ID
large_stringlengths 10
61
| year
int64 1.96k
2.03k
| title
large_stringlengths 4
560
| abstract
large_stringlengths 0
12.8k
|
|---|---|---|---|
bonaldi-etal-2025-first | 2,025 | The First Workshop on Multilingual Counterspeech Generation at COLING 2025: Overview of the Shared Task | This paper presents an overview of the Shared Task organized in the First Workshop on Multilingual Counterspeech Generation at COLING 2025. While interest in automatic approaches to Counterspeech generation has been steadily growing, the large majority of the published experimental work has been carried out for English. This is due to the scarcity of both non-English manually curated training data and to the crushing predominance of English in the generative Large Language Models (LLMs) ecosystem. The task`s goal is to promote and encourage research on Counterspeech generation in a multilingual setting (Basque, English, Italian, and Spanish) potentially leveraging background knowledge provided in the proposed dataset. The task attracted 11 participants, 9 of whom presented a paper describing their systems. Together with the task, we introduce a new multilingual counterspeech dataset with 2384 triplets of hate speech, counterspeech, and related background knowledge covering 4 languages. The dataset is available at: https://huggingface.co/datasets/LanD-FBK/ML{\_}MTCONAN{\_}KN. |
hettiarachchi-etal-2025-overview | 2,025 | Overview of the First Workshop on Language Models for Low-Resource Languages (LoResLM 2025) | The first Workshop on Language Models for Low-Resource Languages (LoResLM 2025) was held in conjunction with the 31st International Conference on Computational Linguistics (COLING 2025) in Abu Dhabi, United Arab Emirates. This workshop mainly aimed to provide a forum for researchers to share and discuss their ongoing work on language models (LMs) focusing on low-resource languages, following the recent advancements in neural language models and their linguistic biases towards high-resource languages. LoResLM 2025 attracted notable interest from the natural language processing (NLP) community, resulting in 35 accepted papers from 52 submissions. These contributions cover a broad range of low-resource languages from eight language families and 13 diverse research areas, paving the way for future possibilities and promoting linguistic inclusivity in NLP. |
shang-etal-2025-atlas | 2,025 | Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect | We introduce Atlas-Chat, the first-ever collection of LLMs specifically developed for dialectal Arabic. Focusing on Moroccan Arabic, also known as Darija, we construct our instruction dataset by consolidating existing Darija language resources, creating novel datasets both manually and synthetically, and translating English instructions with stringent quality control. Atlas-Chat-2B, 9B, and 27B models, fine-tuned on the dataset, exhibit superior ability in following Darija instructions and performing standard NLP tasks. Notably, our models outperform both state-of-the-art and Arabic-specialized LLMs like LLaMa, Jais, and AceGPT, e.g., our 9B model gains a 13{\%} performance boost over a larger 13B model on DarijaMMLU, in our newly introduced evaluation suite for Darija covering both discriminative and generative tasks. Furthermore, we perform an experimental analysis of various fine-tuning strategies and base model choices to determine optimal configurations. All our resources are publicly accessible, and we believe our work offers comprehensive design methodologies of instruction-tuning for low-resource languages, which are often neglected in favor of data-rich languages by contemporary LLMs. |
mokhtarabadi-etal-2025-empowering | 2,025 | Empowering Persian LLMs for Instruction Following: A Novel Dataset and Training Approach | Instruction-tuned large language models have demonstrated remarkable capabilities in following human instructions across various domains. However, their proficiency remains notably deficient in many low-resource languages. To address this challenge, we begin by introducing FarsInstruct: a comprehensive instruction dataset designed to enhance the instruction-following ability of large language models specifically for the Persian language{---}a significant yet underrepresented language globally. FarsInstruct encompasses a wide range of task types and datasets, each containing a mix of straightforward to complex manual written instructions, as well as translations from the Public Pool of Prompts, ensuring a rich linguistic and cultural representation. Furthermore, we introduce Co-CoLA, a framework designed to enhance the multi-task adaptability of LoRA-tuned models. Through extensive experimental analyses, our study showcases the effectiveness of the FarsInstruct dataset coupled with training by the Co-CoLA framework, in improving the performance of large language models within the Persian context. As of the current writing, FarsInstruct comprises 197 templates across 21 distinct datasets, and we intend to update it consistently, thus augmenting its applicability. |
alam-etal-2025-bnsentmix | 2,025 | BnSentMix: A Diverse Bengali-English Code-Mixed Dataset for Sentiment Analysis | The widespread availability of code-mixed data in digital spaces can provide valuable insights into low-resource languages like Bengali, which have limited annotated corpora. Sentiment analysis, a pivotal text classification task, has been explored across multiple languages, yet code-mixed Bengali remains underrepresented with no large-scale, diverse benchmark. Code-mixed text is particularly challenging as it requires the understanding of multiple languages and their interaction in the same text. We address this limitation by introducing BnSentMix, a sentiment analysis dataset on code-mixed Bengali comprising 20,000 samples with 4 sentiment labels, sourced from Facebook, YouTube, and e-commerce sites. By aggregating multiple sources, we ensure linguistic diversity reflecting realistic code-mixed scenarios. We implement a novel automated text filtering pipeline using fine-tuned language models to detect code-mixed samples and expand code-mixed text corpora. We further propose baselines using machine learning, neural networks, and transformer-based language models. The availability of a diverse dataset is a critical step towards democratizing NLP and ultimately contributing to a better understanding of code-mixed languages. |
habibzadeh-asadpour-2025-using | 2,025 | Using Language Models for assessment of users' satisfaction with their partner in Persian | Sentiment analysis, the process of gauging user attitudes and emotions through their textual data, including social media posts and other forms of communication, is a valuable tool for informed decision-making. In other words, a statement conveys positivity, negativity, or neutrality, sentiment analysis offers insights into public sentiment regarding a product, individual, event, or other significant topics. This research focuses on the effectiveness of sentiment analysis techniques, using Machine Learning (ML) and Natural Language Processing (NLP) especially pre-trained language models for Persian, in assessing users' satisfaction with their partner, using data collected from X (formerly Twitter). Our motivation stems from traditional in-person surveys, which periodically analyze societal challenges in Iran. The limitations of these surveys led us to explore Artificial Intelligence (AI) as an alternative solution for addressing contemporary social issues. We collected Persian tweets and utilized data annotation techniques to label them according to our research question, forming the dataset. Our goal also was to provide a benchmark of Persian tweets on this specific topic. To evaluate our dataset, we employed several classification methods to achieve our goal, including classical ML models, Deep Neural Networks, and pre-trained language models for Persian. Following a comprehensive evaluation, our results show that BERTweet-FA (one of the pre-trained language models for Persian) emerged as the best performer among the classifiers for assessing users' satisfaction. This point indicates the ability of language models to understand conversational Persian text and perform sentiment analysis, even in a low-resource language like Persian. |
mutsaddi-choudhary-2025-enhancing | 2,025 | Enhancing Plagiarism Detection in Marathi with a Weighted Ensemble of TF-IDF and BERT Embeddings for Low-Resource Language Processing | Plagiarism involves using another person`s work or concepts without proper attribution, presenting them as original creations. With the growing amount of data communicated in regional languages such as Marathi{---}one of India`s regional languages{---}it is crucial to design robust plagiarism detection systems tailored for low-resource languages. Language models like Bidirectional Encoder Representations from Transformers (BERT) have demonstrated exceptional capability in text representation and feature extraction, making them essential tools for semantic analysis and plagiarism detection. However, the application of BERT for low-resource languages remains underexplored, particularly in the context of plagiarism detection. This paper presents a method to enhance the accuracy of plagiarism detection for Marathi texts using BERT sentence embeddings in conjunction with Term Frequency-Inverse Document Frequency (TF-IDF) feature representation. By combining TF-IDF with BERT, the system`s performance is significantly improved, which is especially pronounced in languages where BERT models are not extremely robust due to a lack of resources and corpora. This approach effectively captures statistical, semantic, and syntactic aspects of text features through a weighted voting ensemble of machine learning models. |
sani-etal-2025-investigating | 2,025 | Investigating the Impact of Language-Adaptive Fine-Tuning on Sentiment Analysis in Hausa Language Using AfriBERTa | Sentiment analysis (SA) plays a vital role in Natural Language Processing (NLP) by identifying sentiments expressed in text. Although significant advances have been made in SA for widely spoken languages, low-resource languages such as Hausa face unique challenges, primarily due to a lack of digital resources. This study investigates the effectiveness of Language-Adaptive Fine-Tuning (LAFT) to improve SA performance in Hausa. We first curate a diverse, unlabeled corpus to expand the model`s linguistic capabilities, followed by applying LAFT to adapt AfriBERTa specifically to the nuances of the Hausa language. The adapted model is then fine-tuned on the labeled NaijaSenti sentiment dataset to evaluate its performance. Our findings demonstrate that LAFT gives modest improvements, which may be attributed to the use of formal Hausa text rather than informal social media data. Nevertheless, the pre-trained AfriBERTa model significantly outperformed models not specifically trained on Hausa, highlighting the importance of using pre-trained models in low-resource contexts. This research emphasizes the necessity for diverse data sources to advance NLP applications for low-resource African languages. We will publish the code and the data set to encourage further research and facilitate reproducibility in low-resource NLP |
zhukova-etal-2025-automated | 2,025 | Automated Collection of Evaluation Dataset for Semantic Search in Low-Resource Domain Language | Domain-specific languages that use a lot of specific terminology often fall into the category of low-resource languages. Collecting test datasets in a narrow domain is time-consuming and requires skilled human resources with domain knowledge and training for the annotation task. This study addresses the challenge of automated collecting test datasets to evaluate semantic search in low-resource domain-specific German language of the process industry. Our approach proposes an end-to-end annotation pipeline for automated query generation to the score reassessment of query-document pairs. To overcome the lack of text encoders trained in the German chemistry domain, we explore a principle of an ensemble of {\textquotedblleft}weak{\textquotedblright} text encoders trained on common knowledge datasets. We combine individual relevance scores from diverse models to retrieve document candidates and relevance scores generated by an LLM, aiming to achieve consensus on query-document alignment. Evaluation results demonstrate that the ensemble method significantly improves alignment with human-assigned relevance scores, outperforming individual models in both inter-coder agreement and accuracy metrics. These findings suggest that ensemble learning can effectively adapt semantic search systems for specialized, low-resource languages, offering a practical solution to resource limitations in domain-specific contexts. |
gamboa-lee-2025-filipino | 2,025 | Filipino Benchmarks for Measuring Sexist and Homophobic Bias in Multilingual Language Models from Southeast Asia | Bias studies on multilingual models confirm the presence of gender-related stereotypes in masked models processing languages with high NLP resources. We expand on this line of research by introducing Filipino CrowS-Pairs and Filipino WinoQueer: benchmarks that assess both sexist and anti-queer biases in pretrained language models (PLMs) handling texts in Filipino, a low-resource language from the Philippines. The benchmarks consist of 7,074 new challenge pairs resulting from our cultural adaptation of English bias evaluation datasets{---}a process that we document in detail to guide similar forthcoming efforts. We apply the Filipino benchmarks on masked and causal multilingual models, including those pretrained on Southeast Asian data, and find that they contain considerable amounts of bias. We also find that for multilingual models, the extent of bias learned for a particular language is influenced by how much pretraining data in that language a model was exposed to. Our benchmarks and insights can serve as a foundation for future work analyzing and mitigating bias in multilingual models. |
tran-etal-2025-exploiting | 2,025 | Exploiting Word Sense Disambiguation in Large Language Models for Machine Translation | Machine Translation (MT) has made great strides with the use of Large Language Models (LLMs) and advanced prompting techniques. However, translating sentences with ambiguous words remains challenging, especially when LLMs have limited proficiency in the source language. This paper introduces two methods to enhance MT performance by leveraging the word sense disambiguation capabilities of LLMs. The first method integrates all the available senses of an ambiguous word into the prompting template. The second method uses a pre-trained source language model to predict the correct sense of the ambiguous word, which is then incorporated into the prompting template. Additionally, we propose two prompting template styles for providing word sense information to LLMs. Experiments on the HOLLY dataset demonstrate the effectiveness of our approach in improving MT performance. |
rapacz-smywinski-pohl-2025-low | 2,025 | Low-Resource Interlinear Translation: Morphology-Enhanced Neural Models for Ancient Greek | Contemporary machine translation systems prioritize fluent, natural-sounding output with flexible word ordering. In contrast, interlinear translation maintains the source text`s syntactic structure by aligning target language words directly beneath their source counterparts. Despite its importance in classical scholarship, automated approaches to interlinear translation remain understudied. We evaluated neural interlinear translation from Ancient Greek to English and Polish using four transformer-based models: two Ancient Greek-specialized (GreTa and PhilTa) and two general-purpose multilingual models (mT5-base and mT5-large). Our approach introduces novel morphological embedding layers and evaluates text preprocessing and tag set selection across 144 experimental configurations using a word-aligned parallel corpus of the Greek New Testament. Results show that morphological features through dedicated embedding layers significantly enhance translation quality, improving BLEU scores by 35{\%} (44.67 {\textrightarrow} 60.40) for English and 38{\%} (42.92 {\textrightarrow} 59.33) for Polish compared to baseline models. PhilTa achieves state-of-the-art performance for English, while mT5-large does so for Polish. Notably, PhilTa maintains stable performance using only 10{\%} of training data. Our findings challenge the assumption that modern neural architectures cannot benefit from explicit morphological annotations. While preprocessing strategies and tag set selection show minimal impact, the substantial gains from morphological embeddings demonstrate their value in low-resource scenarios. |
merad-etal-2025-language | 2,025 | Language verY Rare for All | In the quest to overcome language barriers, encoder-decoder models like NLLB have expanded machine translation to rare languages, with some models (e.g., NLLB 1.3B) even trainable on a single GPU. While general-purpose LLMs perform well in translation, open LLMs prove highly competitive when fine-tuned for specific tasks involving unknown corpora. We introduce LYRA (Language verY Rare for All), a novel approach that combines open LLM fine-tuning, retrieval-augmented generation (RAG), and transfer learning from related high-resource languages. This study is exclusively focused on single-GPU training to facilitate ease of adoption. Our study focuses on two-way translation between French and Mon{\'e}gasque {---} a rare language unsupported by existing translation tools due to limited corpus availability. Our results demonstrate LYRA`s effectiveness, frequently surpassing and consistently matching state-of-the-art encoder-decoder models in rare language translation. |
donthi-etal-2025-improving | 2,025 | Improving LLM Abilities in Idiomatic Translation | Translating idiomatic expressions remains a challenge for large language models (LLMs), as they often produce literal, semantically incorrect translations{---}for instance, directly converting {\textquotedblleft}break a leg{\textquotedblright} into a nonsensical phrase in the target language. While external resources like IdiomKB can supply the figurative meaning and thus yield semantically accurate translations, this approach does not preserve the cultural and stylistic nuances that make idioms so distinctive. Our study focuses on idiomatic translations across multiple languages, including Chinese (ZH), Urdu (UR), and Hindi (HI), with clearly defined abbreviations for each. We propose two methods for improving idiomatic translation fidelity: a Semantic Idiom Alignment (SIA) approach that uses pre-trained sentence embeddings to identify target-language idioms, and a Language-Model-based Idiom Alignment (LIA) approach that prompts an LLM to suggest appropriate idiom counterparts. Human evaluations across multiple language pairs show that SIA better preserves idiomatic style. To support this work, we introduce idiom datasets in low-resource languages (Urdu and Hindi). Our results indicate that aligning idioms at the semantic level can improve cross-lingual style preservation and cultural authenticity. |
liu-etal-2025-comparative | 2,025 | A Comparative Study of Static and Contextual Embeddings for Analyzing Semantic Changes in Medieval Latin Charters | The Norman Conquest of 1066 C.E. brought profound transformations to England`s administrative, societal, and linguistic practices. The DEEDS (Documents of Early England Data Set) database offers a unique opportunity to explore these changes by examining shifts in word meanings within a vast collection of Medieval Latin charters. While computational linguistics typically relies on vector representations of words like static and contextual embeddings to analyze semantic changes, existing embeddings for scarce and historical Medieval Latin are limited and may not be well-suited for this task. This paper presents the first computational analysis of semantic change pre- and post-Norman Conquest and the first systematic comparison of static and contextual embeddings in a scarce historical data set. Our findings confirm that, consistent with existing studies, contextual embeddings outperform static word embeddings in capturing semantic change within a scarce historical corpus. |
ouattara-etal-2025-bridging | 2,025 | Bridging Literacy Gaps in African Informal Business Management with Low-Resource Conversational Agents | Position paper: In many African countries, the informal business sector represents the backbone of the economy, providing essential livelihoods and opportunities where formal employment is limited. Despite, however, the growing adoption of digital tools, entrepreneurs in this sector often face significant challenges due to lack of literacy and language barriers. These barriers not only limit accessibility but also increase the risk of fraud and financial insecurity. This position paper explores the potential of conversational agents (CAs) adapted to low-resource languages (LRLs), focusing specifically on Moor{\'e}, a language widely spoken in Burkina Faso. By enabling natural language interactions in local languages, AI-driven conversational agents offer a promising solution to enable informal traders to manage their financial transactions independently, thus promoting greater autonomy and security in business, while providing a step towards formalization of their business. Our study examines the main challenges in developing AI for African languages, including data scarcity and linguistic diversity, and reviews viable strategies for addressing them, such as cross-lingual transfer learning and data augmentation techniques. |
sadhu-etal-2025-social | 2,025 | Social Bias in Large Language Models For Bangla: An Empirical Study on Gender and Religious Bias | The rapid growth of Large Language Models (LLMs) has put forward the study of biases as a crucial field. It is important to assess the influence of different types of biases embedded in LLMs to ensure fair use in sensitive fields. Although there have been extensive works on bias assessment in English, such efforts are rare and scarce for a major language like Bangla. In this work, we examine two types of social biases in LLM generated outputs for Bangla language. Our main contributions in this work are: (1) bias studies on two different social biases for Bangla, (2) a curated dataset for bias measurement benchmarking and (3) testing two different probing techniques for bias detection in the context of Bangla. This is the first work of such kind involving bias assessment of LLMs for Bangla to the best of our knowledge. All our code and resources are publicly available for the progress of bias related research in Bangla NLP. |
cruz-2025-extracting | 2,025 | Extracting General-use Transformers for Low-resource Languages via Knowledge Distillation | In this paper, we propose the use of simple knowledge distillation to produce smaller and more efficient single-language transformers from Massively Multilingual Transformers (MMTs) to alleviate tradeoffs associated with the use of such in low-resource settings. Using Tagalog as a case study, we show that these smaller single-language models perform on-par with strong baselines in a variety of benchmark tasks in a much more efficient manner. Furthermore, we investigate additional steps during the distillation process that improves the soft-supervision of the target language, and provide a number of analyses and ablations to show the efficacy of the proposed method. |
bagheri-nezhad-etal-2025-beyond | 2,025 | Beyond Data Quantity: Key Factors Driving Performance in Multilingual Language Models | Multilingual language models (MLLMs) are crucial for handling text across various languages, yet they often show performance disparities due to differences in resource availability and linguistic characteristics. While the impact of pre-train data percentage and model size on performance is well-known, our study reveals additional critical factors that significantly influence MLLM effectiveness. Analyzing a wide range of features, including geographical, linguistic, and resource-related aspects, we focus on the SIB-200 dataset for classification and the Flores-200 dataset for machine translation, using regression models and SHAP values across 204 languages. Our findings identify token similarity and country similarity as pivotal factors, alongside pre-train data and model size, in enhancing model performance. Token similarity facilitates cross-lingual transfer, while country similarity highlights the importance of shared cultural and linguistic contexts. These insights offer valuable guidance for developing more equitable and effective multilingual language models, particularly for underrepresented languages. |
matzopoulos-etal-2025-babylms | 2,025 | BabyLMs for isiXhosa: Data-Efficient Language Modelling in a Low-Resource Context | The BabyLM challenge called on participants to develop sample-efficient language models. Submissions were pretrained on a fixed English corpus, limited to the amount of words children are exposed to in development ({\ensuremath{<}}100m). The challenge produced new architectures for data-efficient language modelling, outperforming models trained on trillions of words. This is promising for low-resource languages, where available corpora are limited to much less than 100m words. In this paper, we explore the potential of BabyLMs for low-resource languages, using the isiXhosa language as a case study. We pretrain two BabyLM architectures, ELC-BERT and MLSM, on an isiXhosa corpus. They outperform a vanilla pretrained model on POS tagging and NER, achieving notable gains (+3.2 F1) for the latter. In some instances, the BabyLMs even outperform XLM-R. Our findings show that data-efficient models are viable for low-resource languages, but highlight the continued importance, and lack of, high-quality pretraining data. Finally, we visually analyse how BabyLM architectures encode isiXhosa. |
tashu-tudor-2025-mapping | 2,025 | Mapping Cross-Lingual Sentence Representations for Low-Resource Language Pairs Using Pre-trained Language Models | In this work, we explore different linear mapping techniques to learn cross-lingual document representations from pre-trained multilingual large language models for low-resource languages. Three different mapping techniques namely Linear Concept Approximation (LCA), Linear Concept Compression (LCC), and Neural Concept Approximation (NCA) and four multilingual language models such as mBERT, mT5, XLM-R, and ErnieM were used to extract embeddings. The inter-lingual representations were created mappings the monolingual representation extracted from multilingual language models. The experimental results showed that LCA and LCC significantly outperform NCA, with models like ErnieM achieving the highest alignment quality. Language pairs exhibit variable performance, influenced by linguistic similarity and data availability, with the Amharic-English pair yielding particularly high scores. The results showed the utility of LCA and LCC in enabling cross-lingual tasks for low-resource languages. |
harju-van-der-goot-2025-age | 2,025 | How to age BERT Well: Continuous Training for Historical Language Adaptation | As the application of computational tools increases to digitalize historical archives, automatic annotation challenges persist due to distinct linguistic and morphological features of historical languages like Old English (OE). Existing tools struggle with the historical language varieties due to insufficient training. Previous research has focused on adapting pre-trained language models to new languages or domains but has rarely explored the modeling of language variety across time. Hence, we investigate the effectiveness of continuous language model training for adapting language models to OE on domain-specific data. We compare the continuous training of an English model (EN) and a multilingual model (ML), and use POS tagging for downstream evaluation. Results show that continuous pre-training substantially improves performance. We retrain a modern English (EN) model and a Multi-lingual (ML) BERT model for OE. We confirmed the effectiveness of continuous pre-training for language adaptation and downstream evaluation utilizing part-of-speech (POS) tagging, advancing the potential to understand the unique grammatical structures of historical OE archives. More concretely, EN BERT initially outperformed ML BERT with an accuracy of 83{\%} during the language modeling phase. However, on the POS tagging task, ML BERT surpassed EN BERT, achieving an accuracy of 94{\%}, which suggests effective performance to the historical language varieties. |
amin-etal-2025-exploiting | 2,025 | Exploiting Task Reversibility of DRS Parsing and Generation: Challenges and Insights from a Multi-lingual Perspective | Semantic parsing and text generation exhibit reversible properties when utilizing Discourse Representation Structures (DRS). However, both processes{---}text-to-DRS parsing and DRS-to-text generation{---}are susceptible to errors. In this paper, we exploit the reversible nature of DRS to explore both error propagation, which is commonly seen in pipeline methods, and the less frequently studied potential for error correction. We investigate two pipeline approaches: Parse-Generate-Parse (PGP) and Generate-Parse-Generate (GPG), utilizing pre-trained language models where the output of one model becomes the input for the next. Our evaluation uses the Parallel Meaning Bank dataset, focusing on Urdu as a low-resource language, Italian as a mid-resource language, and English serving as a high-resource baseline. Our analysis highlights that while pipelines are theoretically suited for error correction, they more often propagate errors, with Urdu exhibiting the greatest sensitivity, Italian showing a moderate effect, and English demonstrating the highest stability. This variation highlights the unique challenges faced by low-resource languages in semantic processing tasks. Further, our findings suggest that these pipeline methods support the development of more linguistically balanced datasets, enabling a comprehensive assessment across factors like sentence structure, length, type, polarity, and voice. Our cross-linguistic analysis provides valuable insights into the behavior of DRS processing in low-resource contexts, demonstrating both the potential and limitations of reversible pipeline approaches. |
bobojonova-etal-2025-bbpos | 2,025 | BBPOS: BERT-based Part-of-Speech Tagging for Uzbek | This paper advances NLP research for the low-resource Uzbek language by evaluating two previously untested monolingual Uzbek BERT models on the part-of-speech (POS) tagging task and introducing the first publicly available UPOS-tagged benchmark dataset for Uzbek. Our fine-tuned models achieve 91{\%} average accuracy, outperforming the baseline multi-lingual BERT as well as the rule-based tagger. Notably, these models capture intermediate POS changes through affixes and demonstrate context sensitivity, unlike existing rule-based taggers. |
dewangan-etal-2025-every | 2,025 | When Every Token Counts: Optimal Segmentation for Low-Resource Language Models | Traditional greedy tokenization methods have been a critical step in Natural Language Processing (NLP), influencing how text is converted into tokens and directly impacting model performance. While subword tokenizers like Byte-Pair Encoding (BPE) are widely used, questions remain about their optimality across model scales and languages. In this work, we demonstrate through extensive experiments that an optimal BPE configuration significantly reduces token count compared to greedy segmentation, yielding improvements in token-saving percentages and performance benefits, particularly for smaller models. We evaluate tokenization performance across various intrinsic and extrinsic tasks, including generation and classification. Our findings suggest that compression-optimized tokenization strategies could provide substantial advantages for multilingual and low-resource (LR) language applications, highlighting a promising direction for further research and inclusive NLP. |
veitsman-hartmann-2025-recent | 2,025 | Recent Advancements and Challenges of Turkic Central Asian Language Processing | Research in NLP for Central Asian Turkic languages - Kazakh, Uzbek, Kyrgyz, and Turkmen - faces typical low-resource language challenges like data scarcity, limited linguistic resources and technology development. However, recent advancements have included the collection of language-specific datasets and the development of models for downstream tasks. Thus, this paper aims to summarize recent progress and identify future research directions. It provides a high-level overview of each language`s linguistic features, the current technology landscape, the application of transfer learning from higher-resource languages, and the availability of labeled and unlabeled data. By outlining the current state, we hope to inspire and facilitate future research. |
lasheras-pinheiro-2025-calquest | 2,025 | CaLQuest.PT: Towards the Collection and Evaluation of Natural Causal Ladder Questions in Portuguese for AI Agents | Large Language Models (LLMs) are increasingly central to the development of generative AI across diverse fields. While some anticipate these models may mark a step toward artificial general intelligence, their ability to handle complex causal reasoning remains unproven. Causal reasoning, particularly at Pearl`s interventional and counterfactual levels, is essential for true general intelligence. In this work, we introduce CaLQuest.PT, a dataset of over 8,000 natural causal questions in Portuguese, collected from real human interactions. Built upon a novel three-axis taxonomy, CaLQuest.PT categorizes questions by causal intent, action requirements, and the level of causal reasoning needed (associational, interventional, or counterfactual). Our findings from evaluating CaLQuest.PT`s seed questions with GPT-4o reveal that this LLM face challenges in handling interventional and relation-seeking causal queries. These results suggest limitations in using GPT-4o for extending causal question annotations and highlight the need for improved LLM strategies in causal reasoning. CaLQuest.PT provides a foundation for advancing LLM capabilities in causal understanding, particularly for the Portuguese-speaking world. |
zeinalipour-etal-2025-persianmcq | 2,025 | PersianMCQ-Instruct: A Comprehensive Resource for Generating Multiple-Choice Questions in Persian | We present PersianMCQ-Instruct, a comprehensive resource that includes a dataset and advanced models for generating multiple-choice questions (MCQs) in standard Iranian Persian, a low-resource language spoken by over 80 million people. This resource features three state-of-the-art models for Persian MCQ generation: PMCQ-Gemma2-9b, PMCQ-Llama3.1-8b, and PMCQ-Mistral-7B. Inspired by the Agent Instruct framework and GPT-4o, we created the dataset by curating over 4,000 unique Persian Wikipedia pages, resulting in three MCQs per page and a total of over 12,000 questions. To ensure the quality of this dataset, we conducted human evaluations and model fine-tuning, both of which demonstrated significant performance improvements in Persian MCQ generation. The dataset and models are publicly available, offering valuable tools for researchers and educators, with particular benefits for advancing Persian-language educational technology. |
turumtaev-2025-stop | 2,025 | Stop Jostling: Adaptive Negative Sampling Reduces the Marginalization of Low-Resource Language Tokens by Cross-Entropy Loss | Neural language models often struggle with low-resource languages due to the limited availability of training data, making tokens from these languages rare in the training set. This paper addresses a specific challenge during training: rare tokens are disproportionately affected by marginalization, which prevents them from learning effectively. We propose a thresholding technique that reduces the impact of this marginalization, allowing rare tokens to benefit from more meaningful alignment. Through experiments with a character-level language model, we demonstrate that this method significantly improves performance on low-resource language validation data. This work is the first to show how negative sampling can be applied to improve the representation of rare tokens by limiting the harmful influence of excessive marginalization, offering a new approach to enhancing language model performance for underrepresented languages. |
nacar-etal-2025-towards | 2,025 | Towards Inclusive Arabic LLMs: A Culturally Aligned Benchmark in Arabic Large Language Model Evaluation | Arabic Large Language Models are usually evaluated using Western-centric benchmarks that overlook essential cultural contexts, making them less effective and culturally misaligned for Arabic-speaking communities. This study addresses this gap by evaluating the Arabic Massive Multitask Language Understanding (MMLU) Benchmark to assess its cultural alignment and relevance for Arabic Large Language Models (LLMs) across culturally sensitive topics. A team of eleven experts annotated over 2,500 questions, evaluating them based on fluency, adequacy, cultural appropriateness, bias detection, religious sensitivity, and adherence to social norms. Through human assessment, the study highlights significant cultural misalignments and biases, particularly in sensitive areas like religion and morality. In response to these findings, we propose annotation guidelines and integrate culturally enriched data sources to enhance the benchmark`s reliability and relevance. The research highlights the importance of cultural sensitivity in evaluating inclusive Arabic LLMs, fostering more widely accepted LLMs for Arabic-speaking communities. |
kryvosheieva-levy-2025-controlled | 2,025 | Controlled Evaluation of Syntactic Knowledge in Multilingual Language Models | Language models (LMs) are capable of acquiring elements of human-like syntactic knowledge. Targeted syntactic evaluation tests have been employed to measure how well they form generalizations about syntactic phenomena in high-resource languages such as English. However, we still lack a thorough understanding of LMs' capacity for syntactic generalizations in low-resource languages, which are responsible for much of the diversity of syntactic patterns worldwide. In this study, we develop targeted syntactic evaluation tests for three low-resource languages (Basque, Hindi, and Swahili) and use them to evaluate five families of open-access multilingual Transformer LMs. We find that some syntactic tasks prove relatively easy for LMs while others (agreement in sentences containing indirect objects in Basque, agreement across a prepositional phrase in Swahili) are challenging. We additionally uncover issues with publicly available Transformers, including a bias toward the habitual aspect in Hindi in multilingual BERT and underperformance compared to similar-sized models in XGLM-4.5B. |
zhu-etal-2025-evaluating | 2,025 | Evaluating Large Language Models for In-Context Learning of Linguistic Patterns In Unseen Low Resource Languages | This paper investigates the ability of Large language Models (LLMs) in capturing linguistic patterns from unseen languages and applying them to translation between the languages and English within an in-context learning framework. Inspired by the International Linguistics Olympiad (IOL), we create test data consisting of translation puzzles between 40 low resource languages and English. We test the LLMs in two different strategies: direct prompting and step-by-step prompting. In the latter, the puzzles are manually decomposed into intermediate steps to allow LLMs learn and apply linguistic rules incrementally. The results show that this strategy can significantly improve the performance of LLMs, achieving comparable or slightly superior results to humans when translating the unseen languages to English. However, LLMs still struggle with translating English into the unseen languages, typically with complex syntactic rules. We further observe that LLMs cannot deal with languages with object-subject and noun-adjective word order compared to others, reflecting the potential impact imposed by typological features of languages in training data. |
dai-etal-2025-next | 2,025 | Next-Level Cantonese-to-Mandarin Translation: Fine-Tuning and Post-Processing with LLMs | Large Language Models (LLMs) have improved performance across various natural language processing tasks. Despite these improvements, LLMs continue to face significant challenges, such as grammatical issues and code-switching to English, when applied to low-resource languages like Cantonese in Machine Translation (MT) scenarios. By addressing the unique linguistic and contextual challenges of Cantonese, we present a novel strategy to improve the understanding and translation capabilities of LLMs for Cantonese-to-Mandarin MT. Our strategy comprises three key components: (1) Syntax and Part-of-Speech (POS) fine-tuning, where we use the Universal Dependencies (UD) corpus to fine-tune LLM, focusing on the linguistic structures of Cantonese; (2) Specialized Cantonese to Mandarin sentence pairs, collected from diverse sources such as Cantonese grammar textbooks and manually translated sentences across various domains, to expose the model to a wide range of linguistic contexts; (3) Post-processing with additional LLMs, where we introduce additional LLMs to improve the initial translations, correcting Mandarin grammar and punctuation. Empirical evaluations on human-created test sets show that our proposed strategy improves translation performance and outperforms existing commercial translation models with at least 3 BLEU scores. Additionally, our strategy also benefits other LLMs and a reversed translation direction, demonstrating its generalization and effectiveness. |
sindhujan-etal-2025-llms | 2,025 | When LLMs Struggle: Reference-less Translation Evaluation for Low-resource Languages | This paper investigates the reference-less evaluation of machine translation for low-resource language pairs, known as quality estimation (QE). Segment-level QE is a challenging cross-lingual language understanding task that provides a quality score (0 -100) to the translated output. We comprehensively evaluate large language models (LLMs) in zero/few-shot scenarios and perform instruction fine-tuning using a novel prompt based on annotation guidelines. Our results indicate that prompt-based approaches are outperformed by the encoder-based fine-tuned QE models. Our error analysis reveals tokenization issues, along with errors due to transliteration and named entities, and argues for refinement in LLM pre-training for cross-lingual tasks. We release the data, and models trained publicly for further research. |
dmonte-etal-2025-machine | 2,025 | Does Machine Translation Impact Offensive Language Identification? The Case of Indo-Aryan Languages | The accessibility to social media platforms can be improved with the use of machine translation (MT). Non-standard features present in user-generated on social media content such as hashtags, emojis, and alternative spellings can lead to mistranslated instances by the MT systems. In this paper, we investigate the impact of MT on offensive language identification in Indo-Aryan languages. We use both original and MT datasets to evaluate the performance of various offensive language models. Our evaluation indicates that offensive language identification models achieve superior performance on original data than on MT data, and that the models trained on MT data identify offensive language more precisely on MT data than the models trained on original data. |
mahlaza-etal-2025-isizulu | 2,025 | IsiZulu noun classification based on replicating the ensemble approach for Runyankore | A noun`s class is a crucial component in NLP, because it governs agreement across the sentence in Niger Congo B (NCB) languages, among others. The phenomenon is ill-documented in most NCB languages, or in a non-reusable format, such as a printed dictionary subject to copyright restrictions. A promising approach by Byamugisha (2022) used a data-driven approach for Runyankore that combined syntax and semantics. The code and data are inaccessible however, and it remains to be seen whether it is suitable for other NCB languages. We aimed to reproduce Byamugisha`s experiment, but then for isiZulu. We conducted this as two independent experiments, so that we also could subject it to a meta-analysis. Results showed that it was reproducible only in part, mainly due to imprecision in the original description, and the current impossibility to generate the same kind of source data set generated from an existing grammar. The different choices made in attempting to reproduce the pipeline as well as differences in choice of training and test data had a large effect on the eventual accuracy of noun class disambiguation but could produce accuracies in the same range as for Runyankore: 80-85{\%}. |
zeinalipour-etal-2025-arabic | 2,025 | From Arabic Text to Puzzles: LLM-Driven Development of Arabic Educational Crosswords | We present an Arabic crossword puzzle generator from a given text that utilizes advanced language models such as GPT-4-Turbo, GPT-3.5-Turbo, and Llama3-8B-Instruct, specifically developed for educational purposes, this innovative generator leverages a meticulously compiled dataset named Arabic-Clue-Instruct with over 50,000 entries encompassing text, answers, clues, and categories. This dataset is intricately designed to aid in the generation of pertinent clues linked to specific texts and keywords within defined categories. This project addresses the scarcity of advanced educational tools tailored for the Arabic language, promoting enhanced language learning and cognitive development. By providing a culturally and linguistically relevant tool, our objective is to make learning more engaging and effective through gamification and interactivity. Integrating state-of-the-art artificial intelligence with contemporary learning methodologies, this tool can generate crossword puzzles from any given educational text, thereby facilitating an interactive and enjoyable learning experience. This tool not only advances educational paradigms but also sets a new standard in interactive and cognitive learning technologies. |
lal-etal-2025-hindi | 2,025 | Hindi Reading Comprehension: Do Large Language Models Exhibit Semantic Understanding? | In this study, we explore the performance of four advanced Generative AI models{---}GPT-3.5, GPT-4, Llama3, and HindiGPT, for the Hindi reading comprehension task. Using a zero-shot, instruction-based prompting strategy, we assess model responses through a comprehensive triple evaluation framework using the HindiRC dataset. Our framework combines (1) automatic evaluation using ROUGE, BLEU, BLEURT, METEOR, and Cosine Similarity; (2) rating-based assessments focussing on correctness, comprehension depth, and informativeness; and (3) preference-based selection to identify the best responses. Human ratings indicate that GPT-4 outperforms the other LLMs on all parameters, followed by HindiGPT, GPT-3.5, and then Llama3. Preference-based evaluation similarly placed GPT-4 (80{\%}) as the best model, followed by HindiGPT(74{\%}). However, automatic evaluation showed GPT-4 to be the lowest performer on n-gram metrics, yet the best performer on semantic metrics, suggesting it captures deeper meaning and semantic alignment over direct lexical overlap, which aligns with its strong human evaluation scores. This study also highlights that even though the models mostly address literal factual recall questions with high precision, they still face the challenge of specificity and interpretive bias at times. |
perera-sumanathilaka-2025-machine | 2,025 | Machine Translation and Transliteration for Indo-Aryan Languages: A Systematic Review | This systematic review paper provides an overview of recent machine translation and transliteration developments for Indo-Aryan languages spoken by a large population across South Asia. The paper examines advancements in translation and transliteration systems for a few language pairs which appear in recently published papers. The review summarizes the current state of these technologies, providing a worthful resource for anyone who is doing research in these fields to understand and find existing systems and techniques for translation and transliteration. |
mutsaddi-etal-2025-bertopic | 2,025 | BERTopic for Topic Modeling of Hindi Short Texts: A Comparative Study | As short text data in native languages like Hindi increasingly appear in modern media, robust methods for topic modeling on such data have gained importance. This study investigates the performance of BERTopic in modeling Hindi short texts, an area that has been under-explored in existing research. Using contextual embeddings, BERTopic can capture semantic relationships in data, making it potentially more effective than traditional models, especially for short and diverse texts. We evaluate BERTopic using 6 different document embedding models and compare its performance against 8 established topic modeling techniques, such as Latent Dirichlet Allocation (LDA), Non-negative Matrix Factorization (NMF), Latent Semantic Indexing (LSI), Additive Regularization of Topic Models (ARTM), Probabilistic Latent Semantic Analysis (PLSA), Embedded Topic Model (ETM), Combined Topic Model (CTM), and Top2Vec. The models are assessed using coherence scores across a range of topic counts. Our results reveal that BERTopic consistently outperforms other models in capturing coherent topics from short Hindi texts. |
amin-etal-2025-evaluating | 2,025 | Evaluating Structural and Linguistic Quality in Urdu DRS Parsing and Generation through Bidirectional Evaluation | Evaluating Discourse Representation Structure (DRS)-based systems for semantic parsing (Text-to-DRS) and generation (DRS-to-Text) poses unique challenges, particularly in low-resource languages like Urdu. Traditional metrics often fall short, focusing either on structural accuracy or linguistic quality, but rarely capturing both. To address this limitation, we introduce two complementary evaluation methodologies{---}Parse-Generate (PARS-GEN) and Generate-Parse (GEN-PARS){---}designed for a more comprehensive assessment of DRS-based systems. PARS-GEN evaluates the parsing process by converting DRS outputs back to the text, revealing linguistic nuances often missed by structure-focused metrics like SMATCH. Conversely, GEN-PARS assesses text generation by converting generated text into DRS, providing a semantic perspective that complements surface-level metrics such as BLEU, METEOR, and BERTScore. Using the Parallel Meaning Bank (PMB) dataset, we demonstrate our methodology across Urdu, uncovering unique insights into Urdu`s structural and linguistic interplay. Findings show that traditional metrics frequently overlook the complexity of linguistic and semantic fidelity, especially in low-resource languages. Our dual approach offers a robust framework for evaluating DRS-based systems, enhancing semantic parsing and text generation quality. |
goel-sadat-2025-studying | 2,025 | Studying the Effect of Hindi Tokenizer Performance on Downstream Tasks | This paper deals with a study on the effect of training data size and tokenizer performance for Hindi language on the eventual downstream model performance and comprehension. Multiple monolingual Hindi tokenizers are trained for large language models such as BERT and intrinsic and extrinsic evaluations are performed on multiple Hindi datasets. The objective of this study is to understand the precise effects of tokenizer performance on downstream task performance to gain insight on how to develop better models for low-resource languages. |
joshi-etal-2025-adapting | 2,025 | Adapting Multilingual LLMs to Low-Resource Languages using Continued Pre-training and Synthetic Corpus: A Case Study for Hindi LLMs | Multilingual LLMs support a variety of languages; however, their performance is suboptimal for low-resource languages. In this work, we emphasize the importance of continued pre-training of multilingual LLMs and the use of translation-based synthetic pre-training corpora for improving LLMs in low-resource languages. We conduct our study in the context of the low-resource Indic language Hindi. We introduce Nemotron-Mini-Hindi 4B, a bilingual SLM supporting both Hindi and English, based on Nemotron-Mini 4B. The model is trained using a mix of real and synthetic Hindi + English tokens, with continuous pre-training performed on 400B tokens. We demonstrate that both the base and instruct models achieve state-of-the-art results on Hindi benchmarks while remaining competitive on English tasks. Additionally, we observe that the continued pre-training approach enhances the model`s overall factual accuracy. |
parida-etal-2025-ovqa | 2,025 | OVQA: A Dataset for Visual Question Answering and Multimodal Research in Odia Language | This paper introduces OVQA, the first multimodal dataset designed for visual question-answering (VQA), visual question elicitation (VQE), and multimodal research for the low-resource Odia language. The dataset was created by manually translating 6,149 English question-answer pairs, each associated with 6,149 unique images from the Visual Genome dataset. This effort resulted in 27,809 English-Odia parallel sentences, ensuring a semantic match with the corresponding visual information. Several baseline experiments were conducted on the dataset, including visual question answering and visual question elicitation. The dataset is the first VQA dataset for the low-resource Odia language and will be released for multimodal research purposes and also help researchers extend for other low-resource languages. |
sritharan-thayasivam-2025-advancing | 2,025 | Advancing Multilingual Speaker Identification and Verification for Indo-Aryan and Dravidian Languages | Multilingual speaker identification and verification is a challenging task, especially for languages with diverse acoustic and linguistic features such as Indo-Aryan and Dravidian languages. Previous models have struggled to generalize across multilingual environments, leading to significant performance degradation when applied to multiple languages. In this paper, we propose an advanced approach to multilingual speaker identification and verification, specifically designed for Indo-Aryan and Dravidian languages. Empirical results on the Kathbath dataset show that our approach significantly improves speaker identification accuracy, reducing the performance gap between monolingual and multilingual systems from 15{\%} to just 1{\%}. Additionally, our model reduces the equal error rate for speaker verification from 15{\%} to 5{\%} in noisy conditions. Our method demonstrates strong generalization capabilities across diverse languages, offering a scalable solution for multilingual voice-based biometric systems. |
bandaranayake-usoof-2025-sentiment | 2,025 | Sentiment Analysis of Sinhala News Comments Using Transformers | Sentiment analysis has witnessed significant advancements with the emergence of deep learning models such as transformer models. Transformer models adopt the mechanism of self-attention and have achieved state-of-the-art performance across various natural language processing (NLP) tasks, including sentiment analysis. However, limited studies are exploring the application of these recent advancements in sentiment analysis of Sinhala text. This study addresses this research gap by employing transformer models such as BERT, DistilBERT, RoBERTa, and XLM-RoBERTa (XLM-R) for sentiment analysis of Sinhala News comments. This study was conducted for 4 classes: positive, negative, neutral, and conflict, as well as for 3 classes: positive, negative, and neutral. It revealed that the XLM-R-large model outperformed the other four models, and the transformer models used in previous studies for the Sinhala language. The XLM-R-large model achieved an accuracy of 65.84{\%} and a macro-F1 score of 62.04{\%} for sentiment analysis with four classes and an accuracy of 75.90{\%} and a macro-F1 score of 72.31{\%} for three classes. |
debnath-etal-2025-exmute | 2,025 | ExMute: A Context-Enriched Multimodal Dataset for Hateful Memes | In this paper, we introduce ExMute, an extended dataset for classifying hateful memes that incorporates critical contextual information, addressing a significant gap in existing resources. Building on a previous dataset of 4,158 memes without contextual annotations, ExMute expands the collection by adding 2,041 new memes and providing comprehensive annotations for all 6,199 memes. Each meme is labeled across six defined contexts with language markers indicating code-mixing, code-switching, and Bengali captions, enhancing its value for linguistic and cultural research. These memes are systematically labeled across six contexts: religion, politics, celebrity, male, female, and others, facilitating a more nuanced understanding of meme content and intent. To evaluate ExMute, we apply state-of-the-art textual, visual, and multimodal approaches, leveraging models including BanglaBERT, Visual Geometry Group (VGG), Inception, ResNet, and Vision Transformer (ViT). Our experiments show that our custom LSTM-based attention-based textual model achieves an accuracy of 0.60, while VGG-based visual models reach up to 0.63. Multimodal models, which combine visual and textual features, consistently achieve accuracy scores of around 0.64, demonstrating the dataset`s robustness for advancing multimodal classification tasks. ExMute establishes a valuable benchmark for future NLP research, particularly in low-resource language settings, highlighting the importance of context-aware labeling in improving classification accuracy and reducing bias. |
kumar-roy-2025-studying | 2,025 | Studying the capabilities of Large Language Models in solving Combinatorics Problems posed in Hindi | There are serious attempts at improving the mathematical acumen of LLMs in questions posed in English. In India, where a large fraction of the students study in regional languages, there is a need to assess and improve these state-of-the-art LLMs in their reasoning abilities in regional languages as well. As Hindi is a language predominantly used in India, this study proposes a new dataset on mathematical combinatorics problems consisting of a parallel corpus of problems in English and Hindi collected from NCERT textbooks. We evaluate the {\textquotedblleft}raw{\textquotedblright} single-shot capabilities of these LLMs in solving problems posed in Hindi. Then we apply a chain-of-thought approach to evaluate the improvement in the abilities of the LLMs at solving combinatorics problems posed in Hindi. Our study reveals that while smaller LLMs like LLaMa3-8B shows a significant drop in performance when questions are posed in Hindi, versus questions posed in English, larger LLMs like GPT4-turbo shows excellent capabilities at solving problems posed in Hindi, almost at par its abilities in English. We make two primary inferences from our study: (1) large models like GPT4 can be readily deployed in schools where Hindi is the primary language of study, especially in rural India; (2) there is a need to improve the multilingual capabilities of smaller models. As these smaller open-source models can be deployed on not so expensive GPUs, it is easier for schools to provide these models to the students, and hence, the latter is an important direction for future research. |
shibu-etal-2025-scarcity | 2,025 | From Scarcity to Capability: Empowering Fake News Detection in Low-Resource Languages with LLMs | The rapid spread of fake news presents a significant global challenge, particularly in low-resource languages like Bangla, which lack adequate datasets and detection tools. Although manual fact-checking is accurate, it is expensive and slow to prevent the dissemination of fake news. Addressing this gap, we introduce BanFakeNews-2.0, a robust dataset to enhance Bangla fake news detection. This version includes 11,700 additional, meticulously curated fake news articles validated from credible sources, creating a proportional dataset of 47,000 authentic and 13,000 fake news items across 13 categories. In addition, we created a manually curated independent test set of 460 fake and 540 authentic news items for rigorous evaluation. We invest efforts in collecting fake news from credible sources and manually verified while preserving the linguistic richness. We develop a benchmark system utilizing transformer-based architectures, including fine-tuned Bidirectional Encoder Representations from Transformers variants (F1-87{\%}) and Large Language Models with Quantized Low-Rank Approximation (F1-89{\%}), that significantly outperforms traditional methods. BanFakeNews-2.0 offers a valuable resource to advance research and application in fake news detection for low-resourced languages. We publicly release our dataset and model on GitHub to foster research in this direction. |
zhao-etal-2025-enhancing | 2,025 | Enhancing Participatory Development Research in South Asia through LLM Agents System: An Empirically-Grounded Methodological Initiative from Field Evidence in Sri Lankan | The integration of artificial intelligence into development research methodologies offers unprecedented opportunities to address persistent challenges in participatory research, particularly in linguistically diverse regions like South Asia. Drawing on empirical implementation in Sri Lanka`s Sinhala-speaking communities, this study presents a methodological framework designed to transform participatory development research in the multilingual context of Sri Lanka`s flood-prone Nilwala River Basin. Moving beyond conventional translation and data collection tools, the proposed framework leverages a multi-agent system architecture to redefine how data collection, analysis, and community engagement are conducted in linguistically and culturally complex research settings. This structured, agent-based approach facilitates participatory research that is both scalable and adaptive, ensuring that community perspectives remain central to research outcomes. Field experiences underscore the immense potential of LLM-based systems in addressing long-standing issues in development research across resource-limited regions, delivering both quantitative efficiencies and qualitative improvements in inclusivity. At a broader methodological level, this research advocates for AI-driven participatory research tools that prioritize ethical considerations, cultural sensitivity, and operational efficiency. It highlights strategic pathways for deploying AI systems to reinforce community agency and equitable knowledge generation, offering insights that could inform broader research agendas across the Global South. |
kancharla-etal-2025-identifying | 2,025 | Identifying Aggression and Offensive Language in Code-Mixed Tweets: A Multi-Task Transfer Learning Approach | The widespread use of social media has contributed to the increase in hate speech and offensive language, impacting people of all ages. This issue is particularly difficult to address when the text is in a code-mixed language. Twitter is commonly used to express opinions in code-mixed language. In this paper, we introduce a novel Multi-Task Transfer Learning (MTTL) framework to detect aggression and offensive language. By focusing on the dual facets of cyberbullying, aggressiveness and offensiveness, our model leverages the MTTL approach to enhance the performance of the model on the aggression and offensive language detection. Results show that our Multi-Task Transfer Learning (MTTL) setup significantly enhances the performance of state-of-the-art pretrained language models, BERT, RoBERTa, and Hing-RoBERTa for Hindi-English code-mixed data from Twitter. |
kumar-etal-2025-team | 2,025 | Team IndiDataMiner at IndoNLP 2025: Hindi Back Transliteration - Roman to Devanagari using LLaMa | The increasing use of Romanized typing for Indo-Aryan languages on social media poses challenges due to its lack of standardization and loss of linguistic richness. To address this, we propose a sentence-level back-transliteration approach using the LLaMa 3.1 model for Hindi. Leveraging fine-tuning with the Dakshina dataset, our approach effectively resolves ambiguities in Romanized Hindi text, offering a robust solution for converting it into the native Devanagari script. |
perera-etal-2025-indonlp | 2,025 | IndoNLP 2025 Shared Task: Romanized Sinhala to Sinhala Reverse Transliteration Using BERT | The Romanized text has become popular with the growth of digital communication platforms, largely due to the familiarity with English keyboards. In Sri Lanka, Romanized Sinhala, commonly referred to as {\textquotedblleft}Singlish{\textquotedblright} is widely used in digital communications. This paper introduces a novel context-aware back-transliteration system designed to address the ad-hoc typing patterns and lexical ambiguity inherent in Singlish. The proposed system com bines dictionary-based mapping for Singlish words, a rule-based transliteration for out of-vocabulary words and a BERT-based language model for addressing lexical ambiguities. Evaluation results demonstrate the robustness of the proposed approach, achieving high BLEU scores along with low Word Error Rate (WER) and Character Error Rate (CER) across test datasets. This study provides an effective solution for Romanized Sinhala back-transliteration and establishes the foundation for improving NLP tools for similar low-resourced languages. |
kazi-etal-2025-crossing | 2,025 | Crossing Language Boundaries: Evaluation of Large Language Models on Urdu-English Question Answering | This study evaluates the question-answering capabilities of Large Language Models (LLMs) in Urdu, addressing a critical gap in low-resource language processing. Four models GPT-4, mBERT, XLM-R, and mT5 are assessed across monolingual, cross-lingual, and mixed-language settings using the UQuAD1.0 and SQuAD2.0 datasets. Results reveal significant performance gaps between English and Urdu processing, with GPT-4 achieving the highest F1 scores (89.1{\%} in English, 76.4{\%} in Urdu) while demonstrating relative robustness in cross-lingual scenarios. Boundary detection and translation mismatches emerge as primary challenges, particularly in cross-lingual settings. The study further demonstrates that question complexity and length significantly impact performance, with factoid questions yielding 14.2{\%} higher F1 scores compared to complex questions. These findings establish important benchmarks for enhancing LLM performance in low-resource languages and identify key areas for improvement in multilingual question-answering systems. |
das-etal-2025-investigating | 2,025 | Investigating the Effect of Backtranslation for Indic Languages | Neural machine translation (NMT) is becoming increasingly popular as an effective method of automated language translation. However, due to a scarcity of training datasets, its effectiveness is limited when used with low-resource languages, such as Indian Languages (ILs). The lack of parallel datasets in Natural Language Processing (NLP) makes it difficult to investigate many ILs for Machine Translation (MT). A data augmentation approach such as Backtranslation (BT) can be used to enhance the size of the training dataset. This paper presents the development of a NMT model for ILs within the context of a MT system. To address the issue of data scarcity, the paper examines the effectiveness of a BT approach for ILs that uses both monolingual and parallel datasets. Experimental results reveal that while the BT has improved the model`s performance, however, it is not as significant as expected. It has also been observed that, even though the English-ILs and ILs-English models are trained on the same dataset, the ILs-English models perform better in all evaluation metrics. The reason for this is that ILs frequently differ in sentence structure, word order, and morphological richness from English. The paper also includes error analysis for translations between languages that were utilized in experiments utilizing the Multidimensional Quality Metrics (MQM) framework. |
de-mel-etal-2025-sinhala | 2,025 | Sinhala Transliteration: A Comparative Analysis Between Rule-based and Seq2Seq Approaches | Due to reasons of convenience and lack of tech literacy, transliteration (i.e., Romanizing native scripts instead of using localization tools) is eminently prevalent in the context of low-resource languages such as Sinhala, which have their own writing script. In this study, our focus is on Romanized Sinhala transliteration. We propose two methods to address this problem: Our baseline is a rule-based method, which is then compared against our second method where we approach the transliteration problem as a sequence-to-sequence task akin to the established Neural Machine Translation (NMT) task. For the latter, we propose a Transformer based Encode-Decoder solution. We witnessed that the Transformer-based method could grab many ad-hoc patterns within the Romanized scripts compared to the rule-based method. |
baiju-etal-2025-romanized | 2,025 | Romanized to Native Malayalam Script Transliteration Using an Encoder-Decoder Framework | In this work, we present the development of a reverse transliteration model to convert romanized Malayalam to native script using an encoder-decoder framework built with attention-based bidirectional Long Short Term Memory (Bi-LSTM) architecture. To train the model, we have used curated and combined collection of 4.3 million transliteration pairs derived from publicly available Indic language translitertion datasets, Dakshina and Aksharantar. We evaluated the model on two different test dataset provided by IndoNLP-2025-Shared-Task that contain, (1) General typing patterns and (2) Adhoc typing patterns, respectively. On the Test Set-1, we obtained a character error rate (CER) of 7.42{\%}. However upon Test Set-2, with adhoc typing patterns, where most vowel indicators are missing, our model gave a CER of 22.8{\%}. |
saini-nayak-2025-effective | 2,025 | Effective Modeling of Generative Framework for Document-level Relational Triple Extraction | Document-level relation triple extraction (DocRTE) is a complex task that involves three key sub-tasks: entity mention extraction, entity clustering, and relation triple extraction. Past work has applied discriminative models to address these three sub-tasks, either by training them sequentially in a pipeline fashion or jointly training them. However, while end-to-end discriminative or generative models have proven effective for sentence-level relation triple extraction, they cannot be trivially extended to the document level, as they only handle relation extraction without addressing the remaining two sub-tasks, entity mention extraction or clustering. In this paper, we propose a three-stage generative framework leveraging a pre-trained BART model to address all three tasks required for document-level relation triple extraction. Tested on the widely used DocRED dataset, our approach outperforms previous generative methods and achieves competitive performance against discriminative models. |
martynova-etal-2025-learn | 2,025 | Learn Together: Joint Multitask Finetuning of Pretrained KG-enhanced LLM for Downstream Tasks | Recent studies have shown that a knowledge graph (KG) can enhance text data by providing structured background knowledge, which can significantly improve the language understanding skills of the LLM. Besides, finetuning of such models shows solid results on commonsense reasoning benchmarks. In this work, we introduce expandable Joint Multitask Finetuning on Pretrained KG-enchanced LLM approach for Question Answering (QA), Machine Reading Comprehension (MRC) and Knowledge Graph Question Answering (KGQA) tasks. Extensive experiments show competitive performance of joint finetuning QA+MRC+KGQA over single task approach with a maximum gain of 30{\%} accuracy. |
jamshidi-chali-2025-gnet | 2,025 | GNET-QG: Graph Network for Multi-hop Question Generation | Multi-hop question generation is a challenging task in natural language processing (NLP) that requires synthesizing information from multiple sources. We propose GNET-QG, a novel approach that integrates Graph Attention Networks (GAT) with sequence-to-sequence models, enabling structured reasoning over multiple information sources to generate complex questions. Our experiments demonstrate that GNET-QG outperforms previous state-of-the-art models across several evaluation metrics, particularly excelling in METEOR, showing its effectiveness in enhancing machine reasoning capabilities. |
mahalingam-etal-2025-sketch | 2,025 | SKETCH: Structured Knowledge Enhanced Text Comprehension for Holistic Retrieval | This paper discusses about the SKETCH approach which enhances text retrieval and context relevancy on large corpuses compared to the traditional baseline methods. The abstract attached below discusses this further. Abstract: Retrieval-Augmented Generation (RAG) systems have become pivotal in leveraging vast corpora to generate informed and contextually relevant responses, notably reducing hallucinations in Large Language Models. Despite significant advancements, these systems struggle to efficiently process and retrieve information from large datasets while maintaining a comprehensive understanding of the context. This paper introduces SKETCH, a novel methodology that enhances the RAG retrieval process by integrating semantic text retrieval with knowledge graphs, thereby merging structured and unstructured data for a more holistic comprehension. SKETCH, demonstrates substantial improvements in retrieval performance and maintains superior context integrity compared to traditional methods. Evaluated across four diverse datasets: QuALITY, QASPER, NarrativeQA, and Italian Cuisine{---}SKETCH consistently outperforms baseline approaches on key RAGAS metrics such as answer relevancy, faithfulness, context precision and context recall. Notably, on the Italian Cuisine dataset, SKETCH achieved an answer relevancy of 0.94 and a context precision of 0.99, representing the highest performance across all evaluated metrics. These results highlight SKETCH`s capability in delivering more accurate and contextually relevant responses, setting new benchmarks for future retrieval systems. |
iarosh-etal-2025-reducing | 2,025 | On Reducing Factual Hallucinations in Graph-to-Text Generation Using Large Language Models | Recent work in Graph-to-Text generation has achieved impressive results, but it still suffers from hallucinations in some cases, despite extensive pretraining stages and various methods for working with graph data. While the commonly used metrics for evaluating the quality of Graph-to-Text models show almost perfect results, it makes it challenging to compare different approaches. This paper demonstrates the challenges of recent Graph-to-Text systems in terms of hallucinations and proposes a simple yet effective approach to using a general LLM, which has shown state-of-the-art results and reduced the number of factual hallucinations. We provide step-by-step instructions on how to develop prompts for language models and a detailed analysis of potential factual errors in the generated text. |
barry-etal-2025-graphrag | 2,025 | GraphRAG: Leveraging Graph-Based Efficiency to Minimize Hallucinations in LLM-Driven RAG for Finance Data | This study explores the integration of graph-based methods into Retrieval-Augmented Generation (RAG) systems to enhance efficiency, reduce hallucinations, and improve explainability, with a particular focus on financial and regulatory document retrieval. We propose two strategies{---}FactRAG and HybridRAG{---}which leverage knowledge graphs to improve RAG performance. Experiments conducted using Finance Bench, a benchmark for AI in finance, demonstrate that these approaches achieve a 6{\%} reduction in hallucinations and an 80{\%} decrease in token usage compared to conventional RAG methods. Furthermore, we evaluate HybridRAG by comparing the Digital Operational Resilience Act (DORA) from the European Union with the Federal Financial Institutions Examination Council (FFIEC) guidelines from the United States. The results reveal a significant improvement in computational efficiency, reducing contradiction detection complexity from $O(n^2)$ to $O(k \cdot n)${---}where $n$ is the number of chunks{---}and a remarkable 734-fold decrease in token consumption. Graph-based retrieval methods can improve the efficiency and cost-effectiveness of large language model (LLM) applications, though their performance and token usage depend on the dataset, knowledge graph design, and retrieval task. |
eldessouky-etal-2025-structured | 2,025 | Structured Knowledge meets GenAI: A Framework for Logic-Driven Language Models | Large Language Models (LLMs) excel at generating fluent text but struggle with context sensitivity, logical reasoning, and personalization without extensive fine-tuning. This paper presents a logical modulator: an adaptable communication layer between Knowledge Graphs (KGs) and LLMs as a way to address these limitations. Unlike direct KG-LLM integrations, our modulator is domain-agnostic and incorporates logical dependencies and commonsense reasoning in order to achieve contextual personalization. By enhancing KG interaction, this method will produce linguistically coherent and logically sound outputs, increasing interpretability and reliability in generative AI. |
mecharnia-daquin-2025-performance | 2,025 | Performance and Limitations of Fine-Tuned LLMs in SPARQL Query Generation | Generative AI has simplified information access by enabling natural language-driven interactions between users and automated systems. In particular, Question Answering (QA) has emerged as a key application of AI, facilitating efficient access to complex information through dialogue systems and virtual assistants. The Large Language Models (LLMs) combined with Knowledge Graphs (KGs) have further enhanced QA systems, allowing them to not only correctly interpret natural language but also retrieve precise answers from structured data sources such as Wikidata and DBpedia. However, enabling LLMs to generate machine-readable SPARQL queries from natural language questions (NLQs) remains challenging, particularly for complex questions. In this study, we present experiments in fine-tuning LLMs for the task of NLQ-to-SPARQL transformation. We rely on benchmark datasets for training and testing the fine-tuned models, generating queries directly from questions written in English (without further processing of the input or output). By conducting an analytical study, we examine the effectiveness of each model, as well as the limitations associated with using fine-tuned LLMs to generate SPARQL. |
dong-etal-2025-refining | 2,025 | Refining Noisy Knowledge Graph with Large Language Models | Knowledge graphs (KGs) represent structured real-world information composed by triplets of head entity, relation, and tail entity. These graphs can be constructed automatically from text or manually curated. However, regardless of the construction method, KGs often suffer from misinformation, incompleteness, and noise, which hinder their reliability and utility. This study addresses the challenge of noisy KGs, where incorrect or misaligned entities and relations degrade graph quality. Leveraging recent advancements in large language models (LLMs) with strong capabilities across diverse tasks, we explore their potential to detect and refine noise in KGs. Specifically, we propose a novel method, LLM{\_}sim, to enhance the detection and refinement of noisy triples. Our results confirm the effectiveness of this approach in elevating KG quality in noisy environments. Additionally, we apply our proposed method to Knowledge Graph Completion (KGC), a downstream KG task that aims to predict missing links and improve graph completeness. Traditional KGC methods assume that KGs are noise-free, which is unrealistic in practical scenarios. Our experiments analyze the impact of varying noise levels on KGC performance, revealing that LLMs can mitigate noise by identifying and refining incorrect entries, thus enhancing KG quality. |
regino-dos-reis-2025-llms | 2,025 | Can LLMs be Knowledge Graph Curators for Validating Triple Insertions? | As Knowledge Graphs (KGs) become central to modern applications, automated methods for validating RDF triples before insertion into these graphs are essential. The complexity and scalability challenges in manual validation processes have led researchers to explore Large Language Models (LLMs) as potential automated validators. This study investigates the feasibility of using LLMs to validate RDF triples by focusing on four distinct and complementary validation tasks: class and property alignment, URI standardization, semantic consistency, and syntactic correctness. We propose a systematic validation method that uses prompts to guide LLMs through each stage of the triple evaluation of the RDF. In our experiments, four models are evaluated across these tasks. Our results reveal that more advanced models like Llama-3-70B-Instruct offer superior accuracy and consistency. Our findings emphasize the practical open challenges of deploying LLMs in real-world RDF validation scenarios, including domain generalization, semantic drift, and the need for human-in-the-loop interventions. This investigation advances the research on the refinement and integration of LLM-based RDF validation techniques into KG management workflows. |
ozsoy-etal-2025-text2cypher | 2,025 | Text2Cypher: Bridging Natural Language and Graph Databases | Knowledge graphs use nodes, relationships, and properties to represent arbitrarily complex data. When stored in a graph database, the Cypher query language enables efficient modeling and querying of knowledge graphs. However, using Cypher requires specialized knowledge, which can present a challenge for non-expert users. Our work Text2Cypher aims to bridge this gap by translating natural language queries into Cypher query language and extending the utility of knowledge graphs to non-technical expert users. While large language models (LLMs) can be used for this purpose, they often struggle to capture complex nuances, resulting in incomplete or incorrect outputs. Fine-tuning LLMs on domain-specific datasets has proven to be a more promising approach, but the limited availability of high-quality, publicly available Text2Cypher datasets makes this challenging. In this work, we show how we combined, cleaned and organized several publicly available datasets into a total of 44,387 instances, enabling effective fine-tuning and evaluation. Models fine-tuned on this dataset showed significant performance gains, with improvements in Google-BLEU and Exact Match scores over baseline models, highlighting the importance of high-quality datasets and fine-tuning in improving Text2Cypher performance. |
kumar-etal-2025-kgfakenet | 2,025 | KGFakeNet: A Knowledge Graph-Enhanced Model for Fake News Detection | The proliferation of fake news on social media has intensified the spread of misinformation, promoting societal biases, hate, and violence. While recent advancements in Generative AI (GenAI), particularly large language models (LLMs), have shown promise, these models often need more structured representation for accurate verification, as they rely on pre-trained data patterns without access to real-time or validated information. This study presents a framework that utilizes Open Information Extractor 6 (OpenIE6) to extract triplet relationships (subject-predicate-object) from statements and justifications to compute the cosine similarity between the Knowledge Graphs (KGs) of the statements and their supporting justification to precisely measure the relevance and alignment between them. This similarity feature is integrated with an attention mechanism over GenAI-generated embeddings to enhance the model`s ability to capture semantic features accurately. In addition, a Multi-Layer Perceptron (MLP) classifier is employed to integrate all features, resulting in a 4{\%} improvement in accuracy and a 5{\%} increase in F1-score over state-of-the-art LLM-based approaches. |
toshevska-etal-2025-style | 2,025 | Style Knowledge Graph: Augmenting Text Style Transfer with Knowledge Graphs | Text style transfer is the task of modifying the stylistic attributes of a given text while preserving its original meaning. This task has also gained interest with the advent of large language models. Although knowledge graph augmentation has been explored in various tasks, its potential for enhancing text style transfer has received limited attention. This paper proposes a method to create a Style Knowledge Graph (SKG) to facilitate and improve text style transfer. The SKG captures words, their attributes, and relations in a particular style, that serves as a knowledge resource to augment text style transfer. We conduct baseline experiments to evaluate the effectiveness of the SKG for augmenting text style transfer by incorporating relevant parts from the SKG in the prompt. The preliminary results demonstrate its potential for enhancing content preservation and style transfer strength in text style transfer tasks, while the results on fluency indicate promising outcomes with some room for improvement. We hope that the proposed SKG and the initial experiments will inspire further research in the field. |
kamaladdini-ezzabady-benamara-2025-entity | 2,025 | Entity Quality Enhancement in Knowledge Graphs through LLM-based Question Answering | Most models for triple extraction from texts primarily focus on named entities. However, real-world applications often comprise non-named entities that pose serious challenges for entity linking and disambiguation. We focus on these entities and propose the first LLM-based entity revision framework to improve the quality of extracted triples via a multi-choice question-answering mechanism. When evaluated on two benchmark datasets, our results show a significant improvement, thereby generating more reliable triples for knowledge graphs. |
kavas-etal-2025-multilingual | 2,025 | Multilingual Skill Extraction for Job Vacancy--Job Seeker Matching in Knowledge Graphs | In the modern labor market, accurate matching of job vacancies with suitable candidate CVs is critical. We present a novel multilingual knowledge graph-based framework designed to enhance the matching by accurately extracting the skills requested by a job and provided by a job seeker in a multilingual setting and aligning them via the standardized skill labels of the European Skills, Competences, Qualifications and Occupations (ESCO) taxonomy. The proposed framework employs a combination of state-of-the-art techniques to extract relevant skills from job postings and candidate experiences. These extracted skills are then filtered and mapped to the ESCO taxonomy and integrated into a multilingual knowledge graph that incorporates hierarchical relationships and cross-linguistic variations through embeddings. Our experiments demonstrate a significant improvement of the matching quality compared to the state of the art. |
creo-pudasaini-2025-silverspeak | 2,025 | SilverSpeak: Evading AI-Generated Text Detectors using Homoglyphs | The advent of Large Language Models (LLMs) has enabled the generation of text that increasingly exhibits human-like characteristics. As the detection of such content is of significant importance, substantial research has been conducted with the objective of developing reliable AI-generated text detectors. These detectors have demonstrated promising results on test data, but recent research has revealed that they can be circumvented by employing different techniques. In this paper, we present homoglyph-based attacks ({\textquoteleft}A' {\textrightarrow} Cyrillic {\textquoteleft}{\CYRA}') as a means of circumventing existing detectors. We conduct a comprehensive evaluation to assess the effectiveness of these attacks on seven detectors, including ArguGPT, Binoculars, DetectGPT, Fast-DetectGPT, Ghostbuster, OpenAI`s detector, and watermarking techniques, on five different datasets. Our findings demonstrate that homoglyph-based attacks can effectively circumvent state-of-the-art detectors, leading them to classify all texts as either AI-generated or human-written (decreasing the average Matthews Correlation Coefficient from 0.64 to -0.01). Through further examination, we extract the technical justification underlying the success of the attacks, which varies across detectors. Finally, we discuss the implications of these findings and potential defenses against such attacks. |
moessner-adel-2025-human | 2,025 | Human vs. AI: A Novel Benchmark and a Comparative Study on the Detection of Generated Images and the Impact of Prompts | With the advent of publicly available AI-based text-to-image systems, the process of creating photorealistic but fully synthetic images has been largely democratized. This can pose a threat to the public through a simplified spread of disinformation. Machine detectors and human media expertise can help to differentiate between AI-generated (fake) and real images and counteract this danger. Although AI generation models are highly prompt-dependent, the impact of the prompt on the fake detection performance has rarely been investigated yet. This work therefore examines the influence of the prompt`s level of detail on the detectability of fake images, both with an AI detector and in a user study. For this purpose, we create a novel dataset, COCOXGEN, which consists of real photos from the COCO dataset as well as images generated with SDXL and Fooocus using prompts of two standardized lengths. Our user study with 200 participants shows that images generated with longer, more detailed prompts are detected significantly more easily than those generated with short prompts. Similarly, an AI-based detection model achieves better performance on images generated with longer prompts. However, humans and AI models seem to pay attention to different details, as we show in a heat map analysis. |
baradia-etal-2025-mirror | 2,025 | Mirror Minds : An Empirical Study on Detecting LLM-Generated Text via LLMs | The use of large language models (LLMs) is inevitable in text generation. LLMs are intelligent and slowly replacing the search engines. LLMs became the de facto choice for conversation, knowledge extraction, and brain storming. This study focuses on a question: {\textquoteleft}Can we utilize the generative capabilities of LLMs to detect AI-generated content?' We present a methodology and empirical results on four publicly available data sets. The result shows, with 90{\%} accuracy it is possible to detect AI-generated content by a zero-shot detector utilizing multiple LLMs. |
pudasaini-etal-2025-benchmarking | 2,025 | Benchmarking AI Text Detection: Assessing Detectors Against New Datasets, Evasion Tactics, and Enhanced LLMs | The rapid advancement of Large Language Models (LLMs), such as GPT-4, has sparked concerns regarding academic misconduct, misinformation, and the erosion of originality. Despite the growing number of AI detection tools, their effectiveness is often undermined by sophisticated evasion tactics and the continuous evolution of LLMs. This research benchmarks the performance of leading AI detectors, including OpenAI Detector, RADAR, and ArguGPT, across a variety of text domains, evaded content, and text generated by cutting-edge LLMs. Our experiments reveal that current detection models show considerable unreliability in real-world scenarios, particularly when tested against diverse data domains and novel evasion strategies. The study underscores the need for enhanced robustness in detection systems and provides valuable insights into areas of improvement for these models. Additionally, this work lays the groundwork for future research by offering a comprehensive evaluation of existing detectors under challenging conditions, fostering a deeper understanding of their limitations. The experimental code and datasets are publicly available for further benchmarking on Github. |
kindji-etal-2025-cross | 2,025 | Cross-table Synthetic Tabular Data Detection | Detecting synthetic tabular data is essential to prevent the distribution of false or manipulated datasets that could compromise data-driven decision-making. This study explores whether synthetic tabular data can be reliably identified {\textquotedblleft}in the wild{\textquotedblright}{---}meaning across different generators, domains, and table formats. This challenge is unique to tabular data, where structures (such as number of columns, data types, and formats) can vary widely from one table to another. We propose three cross-table baseline detectors and four distinct evaluation protocols, each corresponding to a different level of {\textquotedblleft}wildness{\textquotedblright}. Our very preliminary results confirm that cross-table adaptation is a challenging task. |
mcgovern-etal-2025-large | 2,025 | Your Large Language Models are Leaving Fingerprints | It has been shown that fine-tuned transformers and other supervised detectors are effective for distinguishing between human and machine-generated texts in non-adversarial settings, but we find that even simple classifiers on top of n-gram and part-of-speech features can achieve very robust performance on both in- and out-of-domain data. To understand how this is possible, we analyze machine-generated output text in four datasets, finding that LLMs possess unique fingerprints that manifest as slight differences in the frequency of certain lexical and morphosyntactic features. We show how to visualize such fingerprints, describe how they can be used to detect machine-generated text and find that they are even robust across text domains. We find that fingerprints are often persistent across models in the same model family (e.g. 13B parameter LLaMA`s fingerprint is similar to that of 65B parameter LLaMA) and that while a detector trained on text from one model can easily recognize text generated by a model in the same family, it struggles to detect text generated by an unrelated model. |
rathi-etal-2025-gpt | 2,025 | GPT-4 is Judged More Human than Humans in Displaced and Inverted Turing Tests | Everyday AI detection requires differentiating between humans and AI in informal, online conversations. At present, human users most often do not interact directly with bots but instead read their conversations with other humans. We measured how well humans and large language models can discriminate using two modified versions of the Turing test: inverted and displaced. GPT-3.5, GPT-4, and displaced human adjudicators judged whether an agent was human or AI on the basis of a Turing test transcript. We found that both AI and displaced human judges were less accurate than interactive interrogators, with below chance accuracy overall. Moreover, all three judged the best-performing GPT-4 witness to be human more often than human witnesses. This suggests that both humans and current LLMs struggle to distinguish between the two when they are not actively interrogating the person, underscoring an urgent need for more accurate tools to detect AI in conversations. |
varadarajan-etal-2025-consistent | 2,025 | The Consistent Lack of Variance of Psychological Factors Expressed by LLMs and Spambots | In recent years, the proliferation of chatbots like ChatGPT and Claude has led to an increasing volume of AI-generated text. While the text itself is convincingly coherent and human-like, the variety of expressed of human attributes may still be limited. Using theoretical individual differences, the fundamental psychological traits which distinguish people, this study reveals a distinctive characteristic of such content: AI-generations exhibit remarkably limited variation in inferrable psychological traits compared to human-authored texts. We present a review and study across multiple datasets spanning various domains. We find that AI-generated text consistently models the authorship of an {\textquotedblleft}average{\textquotedblright} human with such little variation that, on aggregate, it is clearly distinguishable from human-written texts using unsupervised methods (i.e., without using ground truth labels). Our results show that (1) fundamental human traits are able to accurately distinguish human- and machine-generated text and (2) current generation capabilities fail to capture a diverse range of human traits |
masrour-etal-2025-damage | 2,025 | DAMAGE: Detecting Adversarially Modified AI Generated Text | AI humanizers are a new class of online software tools meant to paraphrase and rewrite AI-generated text in a way that allows them to evade AI detection software. We study 19 AI humanizer and paraphrasing tools and qualitatively assess their effects and faithfulness in preserving the meaning of the original text. We show that many existing AI detectors fail to detect humanized text. Finally, we demonstrate a robust model that can detect humanized AI text while maintaining a low false positive rate using a data-centric augmentation approach. We attack our own detector, training our own fine-tuned model optimized against our detector`s predictions, and show that our detector`s cross-humanizer generalization is sufficient to remain robust to this attack. |
valdez-gomez-adorno-2025-text | 2,025 | Text Graph Neural Networks for Detecting AI-Generated Content | The widespread availability of Large Language Models (LLMs) such as GPT-4 and Llama-3, among others, has led to a surge in machine-generated content across various platforms, including social media, educational tools, and academic settings. While these models demonstrate remarkable capabilities in generating coherent text, their misuse raises significant concerns. For this reason, detecting machine-generated text has become a pressing need to mitigate these risks. This research proposed a novel classification method combining text-graph representations with Graph Neural Networks (GNNs) and different node feature initialization strategies to distinguish between human-written and machine-generated content. Experimental results demonstrate that the proposed approach outperforms traditional machine learning classifiers, highlighting the effectiveness of integrating structural and semantic relationships in text. |
keles-etal-2025-know | 2,025 | I Know You Did Not Write That! A Sampling Based Watermarking Method for Identifying Machine Generated Text | Potential harms of Large Language Models such as mass misinformation and plagiarism can be partially mitigated if there exists a reliable way to detect machine generated text. In this paper, we propose a new watermarking method to detect machine-generated texts. Our method embeds a unique pattern within the generated text, ensuring that while the content remains coherent and natural to human readers, it carries distinct markers that can be identified algorithmically. Specifically, we intervene with the token sampling process in a way which enables us to trace back our token choices during the detection phase. We show how watermarking affects textual quality and compare our proposed method with a state-of-the-art watermarking method in terms of robustness and detectability. Through extensive experiments, we demonstrate the effectiveness of our watermarking scheme in distinguishing between watermarked and non-watermarked text, achieving high detection rates while maintaining textual quality. |
zhang-etal-2025-dcbu | 2,025 | DCBU at GenAI Detection Task 1: Enhancing Machine-Generated Text Detection with Semantic and Probabilistic Features | This paper presents our approach to the MGT Detection Task 1, which focuses on detecting AI-generated content. The objective of this task is to classify texts as either machine-generated or human-written. We participated in Subtask A, which concentrates on English-only texts. We utilized the RoBERTa model for semantic feature extraction and the LLaMA3 model for probabilistic feature analysis. By integrating these features, we aimed to enhance the system`s classification accuracy. Our approach achieved strong results, with an F1 score of 0.7713 on Subtask A, ranking ninth among 36 teams. These results demonstrate the effectiveness of our feature integration strategy. |
tran-nam-2025-l3i | 2,025 | L3i++ at GenAI Detection Task 1: Can Label-Supervised LLaMA Detect Machine-Generated Text? | The widespread use of large language models (LLMs) influences different social media and educational contexts through the overwhelming generated text with a certain degree of coherence. To mitigate their potential misuse, this paper explores the feasibility of finetuning LLaMA with label supervision (named LS-LLaMA) in unidirectional and bidirectional settings, to discriminate the texts generated by machines and humans in monolingual and multilingual corpora. Our findings show that unidirectional LS-LLaMA outperformed the sequence language models as the benchmark by a large margin. Our code is publicly available at https://github.com/honghanhh/llama-as-a-judge. |
mehak-etal-2025-techexperts | 2,025 | TechExperts(IPN) at GenAI Detection Task 1: Detecting AI-Generated Text in English and Multilingual Contexts | The ever-increasing spread of AI-generated text, driven by the considerable progress in large language models, entails a real problem for all digital platforms: how to ensure con tent authenticity. The team TechExperts(IPN) presents a method for detecting AI-generated content in English and multilingual contexts, using the google/gemma-2b model fine-tuned for COLING 2025 shared task 1 for English and multilingual. Training results show peak F1 scores of 97.63{\%} for English and 97.87{\%} for multilingual detection, highlighting the model`s effectiveness in supporting content integrity across platforms. |
kiss-berend-2025-szegedai | 2,025 | SzegedAI at GenAI Detection Task 1: Beyond Binary - Soft-Voting Multi-Class Classification for Binary Machine-Generated Text Detection Across Diverse Language Models | This paper describes the participation of the SzegedAI team in Subtask A of Task 1 at the COLING 2025 Workshop on Detecting AI-Generated Content. Our solutions investigate the effectiveness of combining multi-class approaches with ensemble methods for detecting machine-generated text. This approach groups models into multiple classes based on properties such as model size or generative capabilities. Additionally, we employ a length-based method, utilizing specialized expert models designed for specific text length ranges. During inference, we condense multi-class predictions into a binary outcome, categorizing any label other than human as AI-generated. The effectiveness of both standard and snapshot ensemble techniques is evaluated. Although not all multi-class configurations outperformed the binary setup, our findings indicate that the combination of multi-class training and ensemble methods can enhance performance over single-method or binary approaches. |
creanga-etal-2025-team | 2,025 | Team Unibuc - NLP at GenAI Detection Task 1: Qwen it detect machine-generated text? | We explored both masked language models and causal models. For Subtask A, our best model achieved first-place out of 36 teams when looking at F1 Micro (Auxiliary Score) of 0.8333, and second-place when looking at F1 Macro (Main Score) of 0.8301. For causal models, our best model was a fine-tuned version of Qwen and for masked models, our best model was a fine-tuned version of XLM-Roberta-Base. |
schaefer-steinebach-2025-fraunhofer | 2,025 | Fraunhofer SIT at GenAI Detection Task 1: Adapter Fusion for AI-generated Text Detection | The detection of AI-generated content is becoming increasingly important with the growing prevalence of tools such as ChatGPT. This paper presents our results in the GenAI Content Detection Task 1, focusing on binary English and multilingual AI-generated text detection. We trained and tested transformers, adapters and adapter fusion. In the English setting (Subtask A), the combination of our own adapter on AI-generated text detection based on RoBERTa with a task adapter on multi-genre NLI yielded a macro F1 score of 0.828 on the challenge test set, ranking us third out of 35 teams. In the multilingual setting (Subtask B), adapter fusion resulted in a deterioration of the results. Consequently, XLM-RoBERTa, fine-tuned on the training set, was employed for the final evaluation, attaining a macro F1 score of 0.7258 and ranking tenth out of 25 teams. |
agrahari-ranbir-singh-2025-osint | 2,025 | OSINT at GenAI Detection Task 1: Multilingual MGT Detection: Leveraging Cross-Lingual Adaptation for Robust LLMs Text Identification | Detecting AI-generated text has become in- creasingly prominent. This paper presents our solution for the DAIGenC Task 1 Subtask 2, where we address the challenge of distin- guishing human-authored text from machine- generated content, especially in multilingual contexts. We introduce Multi-Task Detection (MLDet), a model that leverages Cross-Lingual Adaptation and Model Generalization strate- gies for Multilingual Machine-Generated Text (MGT) detection. By combining language- specific embeddings with fusion techniques, MLDet creates a unified, language-agnostic feature representation, enhancing its ability to generalize across diverse languages and mod- els. Our approach demonstrates strong perfor- mance, achieving macro and micro F1 scores of 0.7067 and 0.7187, respectively, and ranking 15th in the competition1. We also evaluate our model across datasets generated by different distinct models in many languages, showcasing its robustness in multilingual and cross-model scenarios. |
park-etal-2025-nota | 2,025 | Nota AI at GenAI Detection Task 1: Unseen Language-Aware Detection System for Multilingual Machine-Generated Text | Recently, large language models (LLMs) have demonstrated unprecedented capabilities in language generation, yet they still often produce incorrect information. Therefore, determining whether a text was generated by an LLM has become one of the factors that must be considered when evaluating its reliability. In this paper, we discuss methods to determine whether texts written in various languages were authored by humans or generated by LLMs. We have discovered that the classification accuracy significantly decreases for texts written in languages not observed during the training process, and we aim to address this issue. We propose a method to improve performance for unseen languages by using token-level predictive distributions extracted from various LLMs and text embeddings from a multilingual pre-trained langauge model. With the proposed method, we achieved third place out of 25 teams in Subtask B (binary multilingual machine-generated text detection) of Shared Task 1, with an F1 macro score of 0.7532. |
yadagiri-etal-2025-cnlp | 2,025 | CNLP-NITS-PP at GenAI Detection Task 1: AI-Generated Text Using Transformer-Based Approaches | In the current digital landscape, distinguishing between text generated by humans and that created by large language models has become increasingly complex. This challenge is exacerbated by advanced LLMs such as the Gemini, ChatGPT, GPT-4, and LLaMa, which can produce highly sophisticated, human-like text. This indistinguishability introduces a range of challenges across different sectors. Cybersecurity increases the risk of social engineering and misinformation, while social media aids the spread of biased or false content. The educational sector faces issues of academic integrity, and within large, multi-team environments, these models add complexity to managing interactions between human and AI agents. To address these challenges, we approached the problem as a binary classification task using an English-language benchmark COLING dataset. We employed transformer-based neural network models, including BERT, DistilBERT, and RoBERTa, fine-tuning each model with optimized hyperparameters to maximize classification accuracy. Our team CNLP-NITS-PP has achieved the 23rd rank in subtask 1 at COLING-2025 for machine-generated text detection in English with a Main Score F1 Macro of 0.6502 and micro-F1 score of 0.6876. |
mobin-islam-2025-luxveri | 2,025 | LuxVeri at GenAI Detection Task 1: Inverse Perplexity Weighted Ensemble for Robust Detection of AI-Generated Text across English and Multilingual Contexts | This paper presents a system developed for Task 1 of the COLING 2025 Workshop on Detecting AI-Generated Content, focusing on the binary classification of machine-generated versus human-written text. Our approach utilizes an ensemble of models, with weights assigned according to each model`s inverse perplexity, to enhance classification accuracy. For the English text detection task, we combined RoBERTa-base, RoBERTa-base with the OpenAI detector, and BERT-base-cased, achieving a Macro F1-score of 0.7458, which ranked us 12th out of 35 teams. We ensembled RemBERT, XLM-RoBERTa-base, and BERT-base-multilingual-case for the multilingual text detection task, employing the same inverse perplexity weighting technique. This resulted in a Macro F1-score of 0.7513, positioning us 4th out of 25 teams. Our results demonstrate the effectiveness of inverse perplexity weighting in improving the robustness of machine-generated text detection across both monolingual and multilingual settings, highlighting the potential of ensemble methods for this challenging task. |
doan-inui-2025-grape | 2,025 | Grape at GenAI Detection Task 1: Leveraging Compact Models and Linguistic Features for Robust Machine-Generated Text Detection | In this project, we aim to address two subtasks of Task 1: Binary Multilingual Machine-Generated Text (MGT) Detection (Human vs. Machine) as part of the COLING 2025 Workshop on MGT Detection (Wang et al., 2025) using different approaches. The first method involves separately fine-tuning small language models tailored to the specific subtask. The second approach builds on this methodology by incorporating linguistic, syntactic, and semantic features, leveraging ensemble learning to integrate these features with model predictions for more robust classification. By evaluating and comparing these approaches, we aim to identify the most effective techniques for detecting machine-generated content across languages, providing insights into improving automated verification tools amidst the rapid growth of LLM-generated text in digital spaces. |
bhandarkar-etal-2025-aaig | 2,025 | AAIG at GenAI Detection Task 1: Exploring Syntactically-Aware, Resource-Efficient Small Autoregressive Decoders for AI Content Detection | This paper presents a lightweight and efficient approach to AI-generated content detection using small autoregressive fine-tuned decoders (AFDs) for secure, on-device deployment. Motivated by resource-efficiency, syntactic awareness, and bias mitigation, our model employs small language models (SLMs) with autoregressive pre-training and loss fusion to accurately distinguish between human and AI-generated content while significantly reducing computational demands. The system achieved highest macro-F1 score of 0.8186, with the submitted model scoring 0.7874{---}both significantly outperforming the task baseline while reducing model parameters by {\textasciitilde}60{\%}. Notably, our approach mitigates biases, improving recall for human-authored text by over 60{\%}. Ranking 8th out of 36 participants, these results confirm the feasibility and competitiveness of small AFDs in challenging, adversarial settings, making them ideal for privacy-preserving, on-device deployment suitable for real-world applications. |
keles-kutlu-2025-turquaz | 2,025 | TurQUaz at GenAI Detection Task 1:Dr. Perplexity or: How I Learned to Stop Worrying and Love the Finetuning | This paper details our methods for addressing Task 1 of the GenAI Content Detection shared tasks, which focus on distinguishing AI-generated text from human-written content. The task comprises two subtasks: Subtask A, centered on English-only datasets, and Subtask B, which extends the challenge to multilingual data. Our approach uses a fine-tuned XLM-RoBERTa model for classification, complemented by features including perplexity and TF-IDF. While perplexity is commonly regarded as a useful indicator for identifying machine-generated text, our findings suggest its limitations in multi-model and multilingual contexts. Our approach ranked 6th in Subtask A, but a submission issue left our Subtask B unranked, where it would have placed 23rd. |
singh-etal-2025-ai | 2,025 | AI-Monitors at GenAI Detection Task 1: Fast and Scalable Machine Generated Text Detection | We describe the work carried out by our team, AI-Monitors, on the Binary Multilingual Machine-Generated Text Detection (Human vs. Machine) task at COLING 2025. This task aims to determine whether a given text is generated by a machine or authored by a human. We propose a lightweight, simple, and scalable approach using encoder models such as RoBERTa and XLM-R We provide an in-depth analysis based on our experiments. Our study found that carefully exploring fine-tuned parameters such as i) no. of training epochs, ii) maximum input size, iii) handling class imbalance etc., plays an important role in building an effective system to achieve good results and can significantly impact the underlying tasks. We found the optimum setting of these parameters can lead to a difference of about 5-6{\%} in absolute terms for measure such as accuracy and F1 measure. The paper presents crucial insights into optimal parameter selection for fine-tuning RoBERTa and XLM-R based models to detect whether a given text is generated by a machine or a human. |
gritsai-etal-2025-advacheck | 2,025 | Advacheck at GenAI Detection Task 1: AI Detection Powered by Domain-Aware Multi-Tasking | The paper describes a system designed by Advacheck team to recognise machine-generated and human-written texts in the monolingual subtask of GenAI Detection Task 1 competition. Our developed system is a multi-task architecture with shared Transformer Encoder between several classification heads. One head is responsible for binary classification between human-written and machine-generated texts, while the other heads are auxiliary multiclass classifiers for texts of different domains from particular datasets. As multiclass heads were trained to distinguish the domains presented in the data, they provide a better understanding of the samples. This approach led us to achieve the first place in the official ranking with 83.07{\%} macro F1-score on the test set and bypass the baseline by 10{\%}. We further study obtained system through ablation, error and representation analyses, finding that multi-task learning outperforms single-task mode and simultaneous tasks form a cluster structure in embeddings space. |
wang-etal-2025-genai | 2,025 | GenAI Content Detection Task 1: English and Multilingual Machine-Generated Text Detection: AI vs. Human | We present the GenAI Content Detection Task 1 {--} a shared task on binary machine generated text detection, conducted as a part of the GenAI workshop at COLING 2025. The task consists of two subtasks: Monolingual (English) and Multilingual. The shared task attracted many participants: 36 teams made official submissions to the Monolingual subtask during the test phase and 27 teams {--} to the Multilingual. We provide a comprehensive overview of the data, a summary of the results {--} including system rankings and performance scores {--} detailed descriptions of the participating systems, and an in-depth analysis of submissions. |
abiola-etal-2025-cic | 2,025 | CIC-NLP at GenAI Detection Task 1: Advancing Multilingual Machine-Generated Text Detection | Machine-written texts are gradually becoming indistinguishable from human-generated texts, leading to the need to use sophisticated methods to detect them. Team CIC-NLP presents work in the Gen-AI Content Detection Task 1 at COLING 2025 Workshop: the focus of our work is on Subtask B of Task 1, which is the classification of text written by machines and human authors, with particular attention paid to identifying multilingual binary classification problem. Usng mBERT, we addressed the binary classification task using the dataset provided by the GenAI Detection Task team. mBERT acchieved a macro-average F1-score of 0.72 as well as an accuracy score of 0.73. |
Subsets and Splits