we survey the initial safety needs of some high-profile LLMs, and then dive into the different components of a red teaming system as well as software packages for implementing them. We cover various attack methods, strategies for attack-success evaluation, metrics for assessing experiment outcomes, as well as a host of other considerations. Our survey will be useful for any reader who wants to rapidly obtain a grasp of the major red teaming concepts for their own use in practical applications. |
angelov-etal-2025-difficulty | 2,025 | Difficulty Estimation in Natural Language Tasks with Action Scores | This study investigates the effectiveness of the action score, a metric originally developed for computer vision tasks, in estimating sample difficulty across various natural language processing (NLP) tasks. Using transformer-based models, the action score is applied to sentiment analysis, natural language inference, and abstractive text summarization. The results demonstrate that the action score can effectively identify challenging samples in sentiment analysis and natural language inference, often capturing difficult instances that are missed by more established metrics like entropy. However, the effectiveness of the action score appears to be task-dependent, as evidenced by its performance in the abstractive text summarization task, where it exhibits a nearly linear relationship with entropy. The findings suggest that the action score can provide valuable insights into the characteristics of challenging samples in NLP tasks, particularly in classification settings. However, its application should be carefully considered in the context of each specific task and in light of emerging research on the potential value of hard samples in machine learning. |
sinha-etal-2025-small | 2,025 | Are Small Language Models Ready to Compete with Large Language Models for Practical Applications? | The rapid rise of Language Models (LMs) has expanded their use in several applications. Yet, due to constraints of model size, associated cost, or proprietary restrictions, utilizing state-of-the-art (SOTA) LLMs is not always feasible. With open, smaller LMs emerging, more applications can leverage their capabilities, but selecting the right LM can be challenging as smaller LMs don`t perform well universally. This work tries to bridge this gap by proposing a framework to experimentally evaluate small, open LMs in practical settings through measuring semantic correctness of outputs across three practical aspects: task types, application domains and reasoning types, using diverse prompt styles. It also conducts an in-depth comparison of 10 small, open LMs to identify best LM and prompt style depending on specific application requirement using the proposed framework. We also show that if selected appropriately, they can outperform SOTA LLMs like DeepSeek-v2, GPT-4o-mini, Gemini-1.5-Pro, and even compete with GPT-4o. |
bodhwani-etal-2025-calibrated | 2,025 | A Calibrated Reflection Approach for Enhancing Confidence Estimation in LLMs | A critical challenge in deploying Large Language Models (LLMs) is developing reliable mechanisms to estimate their confidence, enabling systems to determine when to trust model outputs and when to seek human intervention. In this paper, we present a Calibrated Reflection Approach for Enhancing Confidence Estimation in LLMs, a framework that combines structured reasoning with distance-aware calibration techniques. Our approach introduces three key innovations: (1) a Maximum Confidence Selection (MCS) method that comprehensively evaluates confidence across all possible labels, (2) a reflection-based prompting mechanism that enhances reasoning reliability, and (3) a distance-aware calibration technique that accounts for ordinal relationships between labels. We evaluate our framework across diverse datasets, including HelpSteer2, Llama T-REx, and an internal conversational dataset, demonstrating its effectiveness across both conversational and fact-based classification tasks. This work contributes to the broader goal of developing reliable and well-calibrated confidence estimation methods for LLMs, enabling informed decisions about when to trust model outputs and when to defer to human judgement. |
cao-wang-2025-evaluating | 2,025 | Evaluating Design Choices in Verifiable Generation with Open-source Models | Verifiable generation is introduced to improve the transparency and trustworthiness of outputs produced by large language models (LLMs). Recent studies observe that open-source models struggle to include accurate citations to supporting documents in their generation with in-context learning, in contrast to the strong performance demonstrated by proprietary models. Our work aims to reveal the critical design choices that can benefit open-source models, including generation pipelines, fine-tuning methods, and inference-time compute techniques.We consider three generation pipelines, producing the outputs directly or decomposing the generation into subtasks.These generation pipelines are fine-tuned using supervised fine-tuning and preference-based optimization including further fine-tuning with rejection sampling data and direct preference optimization (DPO).The construction of preference data with varying content and citation diversity is also investigated.Additionally, we examine the benefit of an additional reranking step. With four open-source models, our experiments show that directly generating the outputs achieves the best performance. Compared to other fine-tuning methods, DPO that computes training signals from contrastive pairs consistently yields better performance, and it reaches the peak performance when the contrastive pairs are constructed with sufficient content diversity.We also find that reranking can further boost the performance of verifiable generation systems, but the marginal improvement might not justify the additional cost. |
sakib-etal-2025-battling | 2,025 | Battling Misinformation: An Empirical Study on Adversarial Factuality in Open-Source Large Language Models | Adversarial factuality refers to the deliberate insertion of misinformation into input prompts by an adversary, characterized by varying levels of expressed confidence. In this study, we systematically evaluate the performance of several open-source large language models (LLMs) when exposed to such adversarial inputs. Three tiers of adversarial confidence are considered: strongly confident, moderately confident, and limited confidence. Our analysis encompasses eight LLMs: LLaMA 3.1 (8B), Phi 3 (3.8B), Qwen 2.5 (7B), Deepseek-v2 (16B), Gemma2 (9B), Falcon (7B), Mistrallite (7B), and LLaVA (7B). Empirical results indicate that LLaMA 3.1 (8B) exhibits a robust capability in detecting adversarial inputs, whereas Falcon (7B) shows comparatively lower performance. Notably, for the majority of the models, detection success improves as the adversary`s confidence decreases; however, this trend is reversed for LLaMA 3.1 (8B) and Phi 3 (3.8B), where a reduction in adversarial confidence corresponds with diminished detection performance. Further analysis of the queries that elicited the highest and lowest rates of successful attacks reveals that adversarial attacks are more effective when targeting less commonly referenced or obscure information. |
chance-etal-2025-will | 2,025 | Will the Prince Get True Love`s Kiss? On the Model Sensitivity to Gender Perturbation over Fairytale Texts | In this paper, we study whether language models are affected by learned gender stereotypes during the comprehension of stories. Specifically, we investigate how models respond to gender stereotype perturbations through counterfactual data augmentation. Focusing on Question Answering (QA) tasks in fairytales, we modify the FairytaleQA dataset by swapping gendered character information and introducing counterfactual gender stereotypes during training. This allows us to assess model robustness and examine whether learned biases influence story comprehension. Our results show that models exhibit slight performance drops when faced with gender perturbations in the test set, indicating sensitivity to learned stereotypes. However, when fine-tuned on counterfactual training data, models become more robust to anti-stereotypical narratives. Additionally, we conduct a case study demonstrating how incorporating counterfactual anti-stereotype examples can improve inclusivity in downstream applications. |
karwa-singh-2025-disentangling | 2,025 | Disentangling Linguistic Features with Dimension-Wise Analysis of Vector Embeddings | Understanding the inner workings of neural embeddings, particularly in models such as BERT, remains a challenge because of their high-dimensional and opaque nature. This paper proposes a framework for uncovering the specific dimensions of vector embeddings that encode distinct linguistic properties (LPs). We introduce the Linguistically Distinct Sentence Pairs (LDSP-10) dataset, which isolates ten key linguistic features such as synonymy, negation, tense, and quantity. Using this dataset, we analyze BERT embeddings with various methods, including the Wilcoxon signed-rank test, mutual information, and recursive feature elimination, to identify the most influential dimensions for each LP. We introduce a new metric, the Embedding Dimension Impact (EDI) score, which quantifies the relevance of each embedding dimension to a LP. Our findings show that certain properties, such as negation and polarity, are robustly encoded in specific dimensions, while others, like synonymy, exhibit more complex patterns. This study provides insights into the interpretability of embeddings, which can guide the development of more transparent and optimized language models, with implications for model bias mitigation and the responsible deployment of AI systems. |
zakizadeh-pilehvar-2025-gender | 2,025 | Gender Encoding Patterns in Pretrained Language Model Representations | Gender bias in pretrained language models (PLMs) poses significant social and ethical challenges. Despite growing awareness, there is a lack of comprehensive investigation into how different models internally represent and propagate such biases. This study adopts an information-theoretic approach to analyze how gender biases are encoded within various encoder-based architectures.We focus on three key aspects: identifying how models encode gender information and biases, examining the impact of bias mitigation techniques and fine-tuning on the encoded biases and their effectiveness, and exploring how model design differences influence the encoding of biases.Through rigorous and systematic investigation, our findings reveal a consistent pattern of gender encoding across diverse models. Surprisingly, debiasing techniques often exhibit limited efficacy, sometimes inadvertently increasing the encoded bias in internal representations while reducing bias in model output distributions. This highlights a disconnect between mitigating bias in output distributions and addressing its internal representations. This work provides valuable guidance for advancing bias mitigation strategies and fostering the development of more equitable language models. |
rawte-etal-2025-defining | 2,025 | Defining and Quantifying Visual Hallucinations in Vision-Language Models | The troubling rise of hallucination presents perhaps the most significant impediment to the advancement of responsible AI. In recent times, considerable research has focused on detecting and mitigating hallucination in Large Language Models (LLMs). However, it`s worth noting that hallucination is also quite prevalent in Vision-Language models (VLMs). In this paper, we offer a fine-grained discourse on profiling VLM hallucination based on the image captioning task. We delineate eight fine-grained orientations of visual hallucination: i) Contextual Guessing, ii) Identity Incongruity, iii) Geographical Erratum, iv) Visual Illusion, v) Gender Anomaly, vi) VLM as Classifier, vii) Wrong Reading, and viii) Numeric Discrepancy. We curate Visual HallucInation eLiciTation, a publicly available dataset comprising 2,000 samples generated using eight VLMs across the image captioning task, along with human annotations for the categories as mentioned earlier. To establish a method for quantification and to offer a comparative framework enabling the evaluation and ranking of VLMs according to their vulnerability to producing hallucinations, we propose the Visual Hallucination Vulnerability Index (VHVI). In summary, we introduce the VHILT dataset for image-to-text hallucinations and propose the VHVI metric to quantify hallucinations in VLMs, targeting specific visual hallucination types. A subset sample is available at: \url{https://huggingface.co/datasets/vr25/vhil}. The full dataset will be publicly released upon acceptance. |
etzine-etal-2025-revitalizing | 2,025 | Revitalizing Saturated Benchmarks: A Weighted Metric Approach for Differentiating Large Language Model Performance | Existing benchmarks are becoming saturated and less effective in evaluating model performance due to factors such as data contamination and the advancing capabilities of the Large Language Models (LLMs). This paper introduces EMDM (Enhanced Model Differentiation Metric), a novel weighted metric designed to revitalize existing benchmarks. EMDM implements a weighting schema for samples based on their complexity and requisite knowledge, utilizing the performance of a baseline LLM in two experimental setups: {\textquotedblleft}Unguided{\textquotedblright}, where the model has no prior exposure to test samples, and {\textquotedblleft}Guided{\textquotedblright}, where the model has prior knowledge about the desired answer. This schema is leveraged in an optimization objective to assign weights to test samples, distinguishing instances of varying complexity. EMDM accounts for both answer correctness and the depth and accuracy of reasoning, offering a more nuanced evaluation of model performance. By weighting test examples based on their required reasoning and knowledge, EMDM achieves a distinguishing range of evaluation scores of 46{\%} among various LLMs, compared to just 17{\%} with traditional exact match (EM) metrics, thereby highlighting the saturation of current evaluation methods. |
labrak-etal-2025-synthetic | 2,025 | Synthetic Lyrics Detection Across Languages and Genres | In recent years, the use of large language models (LLMs) to generate music content, particularly lyrics, has gained in popularity. These advances provide valuable tools for artists and enhance their creative processes, but they also raise concerns about copyright violations, consumer satisfaction, and content spamming. Previous research has explored content detection in various domains. However, no work has focused on the text modality, lyrics, in music. To address this gap, we curated a diverse dataset of real and synthetic lyrics from multiple languages, music genres, and artists. The generation pipeline was validated using both humans and automated methods. We performed a thorough evaluation of existing synthetic text detection approaches on lyrics, a previously unexplored data type. We also investigated methods to adapt the best-performing features to lyrics through unsupervised domain adaptation. Following both music and industrial constraints, we examined how well these approaches generalize across languages, scale with data availability, handle multilingual language content, and perform on novel genres in few-shot settings. Our findings show promising results that could inform policy decisions around AI-generated music and enhance transparency for users. |
zhang-etal-2025-lightweight | 2,025 | A Lightweight Multi Aspect Controlled Text Generation Solution For Large Language Models | Multi-Aspect Controllable Text Generation (MCTG) introduces fine-grained multiple constraints in natural language generation, i.e. control attributes in topics, sentiments, and detoxification.MCTG demonstrates application prospects for trustworthy generation of Large Language Models (LLMs) but is limited by generalization issues.Existing work exploits additional structures and strategies for solutions, requiring LLMs' modifications.To activate LLMs' MCTG ability, we propose a lightweight MCTG pipeline based on data augmentation and instruction tuning.We analyze aspect bias and correlations in traditional datasets and address these concerns with augmented control attributes and sentences.Augmented datasets are feasible for instruction tuning.We conduct experiments for various LLMs backbone and parameter sizes, demonstrating general effectiveness on MCTG performance. |
ding-etal-2025-gender | 2,025 | Gender Bias in Large Language Models across Multiple Languages: A Case Study of ChatGPT | With the growing deployment of large language models (LLMs) across various applications, assessing the influence of gender biases embedded in LLMs becomes crucial. The topic of gender bias within the realm of natural language processing (NLP) has gained considerable focus, particularly in the context of English. Nonetheless, the investigation of gender bias in languages other than English is still relatively under-explored and insufficiently analyzed. In this work, We examine gender bias in LLMs-generated outputs for different languages. We use three measurements: 1) gender bias in selecting descriptive words given the gender-related context. 2) gender bias in selecting gender-related pronouns (she/he) given the descriptive words. 3) gender bias in the topics of LLM-generated dialogues. We investigate the outputs of the GPT series of LLMs in various languages using our three measurement methods. Our findings revealed significant gender biases across all the languages we examined. |
varshney-etal-2025-investigating | 2,025 | Investigating and Addressing Hallucinations of LLMs in Tasks Involving Negation | Large Language Models (LLMs) have achieved remarkable performance across a wide variety of natural language tasks. However, they have been shown to suffer from a critical limitation pertinent to {\textquoteleft}hallucination' in their output. Recent research has focused on investigating and addressing this problem for a variety of tasks such as biography generation, question answering, abstractive summarization, and dialogue generation. However, the crucial aspect pertaining to {\textquoteleft}negation' has remained considerably underexplored. Negation is important because it adds depth and nuance to the understanding of language and is also crucial for logical reasoning and inference. In this work, we address the above limitation and particularly focus on studying the impact of negation in LLM hallucinations. Specifically, we study four tasks with negation: {\textquoteleft}false premise completion', {\textquoteleft}constrained fact generation', {\textquoteleft}multiple choice question answering', and {\textquoteleft}fact generation'. We show that open-source state-of-the-art LLMs such as LLaMA-2-chat, Vicuna, and Orca-2 hallucinate considerably on all these tasks involving negation which underlines a critical shortcoming of these models. Addressing this problem, we further study numerous strategies to mitigate these hallucinations and demonstrate their impact. |
zevallos-etal-2025-first | 2,025 | The First Multilingual Model For The Detection of Suicide Texts | Suicidal ideation is a serious health problem affecting millions of people worldwide. Social networks provide information about these mental health problems through users' emotional expressions. We propose a multilingual model leveraging transformer architectures like mBERT, XML-R, and mT5 to detect suicidal text across posts in six languages - Spanish, English, German, Catalan, Portuguese and Italian. A Spanish suicide ideation tweet dataset was translated into five other languages using SeamlessM4T. Each model was fine-tuned on this multilingual data and evaluated across classification metrics. Results showed mT5 achieving the best performance overall with F1 scores above 85{\%}, highlighting capabilities for cross-lingual transfer learning. The English and Spanish translations also displayed high quality based on perplexity. Our exploration underscores the importance of considering linguistic diversity in developing automated multilingual tools to identify suicidal risk. Limitations exist around semantic fidelity in translations and ethical implications which provide guidance for future human-in-the-loop evaluations. |
lin-etal-2025-crossin | 2,025 | CrossIn: An Efficient Instruction Tuning Approach for Cross-Lingual Knowledge Alignment | Multilingual proficiency presents a significant challenge for large language models (LLMs). English-centric models are usually suboptimal in other languages, particularly those that are linguistically distant from English. This performance discrepancy mainly stems from the imbalanced distribution of training data across languages during pre-training and instruction tuning stages. To address this problem, we propose a novel approach called CrossIn, which utilizes a mixed composition of cross-lingual instruction tuning data. Our method leverages the compressed representation shared by various languages to efficiently enhance the model`s task-solving capabilities and multilingual proficiency within a single process. In addition, we introduce a multi-task and multi-faceted benchmark to evaluate the effectiveness of CrossIn. Experimental results demonstrate that our method substantially improves performance across tasks and languages, and we provide extensive insights into the impact of cross-lingual data volume and the integration of translation data on enhancing multilingual consistency and accuracy. |
srirag-etal-2025-evaluating | 2,025 | Evaluating Dialect Robustness of Language Models via Conversation Understanding | With an evergrowing number of LLMs reporting superlative performance for English, their ability to perform equitably for different dialects of English (i.e., dialect robustness) needs to be ascertained. Specifically, we use English language (US English or Indian English) conversations between humans who play the word-guessing game of {\textquoteleft}taboo{\textquoteleft}. We formulate two evaluative tasks: target word prediction (TWP) (i.e., predict the masked target word in a conversation) and target word selection (TWS) (i.e., select the most likely masked target word in a conversation, from among a set of candidate words). Extending MD3, an existing dialectic dataset of taboo-playing conversations, we introduce M-MD3, a target-word-masked version of MD3 with the en-US and en-IN subsets. We create two subsets: en-MV (where en-US is transformed to include dialectal information) and en-TR (where dialectal information is removed from en-IN). We evaluate three multilingual LLMs{--}one open source (Llama3) and two closed-source (GPT-4/3.5). LLMs perform significantly better for US English than Indian English for both TWP and TWS tasks, for all settings, exhibiting marginalisation against the Indian dialect of English. While GPT-based models perform the best, the comparatively smaller models work more equitably after fine-tuning. Our evaluation methodology exhibits a novel and reproducible way to examine attributes of language models using pre-existing dialogue datasets with language varieties. Dialect being an artifact of one`s culture, this paper demonstrates the gap in the performance of multilingual LLMs for communities that do not use a mainstream dialect. |
tashu-etal-2025-cross | 2,025 | Cross-Lingual Document Recommendations with Transformer-Based Representations: Evaluating Multilingual Models and Mapping Techniques | Recommendation systems, for documents, have become tools for finding relevant content on the Web. However, these systems have limitations when it comes to recommending documents in languages different from the query language, which means they might overlook resources in non-native languages. This research focuses on representing documents across languages by using Transformer Leveraged Document Representations (TLDRs) that are mapped to a cross-lingual domain. Four multilingual pre-trained transformer models (mBERT, mT5 XLM RoBERTa, ErnieM) were evaluated using three mapping methods across 20 language pairs representing combinations of five selected languages of the European Union. Metrics like Mate Retrieval Rate and Reciprocal Rank were used to measure the effectiveness of mapped TLDRs compared to non-mapped ones. The results highlight the power of cross-lingual representations achieved through pre-trained transformers and mapping approaches suggesting a promising direction for expanding beyond language connections, between two specific languages. |
End of preview. Expand
in Data Studio
ACL abstracts extracted from the official ACL Anthology BibTeX.
- Downloads last month
- 67