ID
large_stringlengths 10
61
| year
int64 1.96k
2.03k
| title
large_stringlengths 4
560
| abstract
large_stringlengths 0
12.8k
|
|---|---|---|---|
abiola-etal-2025-cic-nlp | 2,025 | CIC-NLP at GenAI Detection Task 1: Leveraging DistilBERT for Detecting Machine-Generated Text in English | As machine-generated texts (MGT) become increasingly similar to human writing, these dis- tinctions are harder to identify. In this paper, we as the CIC-NLP team present our submission to the Gen-AI Content Detection Workshop at COLING 2025 for Task 1 Subtask A, which involves distinguishing between text generated by LLMs and text authored by humans, with an emphasis on detecting English-only MGT. We applied the DistilBERT model to this binary classification task using the dataset provided by the organizers. Fine-tuning the model effectively differentiated between the classes, resulting in a micro-average F1-score of 0.70 on the evaluation test set. We provide a detailed explanation of the fine-tuning parameters and steps involved in our analysis. |
lekkala-etal-2025-nits | 2,025 | nits\_teja\_srikar at GenAI Detection Task 2: Distinguishing Human and AI-Generated Essays Using Machine Learning and Transformer Models | This paper presents models to differentiate between human-written and AI-generated essays, addressing challenges posed by advanced AI models like ChatGPT and Claude. Using a structured dataset, we fine-tune multiple machine learning models, including XGBoost and Logistic Regression, along with ensemble learning and k-fold cross-validation. The dataset is processed through TF-IDF vectorization, followed by text cleaning, lemmatization, stemming, and part-of-speech tagging before training. Our team nits{\_}teja{\_}srikar achieves high accuracy, with DistilBERT performing at 77.3{\%} accuracy, standing at 20th position for English, and XLM-RoBERTa excelling in Arabic at 92.2{\%}, standing at 14th position in the official leaderboard, demonstrating the model`s potential for real-world applications. |
alsmadi-2025-integrityai | 2,025 | IntegrityAI at GenAI Detection Task 2: Detecting Machine-Generated Academic Essays in English and Arabic Using ELECTRA and Stylometry | We present a robust system for detecting machine-generated academic essays, leveraging pre-trained, transformer-based models specifically tailored for both English and Arabic texts. Our primary approach utilizes ELECTRA-Small for English and AraELECTRA-Base for Arabic, fine-tuned to deliver high performance while balancing computational efficiency. By incorporating stylometric features, such as word count, sentence length, and vocabulary richness, our models excel at distinguishing between human-written and AI-generated content. Proposed models achieved excellent results with an F1- score of 99.7{\%}, ranking second among of 26 teams in the English subtask, and 98.4{\%}, finishing first out of 23 teams in the Arabic one. Main Contributions include: (1) We develop lightweight and efficient models using ELECTRA-Small and AraELECTRA-Base, achieving an impressive F1-score of 98.5{\%} on the English dataset and 98.4{\%} on the Arabic dataset. This demonstrates the power of combining transformer-based architectures with stylometric analysis. (2) We optimize our system to maintain high performance while being computationally efficient, making it suitable for deployment on GPUs with moderate memory capacity. (3) Additionally, we tested larger models, such as ELECTRA-Large, achieving an even higher F1-score of 99.7{\%} on the English dataset, highlighting the potential for further accuracy gains when using more computationally intensive models. |
jiao-etal-2025-cmi | 2,025 | CMI-AIGCX at GenAI Detection Task 2: Leveraging Multilingual Proxy LLMs for Machine-Generated Text Detection in Academic Essays | This paper presents the approach we proposed for GenAI Detection Task 2, which aims to classify a given text as either machine-generated or human-written, with a particular emphasis on academic essays. We participated in subtasks A and B, which focus on detecting English and Arabic essays, respectively. We propose a simple and efficient method for detecting machine-generated essays, where we use the Llama-3.1-8B as a proxy to capture the essence of each token in the text. These essences are processed and classified using a refined feature classification network. Our approach does not require fine-tuning the LLM. Instead, we leverage its extensive multilingual knowledge acquired during pretraining to significantly enhance detection performance. The results validate the effectiveness of our approach and demonstrate that leveraging a proxy model with diverse multilingual knowledge can significantly enhance the detection of machine-generated text across multiple languages, regardless of model size. In Subtask A, we achieved an F1 score of 99.9{\%}, ranking first out of 26 teams. In Subtask B, we achieved an F1 score of 96.5{\%}, placing fourth out of 22 teams, with the same score as the third-place team. |
agrahari-etal-2025-essaydetect | 2,025 | EssayDetect at GenAI Detection Task 2: Guardians of Academic Integrity: Multilingual Detection of AI-Generated Essays | Detecting AI-generated text in the field of academia is becoming very prominent. This paper presents a solution for Task 2: AI vs. Hu- man {--} Academic Essay Authenticity Challenge in the COLING 2025 DAIGenC Workshop 1. The rise of Large Language models (LLMs) like ChatGPT has posed significant challenges to academic integrity, particularly in detecting AI-generated essays. To address this, we pro- pose a fusion model that combines pre-trained language model embeddings with stylometric and linguistic features. Our approach, tested on both English and Arabic, utilizes adaptive training and attention mechanisms to enhance F1 scores, address class imbalance, and capture linguistic nuances across languages. This work advances multilingual solutions for detecting AI-generated text in academia. |
yadagiri-etal-2025-cnlp-nits | 2,025 | CNLP-NITS-PP at GenAI Detection Task 2: Leveraging DistilBERT and XLM-RoBERTa for Multilingual AI-Generated Text Detection | In today`s digital landscape, distinguishing between human-authored essays and content generated by advanced Large Language Models such as ChatGPT, GPT-4, Gemini, and LLaMa has become increasingly complex. This differentiation is essential across sectors like academia, cybersecurity, social media, and education, where the authenticity of written material is often crucial. Addressing this challenge, the COLING 2025 competition introduced Task 2, a binary classification task to separate AI-generated text from human-authored content. Using a benchmark dataset for English and Arabic, developing a methodology that fine-tuned various transformer-based neural networks, including CNN-LSTM, RNN, Bi-GRU, BERT, DistilBERT, GPT-2, and RoBERTa. Our Team CNLP-NITS-PP achieved competitive performance through meticulous hyperparameter optimization, reaching a Recall score of 0.825. Specifically, we ranked 18th in the English sub-task A with an accuracy of 0.77 and 20th in the Arabic sub-task B with an accuracy of 0.59. These results underscore the potential of transformer-based models in academic settings to detect AI-generated content effectively, laying a foundation for more advanced methods in essay authenticity verification. |
gharib-elgendy-2025-ra | 2,025 | RA at GenAI Detection Task 2: Fine-tuned Language Models For Detection of Academic Authenticity, Results and Thoughts | This paper assesses the performance of {\textquotedblleft}RA{\textquotedblright} in the Academic Essay Authenticity Challenge, which saw nearly 30 teams participating in each subtask. We employed cutting-edge transformer-based models to achieve our results. Our models consistently exceeded both the mean and median scores across the tasks. Notably, we achieved an F1-score of 0.969 in classifying AI-generated essays in English and an F1-score of 0.957 for classifying AI-generated essays in Arabic. Additionally, this paper offers insights into the current state of AI-generated models and argues that the benchmarking methods currently in use do not accurately reflect real-world scenarios. |
indurthi-varma-2025-tesla | 2,025 | Tesla at GenAI Detection Task 2: Fast and Scalable Method for Detection of Academic Essay Authenticity | This paper describes a simple yet effective method to identify if academic essays have been written by students or generated through the language models in English language. We extract a set of style, language complexity, bias and subjectivity, and emotion-based features that can be used to distinguish human-written essays from machine-generated essays. Our methods rank 6th on the leaderboard, achieving an impressive F1-score of 0.986. |
chowdhury-etal-2025-genai | 2,025 | GenAI Content Detection Task 2: AI vs. Human -- Academic Essay Authenticity Challenge | This paper presents a comprehensive overview of the first edition of the Academic Essay Authenticity Challenge, organized as part of the GenAI Content Detection shared tasks collocated with COLING 2025. This challenge focuses on detecting machine-generated \textit{vs} human-authored essays for academic purposes. The task is defined as follows: \textit{{\textquotedblleft}Given an essay, identify whether it is generated by a machine or authored by a human.{\textquotedblright}} The challenge involves two languages: English and Arabic. During the evaluation phase, 25 teams submitted systems for English and 21 teams for Arabic, reflecting substantial interest in the task. Finally, five teams submitted system description papers. The majority of submissions utilized fine-tuned transformer-based models, with one team employing Large Language Models (LLMs) such as Llama 2 and Llama 3. This paper outlines the task formulation, details the dataset construction process, and explains the evaluation framework. Additionally, we present a summary of the approaches adopted by participating teams. Nearly all submitted systems outperformed the n-gram-based baseline, with the top-performing systems achieving F1 scores exceeding 0.98 for both languages, indicating significant progress in the detection of machine-generated text. |
lekkala-etal-2025-cnlp | 2,025 | CNLP-NITS-PP at GenAI Detection Task 3: Cross-Domain Machine-Generated Text Detection Using DistilBERT Techniques | This paper presents a Cross-domain Machine-Generated Text Detection model developed for the COLING 2025 Workshop on Detecting AI-generated Content (DAIGenC). As large language models evolve, detecting machine-generated text becomes increasingly challenging, particularly in contexts like misinformation and academic integrity. While current detectors perform well on unseen data, they remain vulnerable to adversarial strategies, including paraphrasing, homoglyphs, misspellings, synonyms, whitespace manipulations, etc. We introduce a framework to address these adversarial tactics designed to bypass detection systems by adversarial training. Our team DistilBERT-NITS detector placed 7th in the Non-Adversarial Attacks category, and Adversarial-submission-3 achieved 17th in the Adversarial Attacks category. |
edikala-etal-2025-leidos | 2,025 | Leidos at GenAI Detection Task 3: A Weight-Balanced Transformer Approach for AI Generated Text Detection Across Domains | Advancements in Large Language Models (LLMs) blur the distinction between human and machine-generated text (MGT), raising concerns about misinformation and academic dishonesty. Existing MGT detection methods often fail to generalize across domains and generator models. We address this by framing MGT detection as a text classification task using transformer-based models. Utilizing Distil-RoBERTa-Base, we train four classifiers (binary and multi-class, with and without class weighting) on the RAID dataset (Dugan et al., 2024). Our systems placed first to fourth in the COLING 2025 MGT Detection Challenge Task 3 (Dugan et al., 2025). Internal in-domain and zero-shot evaluations reveal that applying class weighting improves detector performance, especially with multi-class classification training. Our best model effectively generalizes to unseen domains and generators, demonstrating that transformer-based models are robust detectors of machine-generated text. |
emi-etal-2025-pangram | 2,025 | Pangram at GenAI Detection Task 3: An Active Learning Approach to Machine-Generated Text Detection | We pretrain an autoregressive LLM-based detector on a wide variety of datasets, domains, languages, prompt schemes, and LLMs used to generate the AI portion of the dataset. We aggressively employ several augmentation strategies and preprocessing strategies to improve robustness. We then mine the RAID train set for the AI examples with the largest error based on the original classifier, and mix those examples and their human-written counterparts back into the training set. We then retrain the detector until convergence. |
mobin-islam-2025-luxveri-genai | 2,025 | LuxVeri at GenAI Detection Task 3: Cross-Domain Detection of AI-Generated Text Using Inverse Perplexity-Weighted Ensemble of Fine-Tuned Transformer Models | This paper presents our approach for Task 3 of the GenAI content detection workshop at COLING-2025, focusing on Cross-Domain Machine-Generated Text (MGT) Detection. We propose an ensemble of fine-tuned transformer models, enhanced by inverse perplexity weighting, to improve classification accuracy across diverse text domains. For Subtask A (Non-Adversarial MGT Detection), we combined a fine-tuned RoBERTa-base model with an OpenAI detector-integrated RoBERTa-base model, achieving an aggregate TPR score of 0.826, ranking 10th out of 23 detectors. In Subtask B (Adversarial MGT Detection), our fine-tuned RoBERTa-base model achieved a TPR score of 0.801, securing 8th out of 22 detectors. Our results demonstrate the effectiveness of inverse perplexity-based weighting for enhancing generalization and performance in both non-adversarial and adversarial MGT detection, highlighting the potential for transformer models in cross-domain AI-generated content detection. |
kandula-etal-2025-bbn | 2,025 | BBN-U.Oregon`s ALERT system at GenAI Content Detection Task 3: Robust Authorship Style Representations for Cross-Domain Machine-Generated Text Detection | This paper presents BBN-U.Oregon`s system, ALERT, submitted to the Shared Task 3: Cross-Domain Machine-Generated Text Detection. Our approach uses robust authorship-style representations to distinguish between human-authored and machine-generated text (MGT) across various domains. We employ an ensemble-based authorship attribution (AA) system that integrates stylistic embeddings from two complementary subsystems: one that focuses on cross-genre robustness with hard positive and negative mining strategies and another that captures nuanced semantic-lexical-authorship contrasts. This combination enhances cross-domain generalization, even under domain shifts and adversarial attacks. Evaluated on the RAID benchmark, our system demonstrates strong performance across genres and decoding strategies, with resilience against adversarial manipulation, achieving 91.8{\%} TPR at FPR=5{\%} on standard test sets and 82.6{\%} on adversarial sets. |
agrahari-etal-2025-random | 2,025 | Random at GenAI Detection Task 3: A Hybrid Approach to Cross-Domain Detection of Machine-Generated Text with Adversarial Attack Mitigation | Machine-generated text (MGT) detection has gained critical importance in the era of large language models, especially for maintaining trust in multilingual and cross-domain applica- tions. This paper presents Task 3 Subtask B: Adversarial Cross-Domain MGT Detection for in the COLING 2025 DAIGenC Workshop. Task 3 emphasizes the complexity of detecting AI-generated text across eight domains, eleven generative models, and four decoding strate- gies, with an added challenge of adversarial manipulation. We propose a robust detection framework transformer embeddings utilizing Domain-Adversarial Neural Networks (DANN) to address domain variability and adversarial robustness. Our model demonstrates strong performance in identifying AI-generated text under adversarial conditions while highlighting condition scope of future improvement. |
dubois-etal-2025-mosaic | 2,025 | MOSAIC at GENAI Detection Task 3 : Zero-Shot Detection Using an Ensemble of Models | MOSAIC introduces a new ensemble approach that combines several detector models to spot AI-generated texts. The method enhances the reliability of detection by integrating insights from multiple models, thus addressing the limitations of using a single detector model which often results in performance brittleness. This approach also involves using a theoretically grounded algorithm to minimize the worst-case expected encoding size across models, thereby optimizing the detection process. In this submission, we report evaluation results on the RAID benchmark, a comprehensive English-centric testbed for machine-generated texts. These results were obtained in the context of the {\textquotedblleft}Cross-domain Machine-Generated Text Detection{\textquotedblright} shared task. We show that our model can be competitive for a variety of domains and generator models, but that it can be challenged by adversarial attacks and by changes in the text generation strategy. |
dugan-etal-2025-genai | 2,025 | GenAI Content Detection Task 3: Cross-Domain Machine Generated Text Detection Challenge | Recently there have been many shared tasks targeting the detection of generated text from Large Language Models (LLMs). However, these shared tasks tend to focus either on cases where text is limited to one particular domain or cases where text can be from many domains, some of which may not be seen during test time. In this shared task, using the newly released RAID benchmark, we aim to answer whether or not models can detect generated text from a large, yet fixed, number of domains and LLMs, all of which are seen during training. Over the course of three months, our task was attempted by 9 teams with 23 detector submissions. We find that multiple participants were able to obtain accuracies of over 99{\%} on machine-generated text from RAID while maintaining a 5{\%} False Positive Rate{---}suggesting that detectors are able to robustly detect text from many domains and models simultaneously. We discuss potential interpretations of this result and provide directions for future research. |
biancotti-etal-2025-chat | 2,025 | Chat Bankman-Fried: an Exploration of LLM Alignment in Finance | Advancements in large language models (LLMs) have renewed concerns about AI alignment{---}the consistency between human and AI goals and values. As various jurisdictions enact legislation on AI safety, the concept of alignment must be defined and measured across different domains. This paper proposes an experimental framework to assess whether LLMs adhere to ethical and legal standards in the relatively unexplored context of finance. We prompt ten LLMs to impersonate the CEO of a financial institution and test their willingness to misuse customer assets to repay outstanding corporate debt. Beginning with a baseline configuration, we adjust preferences, incentives and constraints, analyzing the impact of each adjustment with logistic regression. Our findings reveal significant heterogeneity in the baseline propensity for unethical behavior of LLMs. Factors such as risk aversion, profit expectations, and regulatory environment consistently influence misalignment in ways predicted by economic theory, although the magnitude of these effects varies across LLMs. This paper highlights the benefits and limitations of simulation-based, ex-post safety testing. While it can inform financial authorities and institutions aiming to ensure LLM safety, there is a clear trade-off between generality and cost. |
shukla-etal-2025-graphrag | 2,025 | GraphRAG Analysis for Financial Narrative Summarization and A Framework for Optimizing Domain Adaptation | Large Language Models (LLMs) have shown promise in summarizing complex documents, but their limitations in handling lengthy documents and capturing global information hinder their performance in tasks like Query-Focused Summarization (QFS). We explore GraphRAG, a retrieval-augmented generation approach that utilizes a globally summarized knowledge graph derived from an LLM. We apply GraphRAG to the Financial Narrative Summarization (FNS) dataset, which consists of lengthy financial reports. Our results show that a naive RAG approach outperforms GraphRAG in terms of comprehensiveness, directness, conciseness and completeness. However, we demonstrate that optimizing entity and relation extraction using an LLM as an optimizer can enhance GraphRAG`s performance. Our study highlights the need for domain-specific optimization to improve GraphRAG`s capabilities for summarization tasks in facts-heavy domains like finance. We propose an optimization framework that extends GraphRAG`s original domain adaptation strategy by incorporating entity and relations optimization, leading to improved performance in capturing relevant entities and relationships. Our findings contribute to the development of more effective summarization models for complex documents in finance and other domains. |
wang-etal-2025-buddie | 2,025 | BuDDIE: A Business Document Dataset for Multi-task Information Extraction | The field of visually rich document understanding (VRDU) aims to solve a multitude of well-researched NLP tasks in the multi-modal domain. Several datasets exist for research on specific tasks of VRDU, such as document classification (DC), key entity extraction (KEE), entity linking, visual question answering (VQA), inter alia. These datasets cover documents like invoices and receipts with sparse annotations such that they support one or two co-related tasks (e.g., entity extraction and entity linking). Unfortunately, only focusing on a single specific type of documents or task is not representative of how documents often need to be processed in the wild {--} where variety in style and requirements is expected. In this paper, we introduce BuDDIE: Business Document Dataset for Information Extraction, the first multi-task dataset of 1665 real-world business documents that contains rich and dense annotations for DC, KEE, and VQA. Our dataset consists of publicly available business entity documents from US state government websites. The documents are structured and vary in their style and layout across states and types (e.g., forms, certificates, reports, etc.). We provide data variety and quality metrics for BuDDIE as well as a series of baselines for each task. Our baselines cover traditional textual, multi-modal, and large language model approaches to VRDU. |
zhang-yang-2025-finmoe | 2,025 | FinMoE: A MoE-based Large Chinese Financial Language Model | Large-scale language models have demonstrated remarkable success, achieving strong performance across a variety of general tasks. However, when applied to domain-specific fields, such as finance, these models face challenges due to the need for both specialized knowledge and robust general capabilities. In this paper, we introduce FinMoE, a MOE-based large-scale Chinese financial language model that bridges the gap between general language models and domain-specific requirements. FinMoE employs a dense MoE architecture, where all expert networks are simultaneously activated and dynamically combined to effectively integrate general linguistic understanding with domain-specific financial expertise. Experimental results demonstrate that FinMoE achieves state-of-the-art performance on both general-purpose and financial benchmarks at a comparable scale, validating its ability to balance domain specialization with general knowledge and reasoning. |
kumar-etal-2025-bridging-gap | 2,025 | Bridging the Gap: Efficient Cross-Lingual NER in Low-Resource Financial Domain | We present an innovative and efficient modeling framework for cross-lingual named entity recognition (NER), leveraging the strengths of knowledge distillation and consistency training. Our approach distills knowledge from an XLM-RoBERTa model pre-trained on a high-resource source language (English) to a student model, which then undergoes semi-supervised consistency training with KL divergence loss on a low-resource target language (Arabic). We focus our application on the financial domain, using a small, sourced dataset of financial transactions as seen in SMS messages Using datasets comprising SMS messages in English and Arabic containing financial transaction information, we aim to transfer NER capabilities from English to Arabic with minimal labeled Arabic samples. The framework generalizes named entity recognition from English to Arabic, achieving F1 scores of 0.74 on the Arabic financial transaction dataset and 0.61 on the WikiANN dataset, surpassing or closely competing with models that have 1.7 and 5.3 more parameters, respectively, while efficiently training it on a single T4 GPU. Our experiments show that using a small number of labeled data for low-resource cross-lingual NER applications is a wiser choice than utilizing zero-shot techniques while also using up fewer resources. This framework holds significant potential for developing multilingual applications, particularly in regions where digital interactions span English and low-resource languages. |
figueroa-rosero-etal-2025-evaluating | 2,025 | Evaluating Financial Literacy of Large Language Models through Domain Specific Languages for Plain Text Accounting | Large language models (LLMs) have proven highly effective for a wide range of tasks, including code generation. Recently, advancements in their capabilities have shown promise in areas like mathematical reasoning, chain-of-thought processes and self-reflection. However, their effectiveness in domains requiring nuanced understanding of financial contexts, such as accounting, remains unclear. In this study, we evaluate how well LLMs perform in generating code for domain-specific languages (DSLs) in accounting, using Beancount as a case study. We create a set of tasks based on common financial ratios, to evaluate the numeracy and financial literacy of LLMs. Our findings reveal that while LLMs are state-of-the art in generative tasks, they struggle severely with accounting, often producing inaccurate calculations and misinterpreting financial scenarios. We characterize these shortcomings through a comprehensive evaluation, shedding light on the limitations of LLMs in understanding and handling money-related tasks. |
harsha-etal-2025-synthetic | 2,025 | Synthetic Data Generation Using Large Language Models for Financial Question Answering | Recent research has shown excellent performance of large language models (LLMs) for answering questions requiring multi-step financial reasoning. While the larger models have been used with zero-shot or few-shot prompting, the smaller variants need fine-tuning on training data containing questions and the corresponding answers that includes detailed reasoning demonstrations. To alleviate the significant cost of creating a data set with complex questions and corresponding answers, we explore the use of synthetic data for financial question answering using a multi-step LLM based approach to generate question as well as the answers with reasoning steps. We consider standard as well as conversational financial question answering scenarios. We experiment with synthetic data generation for three different real financial reasoning problems that already have manually collected data sets created with the help of financial experts. Using the same document sources, we use the proposed LLM based approach to generate synthetic questions and answers. To measure the effectiveness, we train multiple small language models (SLMs) on these synthetic data and compare the performance with that of the same SLMs trained on the real data. We further perform extensive experimental analysis generating important evidence on the potential of using synthetic data in financial reasoning tasks. |
lin-jang-2025-concept | 2,025 | Concept-Based RAG Models: A High-Accuracy Fact Retrieval Approach | This study introduces a concept-based methodology to optimize Retrieval-Augmented Generation (RAG) tasks by assessing dataset certainty using entropy-based metrics and concept extraction techniques. Unlike traditional methods focused on reducing LLM hallucinations or modifying data structures, this approach evaluates inherent knowledge uncertainty from an LLM perspective. By pre-processing documents with LLMs, the concept-based method significantly enhances precision in tasks demanding high accuracy, such as legal, finance, or formal document responses. |
uthayasooriyar-etal-2025-training | 2,025 | Training LayoutLM from Scratch for Efficient Named-Entity Recognition in the Insurance Domain | Generic pre-trained neural networks may struggle to produce good results in specialized domains like finance and insurance. This is due to a domain mismatch between training data and downstream tasks, as in-domain data are often scarce due to privacy constraints. In this work, we compare different pre-training strategies for LayoutLM. We show that using domain-relevant documents improves results on a named-entity recognition (NER) problem using a novel dataset of anonymized insurance-related financial documents called PAYSLIPS. Moreover, we show that we can achieve competitive results using a smaller and faster model. |
klimaszewski-etal-2025-avenibench | 2,025 | AveniBench: Accessible and Versatile Evaluation of Finance Intelligence | Over the last few years, there has been great interest in applying large language models (LLMs) to problems in the finance industry, and the field needs a robust LLM benchmark to support this work. Current financial LLM benchmarks contain simple tasks which are not representative of real use cases and have test sets with licences that do not allow commercial use. In response, we release AveniBench, a permissively licensed benchmark that tests a group of six key finance-related skills: tabular reasoning, numerical reasoning, question answering, long context modelling, summarisation and dialogue. We refactor the test sets to ensure that metrics are comparable, providing a unified framework. Furthermore, AveniBench introduces two task difficulty modes, easy and hard, enabling scalable evaluation based on real-world deployment needs. We use our benchmark to evaluate a diverse set of 20 widely used LLMs, from small open-weight models to proprietary systems like GPT-4. This evaluation initiates our public leaderboard, providing valuable insights for future academic research and commercial development. |
drinkall-etal-2025-forecasting | 2,025 | Forecasting Credit Ratings: A Case Study where Traditional Methods Outperform Generative LLMs | Large Language Models (LLMs) have been shown to perform well for many downstream tasks. Transfer learning can enable LLMs to acquire skills that were not targeted during pre-training. In financial contexts, LLMs can sometimes beat well-established benchmarks. This paper investigates how well LLMs perform at forecasting corporate credit ratings. We show that while LLMs are very good at encoding textual information, traditional methods are still very competitive when it comes to encoding numeric and multimodal data. For our task, current LLMs perform worse than a more traditional XGBoost architecture that combines fundamental and macroeconomic data with high-density text-based embedding features. We investigate the degree to which the text encoding methodology affects performance and interpretability. |
yadav-etal-2025-investigating | 2,025 | Investigating the effectiveness of length based rewards in DPO for building Conversational Financial Question Answering Systems | In this paper, we address the numerical reasoning challenges of financial question-answering systems. We propose a two-stage approach where models first generate intermediate calculations and then produce the final answer. We perform two set of experiments to evaluate the performance of our approach. In the first, we compare single-step and multi-step approaches, demonstrating that incorporating intermediate calculations significantly improves numerical accuracy. In the second experiment, we compare traditional DPO and iterative DPO (iDPO) with length-regularized DPO. We show that while traditional DPO reduced parsing errors, it introduces verbosity; iDPO improves reasoning iteratively but faces diminishing returns. On the other hand, Length-regularized DPO reduces verbosity of intermediate calculation as well as enhances numerical accuracy across all models. These results highlight the potential of combining intermediate reasoning steps with domain-specific optimizations to build robust financial question-answering systems. |
yan-zhu-2025-creditllm | 2,025 | CreditLLM: Constructing Financial AI Assistant for Credit Products using Financial LLM and Few Data | Facilitating financial technology with the large-language model (LLM) has been developing in recent years. To address the challenges in one of the biggest world-wide markets, China, Chinese-expertise financial LLM has also been studied. The related works focus on conventional NLP tasks in finance, while developing LLM for specific tasks is also required. Besides, in the credit loan business, the existing AI-based approaches are largely related to Credit like Credit rating and Fraud prediction, while credit product customization is still missing. In China, Inclusive Finance and Rural Finance become two hot topics that raise critical challenges in flexibly customizing credit products to meet the variable fund requirements of small {\&} micro businesses, individual businesses, and agricultural businesses of local character. In this paper, the credit product customization is studied by developing an LLM-based financial AI assistant for the credit loan business. It is proposed to satisfy the business requirements of customer counseling, recommendation, and question-answers regarding credit loans. The proposed LLM is developed by Chinese prompt data automatically constructed based on a small set of real-world credit products. The experiments demonstrate its effectiveness in credit loan-related ability while maintaining comparable performance in conventional finance NLP tasks. |
xu-etal-2025-modeling | 2,025 | Modeling Interactions Between Stocks Using LLM-Enhanced Graphs for Volume Prediction | Accurate trading volume prediction is essential for portfolio optimization, market regulation, and financial risk control. An effective method for predicting trading volume involves building a graph to model relations between stock. Recent research has enhanced these models by integrating stock news to improve forecasting ability. However, existing approaches primarily integrate news data as auxiliary features for nodes in Graph Neural Networks (GNNs), overlooking the relational information between stocks embedded in news. To address this, we propose LLM-Enhanced Dynamic Graph Neural Network (LED-GNN), a framework that constructs dynamic graphs using inter-stock relationships extracted from news via a large language model (LLM)-centered pipeline, combined with graphs learned from historical price-volume data. A dynamic GNN then processes these graphs to generate predictions. Evaluated on a real-world dataset, TOPIX, with Reuters Financial News, LED-GNN consistently outperformed all baseline models, achieving a 2{\%} improvement over the strongest baseline. |
lu-huo-2025-financial | 2,025 | Financial Named Entity Recognition: How Far Can LLM Go? | The surge of large language models (LLMs) has revolutionized the extraction and analysis of crucial information from a growing volume of financial statements, announcements, and business news. Recognition for named entities to construct structured data poses a significant challenge in analyzing financial documents and is a foundational task for intelligent financial analytics. However, how effective are these generic LLMs and their performance under various prompts are yet need a better understanding. To fill in the blank, we present a systematic evaluation of state-of-the-art LLMs and prompting methods in the financial Named Entity Recognition (NER) problem. Specifically, our experimental results highlight their strengths and limitations, identify five representative failure types, and provide insights into their potential and challenges for domain-specific tasks. |
wang-etal-2025-proxy | 2,025 | Proxy Tuning for Financial Sentiment Analysis: Overcoming Data Scarcity and Computational Barriers | Financial sentiment analysis plays a pivotal role in the financial domain. However, the task remains challenging due to the nuanced nature of financial sentiment, the need for high interpretability, and the scarcity of high-quality datasets. To address these issues, we leverage recent advancements in large language models (LLMs) and propose to adapt proxy tuning for financial sentiment analysis. Proxy tuning efficiently transfers knowledge from a pre-trained expert model to a controllable base model by incorporating logit differences, steering the base model toward the desired sentiment representation. Our method offers significant advantages: (1) it is training-free, reducing computational demands and data dependency; (2) it achieves promising performance, with a 36.67{\%} improvement over the base model and over 90{\%} of the tuned model`s performance; and (3) it is highly adaptable, functioning in a plug-and-play manner without requiring access to model architectures or weights. These results demonstrate the potential of proxy tuning as an efficient and practical solution for financial sentiment analysis in data-scarce scenarios. |
ettaleb-etal-2025-contribution | 2,025 | The contribution of LLMs to relation extraction in the economic field | Relation Extraction (RE) is a fundamental task in natural language processing, aimed at deducing semantic relationships between entities in a text. Traditional supervised extraction methods relation extraction methods involve training models to annotate tokens representing entity mentions, followed by predicting the relationship between these entities. However, recent advancements have transformed this task into a sequence-to-sequence problem. This involves converting relationships between entities into target string, which are then generated from the input text. Thus, language models now appear as a solution to this task and have already been used in numerous studies, with various levels of refinement, across different domains. The objective of the present study is to evaluate the contribution of large language models (LLM) to the task of relation extraction in a specific domain (in this case, the economic domain), compared to smaller language models. To do this, we considered as a baseline a model based on the BERT architecture, trained in this domain, and four LLM, namely FinGPT specific to the financial domain, XLNet, ChatGLM, and Llama3, which are generalists. All these models were evaluated on the same extraction task, with zero-shot for the general-purpose LLM, as well as refinements through few-shot learning and fine-tuning. The experiments showedthat the best performance in terms of F-score was achieved with fine-tuned LLM, with Llama3 achieving the highest performance. |
nishida-utsuro-2025-generating | 2,025 | Generating Financial News Articles from Factors of Stock Price Rise / Decline by LLMs | In this paper, we study the task of generating financial news articles related to stock price fluctuations. Traditionally, reporters manually write these articles by identifying the causes behind significant stock price volatility. However, this process is time-consuming, limiting the number of articles produced. To address this, the study explores the use of generative AI to automatically generate such articles. The AI system, similar to human reporters, would analyze stock price volatility and determine the underlying factors contributing to these fluctuations. To support this approach, we introduces a Japanese dataset called JFinSR, which includes stock price fluctuation rankings from {\textquotedblleft}Kabutan{\textquotedblright} and related financial information regarding factors of stock price rise / decline from {\textquotedblleft}Nihon Keizai Shimbun (Nikkei).{\textquotedblright} Using this dataset, we implement the few-shot learning technique on large language models (LLMs) to enable automatic generation of high-quality articles from factors of stock price rise / decline that are available in Nikkei. In the evaluation, we compare zero-shot and few-shot learning approaches, where the few-shot learning achieved the higher F1 scores in terms of ROUGE-1/ROUGE-L metrics. |
wang-brorsson-2025-large | 2,025 | Can Large language model analyze financial statements well? | Since GPT-3.5`s release, large language models (LLMs) have made significant advancements, including in financial analysis. However, their effectiveness in financial calculations and predictions is still uncertain. This study examines LLMs' ability to analyze financial reports, focusing on three questions: their accuracy in calculating financial ratios, the use of these metrics in DuPont analysis and the Z-score model for bankruptcy prediction, and their effectiveness in predicting financial indicators with limited knowledge. We used various methods, including zero-shot and few-shot learning, retrieval-augmented generation (RAG), and fine-tuning, in three advanced LLMs and compared their outputs to ground truth and expert predictions to assess their calculation and predictive abilities. The results highlight both the potential and limitations of LLMs in processing numerical data and performing complex financial analyses. |
abdo-etal-2025-amwal | 2,025 | AMWAL: Named Entity Recognition for Arabic Financial News | Financial Named Entity Recognition (NER) presents a pivotal task in extracting structured information from unstructured financial data, especially when extending its application to languages beyond English. In this paper, we present AMWAL, a named entity recognition system for Arabic financial news. Our approach centered on building a specialized corpus compiled from three major Arabic financial newspapers spanning from 2000 to 2023. Entities were extracted from this corpus using a semi-automatic process that included manual annotation and review to ensure accuracy. The total number of entities identified amounts to 17.1k tokens, distributed across 20 categories, providing a comprehensive coverage of financial entities. To standardize the identified entities, we adopt financial concepts from the Financial Industry Business Ontology (FIBO, 2020), aligning our framework with industry standards. The significance of our work lies not only in the creation of the first customized NER system for Arabic financial data but also in its potential to streamline information extraction processes in the financial domain. Our NER system achieves a Precision score of 96.08, a Recall score of 95.87, and an F1 score of 95.97, which outperforms state-of-the-art general Arabic NER systems as well as other systems for financial NER in other languages. |
moreno-sandoval-etal-2025-financial | 2,025 | The Financial Document Causality Detection Shared Task (FinCausal 2025) | We present the Financial Document Causality Detection Task (FinCausal 2025), a multilingual challenge to identify causal relationships within financial texts. This task comprises English and Spanish subtasks, with datasets compiled from British and Spanish annual reports. Participants were tasked with identifying and generating answers to questions about causes or effects within specific text segments. The dataset combines extractive and generative question-answering (QA) methods, with abstractly formulated questions and directly extracted answers from the text. Systems performance is evaluated using exact matching and semantic similarity metrics. The challenge attracted submissions from 10 teams for the English subtask and 10 teams for the Spanish subtask. FinCausal 2025 is part of the 6th Financial Narrative Processing Workshop (FNP 2025), hosted at COLING 2025 in Abu Dhabi. |
shukla-etal-2025-kulfi | 2,025 | KULFi Framework: Knowledge Utilization for Optimizing Large Language Models for Financial Causal Reasoning | This paper presents our contribution to the Financial Document Causality Detection (FinCausal) task 2025. The FinCausal challenge centers on the extraction of cause-and-effect relationships from financial texts written in both English and Spanish. We introduce KULFi, a novel Knowledge Utilization framework designed to augment the capabilities of Large Language Models (LLMs) by leveraging the expertise of more advanced reasoning models. Through the utilization of Teacher LLMs to generate task-specific instructions, KULFi optimizes the performance of Student LLMs via automated prompt optimization. We evaluate the efficacy of KULFi on the Financial Document Causality Detection Task, where Student LLM achieves a similarity score comparable to human-guided prompt optimization for the same LLM, demonstrating significant improvements in causal reasoning performance. Our results demonstrate that KULFi enables effective knowledge transfer from more robust models to less capable ones, as well as efficient learning from training data, minimizing the need for human input in prompt design and enabling more precise causal analysis in financial contexts. Our system attained SAS and Exact Match scores of 0.92 and 0.35 on the English dataset, and 0.92 and 0.09 on the Spanish dataset, respectively. This framework has far-reaching implications, with potential applications in enhancing decision-making across complex financial environments. |
al-laith-2025-exploring | 2,025 | Exploring the Effectiveness of Multilingual and Generative Large Language Models for Question Answering in Financial Texts | This paper investigates the use of large language models (LLMs) for financial causality detection in the FinCausal 2025 shared task, focusing on generative and multilingual question answering (QA) tasks. Our study employed both generative and discriminative approaches, utilizing GPT-4o for generative QA and BERT-base-multilingual-cased, XLM-RoBerta-large, and XLM-RoBerta-base for multilingual QA across English and Spanish datasets. The datasets consist of financial disclosures where questions reflect causal relationships, paired with extractive answers derived directly from the text. Evaluation was conducted using Semantic Answer Similarity (SAS) and Exact Match (EM) metrics. While the discriminative XLM-RoBerta-large model achieved the best overall performance, ranking 5th in English (SAS: 0.9598, EM: 0.7615) and 4th in Spanish (SAS: 0.9756, EM: 0.8084) among 11 team submissions, our results also highlight the effectiveness of the generative GPT-4o approach. Notably, GPT-4o achieved promising results in few-shot settings, with SAS scores approaching those of fine-tuned discriminative models, demonstrating that the generative approach can provide competitive performance despite lacking task-specific fine-tuning. This comparison underscores the potential of generative LLMs as robust, versatile alternatives for complex QA tasks like financial causality detection. |
k-s-etal-2025-clrg | 2,025 | CLRG@FinCausal2025: Cause-Effect Extraction in Finance Domain | This paper presents our work on Cause-Effect information extraction specifically in the financial domain. Cause and effect information is very much needed for expert decision making. Particularly, in the financial domain, the fund managers, financial analysts, etc. need to have the information on cause-effects for their works. Natural Language Processing (NLP) techniques help in the automatic extraction of cause and effect from a given text. In this work, we build various cause-effect text span detection models using pre-trained transformer-based language models and fine tune these models using the data provided by FinCausal 2025 task organizers. We have only used FinCausal 2025 data sets to train our models. No other external data is used. Our ensemble of sequence tagging models based on the Fine-tuned RoBERTa-Large language model achieves SAS score of 0.9604 and Exact match score of 0.7214 for English. Similarly for Spanish we obtain SAS score of 0.9607 and Exact match score of 0.7166. This is our first time participation in the FinCausal 2025 Task. |
trivedi-etal-2025-sarang | 2,025 | Sarang at FinCausal 2025: Contextual QA for Financial Causality Detection Combining Extractive and Generative Models | This paper describes our approach for the FinCausal 2025 English Shared Task, aimed at detecting and extracting causal relationships from the financial text. The task involved answering context-driven questions to identify causes or effects within specified text segments. Our method utilized a consciousAI RoBERTa-base encoder model, fine-tuned on the SQuADx dataset. We further fine-tuned it using the FinCausal 2025 development set. To enhance the quality and contextual relevance of the answers, we passed outputs from the extractive model through Gemma2-9B, a generative large language model, for answer refinement. This hybrid approach effectively addressed the task`s requirements, showcasing the strength of combining extractive and generative models. We (Team name: Sarang) achieved outstanding results, securing 3rd rank with a Semantic Answer Similarity (SAS) score of 96.74{\%} and an Exact Match (EM) score of 70.14{\%}. |
chatwal-etal-2025-enhancing | 2,025 | Enhancing Causal Relationship Detection Using Prompt Engineering and Large Language Models | This paper explores the use of large language models (LLMs) and prompt engineering to detect causal relationships in financial disclosures. The task was part of the FinCausal 2025 shared competition, which focuses on identifying cause-and-effect relationships in financial texts across languages. The study demonstrates the effectiveness of LLMs, specifically LLaMA 3.2, in tackling causality detection in English and Spanish financial reports. The paper introduces various prompt engineering techniques, including zero-shot, few-shot, and chain-of-thought (CoT) prompting, to improve performance. For English, the best results were achieved using the Few-Shot + CoT approach, while for Spanish, the Few-Shot method provided strong semantic alignment despite lower exact match accuracy. The evaluation used two metrics: Exact Match (EM) and Semantic Alignment Score (SAS). The results showed high SAS scores for both languages, indicating good semantic understanding, with English performing particularly well. The study emphasizes the importance of tailored prompt engineering techniques to handle language-specific nuances in financial contexts and suggests future research directions, including fine-tuning LLaMA 3.2 and testing additional LLM architectures to enhance multilingual causality detection in financial texts. |
niess-etal-2025-addressing | 2,025 | Addressing Hallucination in Causal Q\&A: The Efficacy of Fine-tuning over Prompting in LLMs | This paper presents our approach and findings for participating in the FinCausal 2025 competition, which addresses causal question answering derived from financial documents, specifically English and Spanish annual reports. We investigate the effectiveness of generative models, such as Llama, in contrast to common extractive methods like BERT-based token classification. While prompt optimization and few-shot learning offer some improvements, they were insufficient for consistently outperforming extractive methods in FinCausal, suffering from hallucinations. In contrast, fine-tuning generative models was shown to be essential for minimizing hallucinations and achieving superior performance. Using our fine-tuned multilingual model for both tasks, we outperform our extractive and monolingual approaches, achieving top results for Spanish and second-best for English in the competition. Our findings indicate that fine-tuned large language models are well-suited for causal Q{\&}A from complex financial narratives, offering robust multilingual capabilities and effectively mitigating hallucinations. |
jeenoor-etal-2025-presiuniv | 2,025 | PresiUniv at FinCausal 2025 Shared Task: Applying Fine-tuned Language Models to Explain Financial Cause and Effect with Zero-shot Learning | Transformer-based multilingual question-answering models are used to detect causality in financial text data. This study employs BERT (CITATION) for English text and XLM-RoBERTa (CITATION) for Spanish data, which were fine-tuned on the SQuAD datasets (CITATION) (CITATION). These pre-trained models are used to extract answers to the targeted questions. We design a system using these pre-trained models to answer questions, based on the given context. The results validate the effectiveness of the systems in understanding nuanced financial language and offers a tool for multi-lingual text analysis. Our system is able to achieve SAS scores of 0.75 in Spanish and 0.82 in English. |
moreno-aviles-vaca-2025-extracting | 2,025 | Extracting Financial Causality through QA: Insights from FinCausal 2025 Spanish Subtask | The methodology tested both span extraction and generative tasks, with generative models ultimately proving to be more effective. SuperLenia, a private generative model, was the best-performing model. It is a combination of public models with sizes ranging from 7B to 8B parameters. SuperLenia was fine-tuned using QLoRA in a chat-based framework, and hyperparameter tuned during inference, including adjustments to temperature and sampling, further enhanced its performance. |
liu-etal-2025-finnlp | 2,025 | FinNLP-FNP-LLMFinLegal-2025 Shared Task: Financial Misinformation Detection Challenge Task | Despite the promise of large language models (LLMs) in finance, their capabilities for financial misinformation detection (FMD) remain largely unexplored. To evaluate the capabilities of LLMs in FMD task, we introduce the financial misinformation detection shared task featured at COLING FinNLP-FNP-LLMFinLegal-2024, FMD Challenge. This challenge aims to evaluate the ability of LLMs to verify financial misinformation while generating plausible explanations. In this paper, we provide an overview of this task and dataset, summarize participants' methods, and present their experimental evaluations, highlighting the effectiveness of LLMs in addressing the FMD task. To the best of our knowledge, the FMD Challenge is one of the first challenges for assessing LLMs in the field of FMD. Therefore, we provide detailed observations and draw conclusions for the future development of this field. |
luo-etal-2025-fmd | 2,025 | FMD-Mllama at the Financial Misinformation Detection Challenge Task: Multimodal Reasoning and Evidence Generation | This paper presents our system for the Financial Misinformation Detection Challenge Task. We utilize multimodal reasoning, incorporating textual and image information, to address the task. Our system demonstrates the capability to detect financial misinformation while providing comprehensive explanations. Experimental results show that our final system significantly outperforms the baselines and ranks second on the task leaderboard. |
singh-etal-2025-ask | 2,025 | Ask Asper at the Financial Misinformation Detection Challenge Task: Enhancing Financial Decision-Making: A Dual Approach Using Explainable LLMs for Misinformation Detection | The integrity of the market and investor con- fidence are seriously threatened by the prolif- eration of financial misinformation via digital media. Existing approaches such as fact check, lineage detection and others have demonstrated significant progress in detecting financial mis- information. In this paper, we present a novel two-stage framework leveraging large language models (LLMs) to identify and explain finan- cial misinformation. The framework first em- ploys a GPT-4 model fine-tuned on financial datasets to classify claims as {\textquotedblleft}True,{\textquotedblright} {\textquotedblleft}False,{\textquotedblright} or {\textquotedblleft}Not Enough Information{\textquotedblright} by analyzing rel- evant financial context. To enhance classifi- cation reliability, a second LLM serves as a verification layer, examining and refining the initial model`s predictions. This dual-model approach ensures greater accuracy in misinfor- mation detection through cross-validation. Beyond classification, our methodology empha- sizes generating clear, concise, and actionable explanations that enable users to understand the reasoning behind each determination. By com- bining robust misinformation detection with interpretability, our paradigm advances AI sys- tem transparency and accountability, providing valuable support to investors, regulators, and financial stakeholders in mitigating misinfor- mation risks. |
kawamura-2025-team | 2,025 | Team FMD LLM at the Financial Misinformation Detection Challenge Task: Exploring Task Structuring and Metadata Impact on Performance | The detection of financial misinformation (FMD) is a growing challenge. In this paper, we investigate how task structuring and metadata integration impact the performance of large language models (LLMs) on FMD tasks. We compare two approaches: predicting the label before generating an explanation, and generating the explanation first. Our results reveal that prediction-first models achieve higher F1 scores. We also assess the effect of auxiliary metadata, which surprisingly degraded performance despite its correlation with the labels. Our findings highlight the importance of task order and the need to carefully consider whether to use metadata in limited data settings. |
lee-park-2025-dunamu | 2,025 | Dunamu ML at the Financial Misinformation Detection Challenge Task: Improving Supervised Fine-Tuning with LLM-based Data Augmentation | In this paper, we describe Dunamu ML`s submission to the Financial Misinformation Detection (FMD) 2025 shared task. To address the low-resource challenge in FMD, we augmented a general domain misinformation detection dataset for training. We first collected claims, contexts, and misinformation labels from a public dataset. Then, we generated evidence for each label based on a closed LLM with few-shot examples extracted from the FMD training dataset. Finally, we oversampled the training data specific to the financial domain and augmented it with the generated data to perform supervised fine-tuning (SFT) on the LLM. When evaluated on the blind test dataset, our model achieved an F1 score of 84.67 in misinformation classification and a ROUGE-1 score of 81.21 in evidence generation, ranking first on the leaderboard in both aspects. |
purbey-etal-2025-1 | 2,025 | 1-800-SHARED-TASKS at the Financial Misinformation Detection Challenge Task: Sequential Learning for Claim Verification and Explanation Generation in Financial Domains | This paper presents the system description of our entry for the COLING 2025 FMD challenge, focusing on misinformation detection in financial domains. We experimented with a combination of large language models, including Qwen, Mistral, and Gemma-2, and leveraged pre-processing and sequential learning for not only identifying fraudulent financial content but also generating coherent, and concise explanations that clarify the rationale behind the classifications. Our approach achieved competitive results with an F1-score of 0.8283 for classification, and ROUGE-1 of 0.7253 for explanations. This work highlights the transformative potential of LLMs in financial applications, offering insights into their capabilities for combating misinformation and enhancing transparency while identifying areas for future improvement in robustness and domain adaptation. |
dmonte-etal-2025-gmu | 2,025 | GMU-MU at the Financial Misinformation Detection Challenge Task: Exploring LLMs for Financial Claim Verification | This paper describes the team GMU-MU submission to the Financial Misinformation Detection challenge. The goal of this challenge is to identify financial misinformation and generate explanations justifying the predictions by developing or adapting LLMs. The participants were provided with a dataset of financial claims that were categorized into six financial domain categories. We experiment with the Llama model using two approaches; instruction-tuning the model with the training dataset, and a prompting approach that directly evaluates the off-the-shelf model. Our best system was placed 5th among the 12 systems, achieving an overall evaluation score of 0.6682. |
abburi-etal-2025-deloitte | 2,025 | Deloitte (Drocks) at the Financial Misinformation Detection Challenge Task: Enhancing Misinformation Detection through Instruction-Tuned Models | Large Language Models (LLMs) are capable of producing highly fluent and convincing text; however, they can sometimes include factual errors and misleading information. Consequently, LLMs have emerged as tools for the rapid and cost-effective generation of financial misinformation, enabling bad actors to harm individual investors and attempt to manipulate markets. In this study, we instruction-tune Generative Pre-trained Transformers (GPT-4o-mini) to detect financial misinformation and produce concise explanations for why a given claim or statement is classified as misinformation, leveraging the contextual information provided. Our model achieved fourth place in Financial Misinformation Detection (FMD) shared task with a micro F1 score of 0.788 and a ROUGE-1 score of 0.743 on the private test set of FACT-checking within the FINancial domain (FIN-FACT) dataset provided by the shared task organizers. |
cao-etal-2025-capybara | 2,025 | Capybara at the Financial Misinformation Detection Challenge Task: Chain-of-Thought Enhanced Financial Misinformation Detection | Financial misinformation poses a significant threat to investment decisions and market stability. Recently, the application of Large Language Models (LLMs) for detecting financial misinformation has gained considerable attention within the natural language processing (NLP) community. The Financial Misinformation Detection (FMD) challenge @ Coling 2025 serves as a valuable platform for collaboration and innovation. This paper presents our solution to FMD challenge. Our approach involves using search engines to retrieve the summarized high-quality information as supporting evidence and designing a financial domain-specific chain-of-thought to enhance the reasoning capabilities of LLMs. We evaluated our method on both commercial closed-source LLMs (GPT-family) and open-source models (Llama-3.1-8B and QWen). The experimental results domonstrate that the proposed method improves veracity prediction performance. However, the quality of the generated explanations remains relatively poor. In the paper, we present the experimental findings and provides an in depth analysis of these results. |
martinez-etal-2025-scalable | 2,025 | A Scalable Framework for Legal Text Understanding in Regulatory and Financial Contexts. | This study presents a comprehensive approach to developing a domain-specific large language model (LLM) for regulatory and financial text interpretation. A specialized corpus was constructed through large-scale scraping of financial and regulatory documents across domains such as compliance, licensing, and financial reporting. The data was preprocessed using GPT-4o-mini with prompt engineering to retain critical information and remove noise. We further pre-trained a LLaMA-3.1-8B model on the curated corpus and fine-tuned it using an instruction dataset covering nine tasks from the Coling 2025 Regulations Challenge, including acronym expansion, regulatory question-answering, and XBRL-based financial analytics, employing QLoRA to reduce memory requirements. The model exhibits a slight improvement from baseline answering complex regulatory questions (detailed QA) and expanding acronyms. This study demonstrates the potential of domain-specific LLMs in regulatory text interpretation and lays the groundwork for future research in specialized NLP evaluation methodologies. |
huang-etal-2025-audit | 2,025 | Audit-FT at the Regulations Challenge Task: An Open-Source Large Language Model for Audit | Intelligent auditing represents a crucial advancement in modern audit practices, enhancing both the quality and efficiency of audits within the realm of artificial intelligence. With the rise of large language model (LLM), there is enormous potential for intelligent models to contribute to audit domain. However, general LLMs applied in audit domain face the challenges of lacking specialized knowledge and the presence of data biases. To overcome these challenges, this study introduces AuditWen, an open-source audit LLM by fine-tuning Qwen with constructing instruction data from audit domain. We first outline the application scenarios for LLMs in the audit and extract requirements that shape the development of LLMs tailored for audit purposes. We then propose an audit LLM, called AuditWen, by fine-tuning Qwen with constructing 30k instruction dataset from 15 audit tasks and 3 layers. In evaluation stage, we proposed a benchmark with 5k instructions that covers a set of critical audit tasks derived from the application scenarios. With the benchmark, we compare AuditWen with other existing LLMs from information extraction, question answering and document generation. The experimental results demonstrate superior performance of AuditWen both in question understanding and answer generation, making it an immediately valuable tool for audit. |
chantangphol-etal-2025-finmind | 2,025 | FinMind-Y-Me at the Regulations Challenge Task: Financial Mind Your Meaning based on THaLLE | This paper presents our submission to the COLING 2025 regulation challenge, focusing on nine tasks in the regulatory and financial domains. The challenge aims to advance large language models beyond general-purpose capabilities, adapting them for regulatory and financial tasks using a unified framework of task-specific prompts and input templates. We propose a sequential fine-tuning approach that integrates reasoning-based training, tailored system prompts, and Chain-of-Thought (CoT) inference to optimize task-specific performance. This method improves accuracy and reliability across diverse tasks. Notably, CoT inference demonstrates exceptional effectiveness in handling complex scenarios and tasks requiring specific answer patterns, such as named entity recognition and financial calculations. Our model achieved an overall score of 54.801{\%}, ranking 1st among all teams and becoming the top performer in the challenge. These results highlight the effectiveness of sequential fine-tuning, advanced reasoning techniques, and fine-tuned prompts in improving performance and scalability for complex regulatory and financial applications. |
wang-etal-2025-finnlp | 2,025 | FinNLP-FNP-LLMFinLegal-2025 Shared Task: Regulations Challenge | Financial large language models (FinLLMs) have been applied to various tasks in business, finance, accounting, and auditing. Complex financial regulations and standards are critical to financial services, which LLMs must comply with. However, FinLLMs' performance in understanding and interpreting financial regulations has rarely been studied. Therefore, we organize the Regulations Challenge, a shared task at COLING FinNLP-FNP-LLMFinLegal-2025. It encourages the academic community to explore the strengths and limitations of popular LLMs. We create 9 novel tasks and corresponding question sets. In this paper, we provide an overview of these tasks and summarize participants' approaches and results. We aim to raise awareness of FinLLMs' professional capability in financial regulations and industry standards. |
jiang-etal-2025-intellichain | 2,025 | IntelliChain Stars at the Regulations Challenge Task: A Large Language Model for Financial Regulation | We present our approach to the COLING-2025 Regulations Challenge, which evaluates large language models (LLMs) on nine regulatory tasks, such as abbreviation recognition and financial data extraction. To address challenges like domain-specific terminologies and dynamic regulatory contexts, we developed a robust data construction pipeline, integrating proprietary Chinese regulatory data, Fin-GPT datasets, and financial Q{\&}A data. The pipeline applied, but was not limited to, language filtering, semantic screening, and deduplication, resulting in a 30,000-example dataset combining financial regulations and general financial data. Using this dataset, we fine-tuned Llama 3.2-3B-Instruct to create Reg-LLaMA, a specialized model that outperformed baselines on the Regulations Challenge and PIXIU datasets. These results demonstrate the effectiveness of domain-specific data construction in advancing LLMs for regulatory tasks, paving the way for reliable and interpretable AI in regulated industries. |
nararatwong-etal-2025-fin | 2,025 | Fin-DBQA Shared-task: Database Querying and Reasoning | This paper presents the results of the Fin-DBQA shared task based on a question-answering dataset, focusing on database querying and reasoning. The dataset, consisting of 400 questions grouped into 40 conversations, evaluates language models' abilities to answer sequential questions with complex reasoning and multi-hop queries in a multi-turn conversational question-answering setting. Each sample includes the question, answer, database queries, querying result (tables), and a program (series of operations) that produces the answer from the result. We received 52 submissions from three participants, with scores significantly surpassing the baselines. One participant submitted a paper detailing a prompt-based solution using large language models with additional data preprocessing that helps improve the overall performance. |
strich-2025-adapt | 2,025 | Adapt LLM for Multi-turn Reasoning QA using Tidy Data | This paper presents our submission to the Fin-DBQA shared task at the 9th FinNLP workshop. The task involves answering finance-focused questions in a multi-turn environment, requiring step-by-step reasoning and Python code generation. We propose a novel approach to tackle this multidimensional problem by pre-processing the data into tidy data format so that each column represents a variable and each row an observation. Our experiments demonstrate that using the tidy data format allows all models to surpass SOTA, with GPT-4o achieving a 50.62{\%} accuracy on the DBQR-QA benchmark achieving second place on the shared task leaderboard. These findings suggest that transforming data into the tidy data format enhances reasoning capabilities, reduces syntax errors, and improves performance on table-reasoning QA tasks. The code is available online. |
yu-etal-2025-finnlp | 2,025 | FinNLP-FNP-LLMFinLegal @ COLING 2025 Shared Task: Agent-Based Single Cryptocurrency Trading Challenge | Despite the promise of large language models based agent framework in stock trading task, their capabilities for comprehensive analysis and multiple different financial assets remain largely unexplored, such as cryptocurrency trading. To evaluate the capabilities of LLM-based agent framework in cryptocurrency trading, we introduce an LLMs-based financial shared task featured at COLING 2025 FinNLP-FNP-LLMFinLegal workshop, named Agent-based Single Cryptocurrency Trading Challenge. This challenge includes two cryptocurrencies: BitCoin and Ethereum. In this paper, we provide an overview of these tasks and datasets, summarize participants' methods, and present their experimental evaluations, highlighting the effectiveness of LLMs in addressing cryptocurrency trading challenges. To the best of our knowledge, the Agent-based Single Cryptocurrency Trading Challenge is one of the first challenges for assessing LLMs in the financial area. In consequence, we provide detailed observations and take away conclusions for future development in this area. |
wang-etal-2025-sams | 2,025 | Sam`s Fans at the Crypto Trading Challenge Task: A Threshold-Based Decision Approach Based on FinMem Framework | The advancements of large language models (LLMs) demonstrate the value of pre-training on diverse datasets, enabling these models to excel across a wide range of tasks while adapting effectively to specialized applications. This study presents an approach to enhance LLMs' ability to process and trade based on cryptocurrency data across different time horizons. We fine-tuned two established language models, Llama-3.1-8b and Qwen2.5-7b, to effectively interpret and utilize temporal market data provided by the FinMem framework. Our methodology enables these models to analyze multi-period market data from FinMem, including price movements and momentum indicators, to execute effective cryptocurrency trading decisions. Results show that this fine-tuning approach improves the models' capacity to analyze market conditions and inform trading decisions based on multi-period market dynamics. |
agarkov-etal-2025-300k | 2,025 | 300k/ns team at the Crypto Trading Challenge Task: Enhancing the justification of accurate trading decisions through parameter-efficient fine-tuning of reasoning models | In this paper, we address the Agent-Based Sin- gle Cryptocurrency Trading Challenge, focus- ing on decision-making for trading Bitcoin and Etherium. Our approach utilizes fine- tuning a Mistral AI model on a dataset com- prising summarized cryptocurrency news, en- abling it to make informed {\textquotedblleft}buy,{\textquotedblright} {\textquotedblleft}sell,{\textquotedblright} or {\textquotedblleft}hold{\textquotedblright} decisions and articulate its reasoning. The model integrates textual sentiment analysis and contextual reasoning with real-time mar- ket trends, demonstrating the potential of Large Language Models (LLMs) in high-stakes finan- cial decision-making. The model achieved a notable accuracy, highlighting its capacity to manage risk while optimizing returns. This work contributes to advancing AI-driven so- lutions for cryptocurrency markets and offers insights into the practical deployment of LLMs in real-time trading environments. We made our model publicly available. |
raja-etal-2025-dataset | 2,025 | A Dataset for Programming-based Instructional Video Classification and Question Answering | This work aims to develop an understanding of the rapidly emerging field of VideoQA, particularly in the context of instructional programming videos. It also encourages designing of system that can produce visual answer to programming based natural language questions. We introduce two datasets: CodeVidQA, with 2,104 question-answer pair links with timestamps taken from programming videos of Stack Overflow for Programming Visual Answer Localization task, and CodeVidCL with 4,331 videos (1,751 programming ,2580 non-programming) for Programming Video Classification task. In addition, we proposed a framework that adapts BigBird and SVM for video classification techniques. The proposed approach achieves a significantly high accuracy of 99.61{\%} for video classification. |
pirhadi-etal-2025-cvt5 | 2,025 | CVT5: Using Compressed Video Encoder and UMT5 for Dense Video Captioning | The dense video captioning task aims to detect all events occurring in a video and describe each event using natural language. Unlike most other video processing tasks, where it is typically assumed that videos contain only a single main event, this task deals with long, untrimmed videos. Consequently, the speed of processing videos in dense video captioning is a critical aspect of the system. To the best of our knowledge, all published work on this task uses RGB frames to encode input videos. In this work, we introduce the use of compressed videos for the first time in this task. Our experiments on the SoccerNet challenge demonstrate significant improvements in both processing speed and GPU memory footprint while achieving competitive results. Additionally, we leverage multilingual transcripts, which seems to be effective. The encoder in our proposed method achieves approximately 5.4{\texttimes} higher speed and 5.1{\texttimes} lower GPU memory usage during training, and 4.7{\texttimes} higher speed and 7.8{\texttimes} lower GPU memory usage during inference, compared to its RGB-based counterpart. The code is publicly available at https://github.com/mohammadjavadpirhadi/CVT5. |
bai-pezzelle-2025-feel | 2,025 | If I feel smart, I will do the right thing: Combining Complementary Multimodal Information in Visual Language Models | Generative visual language models (VLMs) have recently shown potential across various downstream language-and-vision tasks. At the same time, it is still an open question whether, and to what extent, these models can properly understand a multimodal context where language and vision provide complementary information{---}a mechanism routinely in place in human language communication. In this work, we test various VLMs on the task of generating action descriptions consistent with both an image`s visual content and an intention or attitude (not visually grounded) conveyed by a textual prompt. Our results show that BLIP-2 is not far from human performance when the task is framed as a generative multiple-choice problem, while other models struggle. Furthermore, the actions generated by BLIP-2 in an open-ended generative setting are better than those by the competitors; indeed, human annotators judge most of them as plausible continuations for the multimodal context. Our study reveals substantial variability among VLMs in integrating complementary multimodal information, yet BLIP-2 demonstrates promising trends across most evaluations, paving the way for seamless human-computer interaction. |
sun-etal-2025-llava | 2,025 | LLaVA-RE: Binary Image-Text Relevancy Evaluation with Multimodal Large Language Model | Multimodal generative AI usually involves generating image or text responses given inputs in another modality. The evaluation of image-text relevancy is essential for measuring the response quality or ranking candidate responses. In particular, binary relevancy evaluation, i.e., {\textquotedblleft}Relevant{\textquotedblright} vs. {\textquotedblleft}Not Relevant{\textquotedblright}, is a fundamental problem. However, this is a challenging task considering that texts have diverse formats and the definition of relevancy varies in different scenarios. We find that Multimodal Large Language Models (MLLMs) are an ideal choice to build such evaluators, as they can flexibly handle complex text formats and take in additional task information. In this paper, we present LLaVA-RE, a first attempt for binary image-text relevancy evaluation with MLLM. It follows the LLaVA architecture and adopts detailed task instructions and multimodal in-context samples. Further, we propose a novel binary relevancy dataset covering diverse tasks. Experimental results validate the effectiveness of our framework. |
farsi-etal-2025-persian | 2,025 | Persian in a Court: Benchmarking VLMs In Persian Multi-Modal Tasks | This study introduces a novel framework for evaluating Large Language Models (LLMs) and Vision-Language Models (VLMs) in Persian, a low-resource language. We develop comprehensive datasets to assess reasoning, linguistic understanding, and multimodal capabilities. Our datasets include Persian-OCR-QA for optical character recognition, Persian-VQA for visual question answering, Persian world-image puzzle for multimodal integration, Visual-Abstraction-Reasoning for abstract reasoning, and Iran-places for visual knowledge of Iranian figures and locations. We evaluate models like GPT-4o, Claude 3.5 Sonnet, and Llama 3.2 90B Vision, revealing their strengths and weaknesses in processing Persian. This research contributes to inclusive language processing by addressing the unique challenges of low-resource language evaluation. |
hsieh-etal-2025-taiwanvqa | 2,025 | TaiwanVQA: A Benchmark for Visual Question Answering for Taiwanese Daily Life | We introduce TaiwanVQA, a novel visual question answering benchmark designed to evaluate vision language models' (VLMs) ability to recognize and reason about Taiwan-specific multimodal content.TaiwanVQA comprises 2,000 image-question pairs covering diverse topics relevant to Taiwanese culture and daily life. We categorize the questions into recognition and reasoning tasks, further sub-classifying reasoning questions based on the level of external knowledge required. We conduct extensive experiments on state-of-the-art VLMs, including GPT-4o, Llama-3.2, LLaVA, Qwen2-VL, and InternVL2 models. Our findings reveal significant limitations in current VLMs when handling culturally specific content. The performance gap widens between recognition tasks (top score 73.60{\%}) and reasoning tasks (top score 49.80{\%}), indicating challenges in cultural inference and contextual understanding.These results highlight the need for more culturally diverse training data and improved model architectures that can better integrate visual and textual information within specific cultural contexts. By providing TaiwanVQA, we aim to contribute to the development of more inclusive and culturally aware AI models, facilitating their deployment in diverse real-world settings. TaiwanVQA can be accessed on our GitHub page. |
sinha-etal-2025-guiding | 2,025 | Guiding Vision-Language Model Selection for Visual Question-Answering Across Tasks, Domains, and Knowledge Types | Visual Question-Answering (VQA) has become key to user experience, particularly after improved generalization capabilities of Vision-Language Models (VLMs). But evaluating VLMs for an application requirement using a standardized framework in practical settings is still challenging. This paper aims to solve that using an end-to-end framework. We present VQA360 - a novel dataset derived from established VQA benchmarks, annotated with task types, application domains, and knowledge types, for a comprehensive evaluation. We also introduce GoEval, a multimodal evaluation metric developed using GPT-4o, achieving a correlation factor of 56.71{\%} with human judgments. Our experiments with state-of-the-art VLMs reveal that no single model excels universally, thus, making a right choice a key design decision. Proprietary models such as Gemini-1.5-Pro and GPT-4o-mini generally outperform others, but open-source models like InternVL-2-8B and CogVLM-2-Llama-3-19B also demonstrate competitive strengths, while providing additional advantages. Our framework can also be extended to other tasks. |
rizzi-etal-2025-bunch | 2,025 | Is a bunch of words enough to detect disagreement in hateful content? | The complexity of the annotation process when adopting crowdsourcing platforms for labeling hateful content can be linked to the presence of textual constituents that can be ambiguous, misinterpreted, or characterized by a reduced surrounding context. In this paper, we address the problem of perspectivism in hateful speech by leveraging contextualized embedding representation of their constituents and weighted probability functions. The effectiveness of the proposed approach is assessed using four datasets provided for the SemEval 2023 Task 11 shared task. The results emphasize that a few elements can serve as a proxy to identify sentences that may be perceived differently by multiple readers, without the need of necessarily exploiting complex Large Language Models. |
yung-demberg-2025-crowdsourcing | 2,025 | On Crowdsourcing Task Design for Discourse Relation Annotation | Interpreting implicit discourse relations involves complex reasoning, requiring the integration of semantic cues with background knowledge, as overt connectives like {\textquotedblleft}because{\textquotedblright} or {\textquotedblleft}then{\textquotedblright} are absent. These relations often allow multiple interpretations, best represented as distributions. In this study, we compare two established methods that crowdsource implicit discourse relation annotation by connective insertion: a free-choice approach, which allows annotators to select any suitable connective, and a forced-choice approach, which asks them to select among a set of predefined options. Specifically, we re-annotate the whole DiscoGeM 1.0 corpus - initially annotated with the free-choice method - using the forced-choice approach. The free-choice approach allows for flexible and intuitive insertion of various connectives, which are context-dependent. Comparison among over 130,000 annotations, however, shows that the free-choice strategy produces less diverse annotations, often converging on common labels. Analysis of the results reveals the interplay between task design and the annotators' abilities to interpret and produce discourse relations. |
dsouza-kovatchev-2025-sources | 2,025 | Sources of Disagreement in Data for LLM Instruction Tuning | In this paper we study the patterns of label disagreement in data used for instruction tuning Large Language models (LLMs). Specifically, we focus on data used for Reinforcement Learning from Human Feedback (RLHF). Our objective is to determine what is the primary source of disagreement: the individual data points, the choice of annotators, or the task formulation. We annotate the same dataset multiple times under different conditions and compare the overall agreement and the patterns of disagreement. For task formulation, we compare {\textquotedblleft}single{\textquotedblright} format where annotators rate LLM responses individually with {\textquotedblleft}preference{\textquotedblright} format where annotators select one of two possible responses. For annotators, we compare data from human labelers with automatic data labeling using LLMs. Our results indicate that: (1) there are very few {\textquotedblleft}universally ambiguous{\textquotedblright} instances. The label disagreement depends largely on the task formulation and the choice of annotators; (2) the overall agreement remains consistent across experiments. We find no evidence that {\textquotedblleft}preference{\textquotedblright} data is of higher quality than {\textquotedblleft}single{\textquotedblright} data; and (3) the change of task formulation and annotators impacts the resulting instance-level labels. The labels obtained in different experiments are correlated, but not identical. |
schlechtweg-etal-2025-comedi | 2,025 | CoMeDi Shared Task: Median Judgment Classification \& Mean Disagreement Ranking with Ordinal Word-in-Context Judgments | We asked task participants to solve two subtasks given a pair of word usages: Ordinal Graded Word-in-Context Classification (OGWiC) and Disagreement in Word-in-Context Ranking (DisWiC). The tasks take a different view on modeling of word meaning by (i) treating WiC as an ordinal classification task, and (ii) making disagreement the explicit detection aim (instead of removing it). OGWiC is solved with relatively high performance while DisWiC proves to be a challenging task. In both tasks, the dominating model architecture uses independently optimized binary Word-in-Context models. |
kuklin-arefyev-2025-deep | 2,025 | Deep-change at CoMeDi: the Cross-Entropy Loss is not All You Need | Manual annotation of edges in Diachronic Word Usage Graphs is a critical step in creation of datasets for Lexical Semantic Change Detection tasks, but a very labour-intensive one. Annotators estimate if two senses of an ambiguous word expressed in two usages of this word are related and how. This is a variation of the Word-in-Context (WiC) task with some peculiarities, including diachronic data, an ordinal scale for annotations consisting of 4 values with pre-defined meanings (e.g. homonymy, polysemy), and special attention to the degree of disagreement between annotators which affects the further processing of the graph. CoMeDi is a shared task aiming at automating this annotation process. Participants are asked to predict the median annotation for a pair of usages in the first subtask, and estimate the disagreement between annotators in the second subtask. Together this gives some idea about the distribution of annotations we can get from humans for a given pair of usages. For the first subtask we tried several ways of adapting a binary WiC model to this 4 class problem. We discovered that further fine-tuning the model as a 4 class classifier on the training data of the shared task works significantly worse than thresholding the original binary model. For the second subtask our best results were achieved by building a model that predicts the whole multinomial distribution of annotations and calculating the disagreement from this distribution. Our solutions for both subtasks have outperformed all other participants of the shared task. |
choppa-etal-2025-predicting | 2,025 | Predicting Median, Disagreement and Noise Label in Ordinal Word-in-Context Data | The quality of annotated data is crucial for Machine Learning models, particularly in word sense annotation in context (Word-in-Context, WiC). WiC datasets often show significant annotator disagreement, and information is lost when creating gold labels through majority or median aggregation. Recent work has addressed this by incorporating disagreement data through new label aggregation methods. Modeling disagreement is important since real-world scenarios often lack clean data and require predictions on inherently difficult samples. Disagreement prediction can help detect complex cases or to reflect inherent data ambiguity. We aim to model different aspects of ordinal Word-in-Context annotations necessary to build a more human-like model: (i) the aggregated label, which has traditionally been the modeling aim, (ii) the disagreement between annotators, and (iii) the aggregated noise label which annotators can choose to exclude data points from annotation. We find that disagreement and noise are impacted by various properties of data like ambiguity, which in turn points to data uncertainty. |
alfter-appelgren-2025-grasp | 2,025 | GRASP at CoMeDi Shared Task: Multi-Strategy Modeling of Annotator Behavior in Multi-Lingual Semantic Judgments | This paper presents the GRASP team`s systems for the CoMeDi 2025 shared task on disagreement prediction in semantic annotation. The task comprises two subtasks: predicting median similarity scores and mean disagreement scores for word usage across multiple languages including Chinese, English, German, Norwegian, Russian, Spanish, and Swedish. For subtask 1, we implement three approaches: Prochain, a probabilistic chain model predicting sequential judgments; FARM, an ensemble of five fine-tuned XLM-RoBERTa models; and THAT, a task-specific model using XL-Lexeme with adaptive thresholds. For subtask 2, we develop three systems: LAMP, combining language-agnostic and monolingual models; BUMBLE, using optimal language combinations; and DRAMA, leveraging disagreement patterns from FARM`s outputs. Our results show strong performance across both subtasks, ranking second overall among participating teams. The probabilistic Prochain model demonstrates surprisingly robust performance when given accurate initial judgments, while our task-specific approaches show varying effectiveness across languages. |
sarumi-etal-2025-funzac | 2,025 | Funzac at CoMeDi Shared Task: Modeling Annotator Disagreement from Word-In-Context Perspectives | In this work, we evaluate annotator disagreement in Word-in-Context (WiC) tasks exploring the relationship between contextual meaning and disagreement as part of the CoMeDi shared task competition. While prior studies have modeled disagreement by analyzing annotator attributes with single-sentence inputs, this shared task incorporates WiC to bridge the gap between sentence-level semantic representation and annotator judgment variability. We describe three different methods that we developed for the shared task, including a feature enrichment approach that combines concatenation, element-wise differences, products, and cosine similarity, Euclidean and Manhattan distances to extend contextual embedding representations, a transformation by Adapter blocks to obtain task-specific representations of contextual embeddings, and classifiers of varying complexities, including ensembles. The comparison of our methods demonstrates improved performance for methods that include enriched and task-specfic features. While the performance of our method falls short in comparison to the best system in subtask 1 (OGWiC), it is competitive to the official evaluation results in subtask 2 (DisWiC) |
chu-2025-fuocchuvip123 | 2,025 | FuocChuVIP123 at CoMeDi Shared Task: Disagreement Ranking with XLM-Roberta Sentence Embeddings and Deep Neural Regression | This paper presents results of our system for CoMeDi Shared Task, focusing on Subtask 2: Disagreement Ranking. Our system leverages sentence embeddings generated by the paraphrase-xlm-r-multilingual-v1 model, combined with a deep neural regression model incorporating batch normalization and dropout for improved generalization. By predicting the mean of pairwise judgment differences between annotators, our method explicitly targets disagreement ranking, diverging from traditional {\textquotedblleft}gold label{\textquotedblright} aggregation approaches. We optimized our system with a tailored architecture and training procedure, achieving competitive performance in Spearman correlation against the mean disagreement labels. Our results highlights the importance of robust embeddings, effective model architecture, and careful handling of judgment differences for ranking disagreement in multilingual contexts. These findings provide insights into leveraging contextualized representations for ordinal judgment tasks and open avenues for further refinement in disagreement prediction models. |
liu-etal-2025-juniperliu | 2,025 | JuniperLiu at CoMeDi Shared Task: Models as Annotators in Lexical Semantics Disagreements | We present the results of our system for the CoMeDi Shared Task, which predicts majority votes (Subtask 1) and annotator disagreements (Subtask 2). Our approach combines model ensemble strategies with MLP-based and threshold-based methods trained on pretrained language models. Treating individual models as virtual annotators, we simulate the annotation process by designing aggregation measures that incorporate continuous relatedness scores and discrete classification labels to capture both majority and disagreement. Additionally, we employ anisotropy removal techniques to enhance performance. Experimental results demonstrate the effectiveness of our methods, particularly for Subtask 2. Notably, we find that standard deviation on continuous relatedness scores among different model manipulations correlates with human disagreement annotations compared to metrics on aggregated discrete labels. The code will be published at https://github.com/RyanLiut/CoMeDi{\_}Solution |
le-van-2025-mmlabuit | 2,025 | MMLabUIT at CoMeDiShared Task: Text Embedding Techniques versus Generation-Based NLI for Median Judgment Classification | This paper presents our approach in the COLING2025-CoMeDi task in 7 languages, focusing on sub-task 1: Median Judgment Classification with Ordinal Word-in-Context Judgments (OGWiC). Specifically, we need to determine the meaning relation of one word in two different contexts and classify the input into 4 labels. To address sub-task 1, we implement and investigate various solutions, including (1) Stacking, Averaged Embedding techniques with a multilingual BERT-based model; and (2) utilizing a Natural Language Inference approach instead of a regular classification process. All the experiments were conducted on the P100 GPU from the Kaggle platform. To enhance the context of input, we perform Improve Known Data Rate and Text Expansion in some languages. For model focusing purposes Custom Token was used in the data processing pipeline. Our best official results on the test set are 0.515, 0.518, and 0.524 in terms of Krippendorff`s {\ensuremath{\alpha}} score on task 1. Our participation system achieved a Top 3 ranking in task 1. Besides the official result, our best approach also achieved 0.596 regarding Krippendorff`s {\ensuremath{\alpha}} score on Task 1. |
loke-etal-2025-abdn | 2,025 | ABDN-NLP at CoMeDi Shared Task: Predicting the Aggregated Human Judgment via Weighted Few-Shot Prompting | Human annotation is notorious for being subjective and expensive. Recently, (CITATION) introduced the CoMeDi shared task aiming to address this issue by predicting human annotations on the semantic proximity between word uses, and estimating the variation of the human annotations. However, distinguishing the proximity between word uses can be challenging, when their semantic difference is subtle. In this work, we focus on predicting the aggregated annotator judgment of semantic proximity by using a large language model fine-tuned on 20 examples with various proximity classes. To distinguish nuanced proximity, we propose a weighted few-shot approach that pays greater attention to the proximity classes identified as important during fine-tuning. We evaluate our approach in the CoMeDi shared task across 7 languages. Our results demonstrate the superiority of our approach over zero-shot and standard few-shot counterparts. While useful, the weighted few-shot should be applied with caution, given that it relies on development sets to compute the importance of proximity classes, and thus may not generalize well to real-world scenarios where the distribution of class importance is different. |
bibal-etal-2025-automating | 2,025 | Automating Annotation Guideline Improvements using LLMs: A Case Study | Annotating texts can be a tedious task, especially when texts are noisy. At the root of the issue, guidelines are not always optimized enough to be able to perform the required annotation task. In difficult cases, complex workflows are designed to be able to reach the best possible guidelines. However, crowdsource workers are commonly recruited to go through these complex workflows, limiting the number of iterations over the workflows, and therefore, the possible results because of the slow speed and the high cost of workers. In this paper, our case study, based on the entity recognition problem, suggests that LLMs can help produce guidelines of high quality (inter-annotator agreement going from 0.593 to 0.84 when improving WNUT-17`s guidelines), while being faster and cheaper than crowdsource workers. |
wein-2025-ambiguity | 2,025 | Ambiguity and Disagreement in Abstract Meaning Representation | Abstract Meaning Representation (AMR) is a graph-based semantic formalism which has been incorporated into a number of downstream tasks related to natural language understanding. Recent work has highlighted the key, yet often ignored, role of ambiguity and implicit information in natural language understanding. As such, in order to effectively leverage AMR in downstream applications, it is imperative to understand to what extent and in what ways ambiguity affects AMR graphs and causes disagreement in AMR annotation. In this work, we examine the role of ambiguity in AMR graph structure by employing a taxonomy of ambiguity types and producing AMRs affected by each type. Additionally, we investigate how various AMR parsers handle the presence of ambiguity in sentences. Finally, we quantify the impact of ambiguity on AMR using disambiguating paraphrases at a larger scale, and compare this to the measurable impact of ambiguity in vector semantics. |
sanchez-montero-etal-2025-disagreement | 2,025 | Disagreement in Metaphor Annotation of Mexican Spanish Science Tweets | Traditional linguistic annotation methods often strive for a gold standard with hard labels as input for natural language processing models, assuming an underlying objective truth for all tasks. However, disagreement among annotators is a common scenario, even for seemingly objective linguistic tasks, and is particularly prominent in figurative language annotation, since multiple valid interpretations can sometimes coexist. This study presents the annotation process for identifying metaphorical tweets within a corpus of 3733 Public Communication of Science texts written in Mexican Spanish, emphasizing inter-annotator disagreement. Using Fleiss' and Cohen`s Kappa alongside agreement percentages, we evaluated metaphorical language detection through binary classification in three situations: two subsets of the corpus labeled by three different non-expert annotators each, and a subset of disagreement tweets, identified in the non-expert annotation phase, re-labeled by three expert annotators. Our results suggest that expert annotation may improve agreement levels, but does not exclude disagreement, likely due to factors such as the relatively novelty of the genre, the presence of multiple scientific topics, and the blending of specialized and non-specialized discourse. Going further, we propose adopting a learning-from-disagreement approach for capturing diverse annotation perspectives to enhance computational metaphor detection in Mexican Spanish. |
fu-etal-2025-preact | 2,025 | PreAct: Prediction Enhances Agent`s Planning Ability | Addressing the disparity between predictions and actual results can enable individuals to expand their thought processes and stimulate self-reflection, thus promoting accurate planning. In this research, we present **PreAct**, an agent framework that integrates **pre**diction, **rea**soning, and **act**ion. By utilizing the information derived from predictions, the large language model (LLM) agent can provide a wider range and more strategically focused reasoning. This leads to more efficient actions that aid the agent in accomplishing intricate tasks. Our experimental results show that PreAct surpasses the ReAct method in completing complex tasks and that PreAct`s performance can be further improved when paired with other memory or selection strategy techniques. We presented the model with varying quantities of historical predictions and discovered that these predictions consistently enhance LLM planning. The variances in single-step reasoning between PreAct and ReAct indicate that PreAct indeed has benefits in terms of diversity and strategic orientation over ReAct. |
alvarez-ojeda-etal-2025-precom | 2,025 | The PRECOM-SM Corpus: Gambling in Spanish Social Media | Gambling addiction is a {\textquotedblleft}silent problem{\textquotedblright} in society, especially among young people in recent years due to the easy access to betting and gambling sites on the Internet through smartphones and personal computers. As online communities in messaging apps, forums and other {\textquotedblleft}teenagers gathering{\textquotedblright} sites keep growing day by day, more textual information is available for its study. This work focuses on collecting text from online Spanish-speaking communities and analysing it in order to find patterns in written language from frequent and infrequent users on the collected platforms so that an emerging gambling addiction problem can be detected. In this paper, a newly built corpus is introduced, as well as an extensive description of how it has been made. Besides, some baseline experiments on the data have been carried on, employing the generated features after the analysis of the text with different machine learning approaches like the bag of words model or deep neural network encodings. |
huang-2025-well | 2,025 | How Well Can a Long Sequence Model Model Long Sequences? Comparing Architectural Inductive Biases on Long-Context Abilities | Long sequences occur in abundance within real-world scenarios, hence properly modelling them opens numerous down-stream use-cases. Deep neural networks, however, have often struggled with these for a variety of reasons. Recent advances, both in system engineering as well as model design, have enabled the scaling up of model that are purported to support extended context length. In particular, the state-space and linear recurrent neural network families of models hypothetically can entend to infinite sequence length. However, is this too good to be true? We conduct an evaluation to show that while such claims may be sound theoretically, there remain large practical gaps that are empirically observed. In particular, recurrent models still suffer in the same settings as long-context LLMs with attention. We further show that different inductive biases have inconsistent extrapolation capabilities, highlighting the need to further study such paradigms and investigate why long-context models seemingly fail to behave as one might expect. |
sun-tian-2025-sequential | 2,025 | Sequential Fusion of Text-close and Text-far Representations for Multimodal Sentiment Analysis | Multimodal Sentiment Analysis (MSA) aims to identify human attitudes from diverse modalities such as visual, audio and text modalities. Recent studies suggest that the text modality tends to be the most effective, which has encouraged models to consider text as its core modality. However, previous methods primarily concentrate on projecting modalities other than text into a space close to the text modality and learning an identical representation, which does not fully make use of the auxiliary information provided by audio and visual modalities. In this paper, we propose a framework, Sequential Fusion of Text-close and Text-far Representations (SFTTR), aiming to refine multimodal representations from multimodal data which should contain both representations close to and far from the text modality. Specifically, we employ contrastive learning to sufficiently explore the information similarities and differences between text and audio/visual modalities. Moreover, to fuse the extracted representations more effectively, we design a sequential cross-modal encoder to sequentially fuse representations that are close to and far from the text modality. |
huang-shen-2025-poembert | 2,025 | PoemBERT: A Dynamic Masking Content and Ratio Based Semantic Language Model For Chinese Poem Generation | Ancient Chinese poetry stands as a crucial treasure in Chinese culture. To address the absence of pre-trained models for ancient poetry, we introduced PoemBERT, a BERT-based model utilizing a corpus of classical Chinese poetry. Recognizing the unique emotional depth and linguistic precision of poetry, we incorporated sentiment and pinyin embeddings into the model, enhancing its sensitivity to emotional information and addressing challenges posed by the phenomenon of multiple pronunciations for the same Chinese character. Additionally, we proposed Character Importance-based masking and dynamic masking strategies, significantly augmenting the model`s capability to extract imagery-related features and handle poetry-specific information. Fine-tuning our PoemBERT model on various downstream tasks, including poem generation and sentiment classification, resulted in state-of-the-art performance in both automatic and manual evaluations. We provided explanations for the selection of the dynamic masking rate strategy and proposed a solution to the issue of a small dataset size. |
xin-etal-2025-cda | 2,025 | CDA\textasciicircum2: Counterfactual Diffusion Augmentation for Cross-Domain Adaptation in Low-Resource Sentiment Analysis | Domain adaptation is widely employed in cross-domain sentiment analysis, enabling the transfer of models from label-rich source domains to target domain with fewer or no labels. However, concerns have been raised regarding their robustness and sensitivity to data distribution shift, particularly when encountering significant disparities in data distribution between the different domains. To tackle this problem, we introduce a framework CDA{\textasciicircum}2 for cross-domain adaptation in low-resource sentiment analysis, which utilizes counterfactual diffusion augmentation. Specifically, it employs samples derived from domain-relevant word substitutions in source domain samples to guide the diffusion model for generating high-quality counterfactual target domain samples. We adopt a soft absorbing state and MMD loss during the training stage, and use advanced ODE solvers to expedite the sampling process. Our experiments demonstrate that CDA{\textasciicircum}2 generates high-quality target samples and achieves state-of-the-art performance in cross-domain sentiment analysis. |
zhao-etal-2025-codejudge | 2,025 | CodeJudge-Eval: Can Large Language Models be Good Judges in Code Understanding? | Recent advancements in large language models (LLMs) have showcased impressive code generation capabilities, primarily evaluated through language-to-code benchmarks. However, these benchmarks may not fully capture a model`s code understanding abilities. We introduce CodeJudge-Eval (CJ-Eval), a novel benchmark designed to assess LLMs' code understanding abilities from the perspective of code judging rather than code generation. CJ-Eval challenges models to determine the correctness of provided code solutions, encompassing various error types and compilation issues. By leveraging a diverse set of problems and a fine-grained judging system, CJ-Eval addresses the limitations of traditional benchmarks, including the potential memorization of solutions. Evaluation of 12 well-known LLMs on CJ-Eval reveals that even state-of-the-art models struggle, highlighting the benchmark`s ability to probe deeper into models' code understanding abilities. Our benchmark is available at https://github.com/CodeLLM-Research/CodeJudge-Eval . |
wang-etal-2025-match | 2,025 | Match, Compare, or Select? An Investigation of Large Language Models for Entity Matching | Entity matching (EM) is a critical step in entity resolution (ER). Recently, entity matching based on large language models (LLMs) has shown great promise. However, current LLM-based entity matching approaches typically follow a binary matching paradigm that ignores the global consistency among record relationships. In this paper, we investigate various methodologies for LLM-based entity matching that incorporate record interactions from different perspectives. Specifically, we comprehensively compare three representative strategies: matching, comparing, and selecting, and analyze their respective advantages and challenges in diverse scenarios. Based on our findings, we further design a compound entity matching framework (ComEM) that leverages the composition of multiple strategies and LLMs. ComEM benefits from the advantages of different sides and achieves improvements in both effectiveness and efficiency. Experimental results on 8 ER datasets and 10 LLMs verify the superiority of incorporating record interactions through the selecting strategy, as well as the further cost-effectiveness brought by ComEM. |
deng-etal-2025-instructgec | 2,025 | InstructGEC: Enhancing Unsupervised Grammatical Error Correction with Instruction Tuning | Recent works have proposed methods of generating synthetic data automatically for unsupervised Grammatical Error Correction (GEC). Although a large amount of synthetic data is generated at a low cost, it is unrealistic and of poor quality. The copying phenomenon of synthetic data prevents GEC models from learning the semantic knowledge of contextual language. In this paper, we design an instruction format and use the masking strategy in both an erroneous sentence and the corresponding instruction consistently to alleviate the impact of the copy phenomenon. We also propose a novel approach, InstructGEC, which integrates the knowledge of grammatical detection into GEC models with instruction tuning to address the low-quality issue. Experiments are conducted on English and Chinese GEC datasets and results demonstrate that our method outperforms state-of-the-art unsupervised GEC methods. |
wang-etal-2025-sibyl | 2,025 | Sibyl: Empowering Empathetic Dialogue Generation in Large Language Models via Sensible and Visionary Commonsense Inference | Recently, there has been a heightened interest in building chatbots based on Large Language Models (LLMs) to emulate human-like qualities in multi-turn conversations. Despite having access to commonsense knowledge to better understand the psychological aspects and causality of dialogue context, even these powerful LLMs struggle to achieve the goals of empathy and emotional support. Current commonsense knowledge derived from dialogue contexts is inherently limited and often fails to adequately anticipate the future course of a dialogue. This lack of foresight can mislead LLMs and hinder their ability to provide effective support. In response to this challenge, we present an innovative framework named Sensible and Visionary Commonsense Knowledge (Sibyl). Designed to concentrate on the immediately succeeding dialogue, this paradigm equips LLMs with the capability to uncover the implicit requirements of the conversation, aiming to elicit more empathetic responses. Experimental results demonstrate that incorporating our paradigm for acquiring commonsense knowledge into LLMs comprehensively enhances the quality of their responses. |
chen-etal-2025-noise | 2,025 | Noise-powered Multi-modal Knowledge Graph Representation Framework | The rise of Multi-modal Pre-training highlights the necessity for a unified Multi-Modal Knowledge Graph (MMKG) representation learning framework. Such a framework is essential for embedding structured knowledge into multi-modal Large Language Models effectively, alleviating issues like knowledge misconceptions and multi-modal hallucinations. In this work, we explore the efficacy of models in accurately embedding entities within MMKGs through two pivotal tasks: Multi-modal Knowledge Graph Completion (MKGC) and Multi-modal Entity Alignment (MMEA). Building on this foundation, we propose a novel SNAG method that utilizes a Transformer-based architecture equipped with modality-level noise masking to robustly integrate multi-modal entity features in KGs. By incorporating specific training objectives for both MKGC and MMEA, our approach achieves SOTA performance across a total of ten datasets, demonstrating its versatility. Moreover, SNAG can not only function as a standalone model but also enhance other existing methods, providing stable performance improvements. Code and data are available at https://github.com/zjukg/SNAG. |
ye-etal-2025-tooleyes | 2,025 | ToolEyes: Fine-Grained Evaluation for Tool Learning Capabilities of Large Language Models in Real-world Scenarios | Existing evaluations of tool learning primarily focus on validating the alignment of selected tools for large language models (LLMs) with expected outcomes. However, these approaches rely on a limited set of scenarios where answers can be pre-determined. Furthermore, a sole emphasis on outcomes disregards the complex capabilities required for LLMs to effectively use tools. To tackle this issue, we propose ToolEyes, a fine-grained system tailored for the evaluation of the LLMs' tool learning capabilities in authentic scenarios. The system meticulously examines seven real-world scenarios, analyzing five dimensions crucial to LLMs in tool learning: format alignment, intent comprehension, behavior planning, tool selection, and answer organization. Additionally, ToolEyes incorporates a tool library boasting approximately 600 tools, serving as an intermediary between LLMs and the physical world. Evaluations involving ten LLMs across three categories reveal a preference for specific scenarios and limited cognitive abilities in tool learning. Intriguingly, expanding the model size even exacerbates the hindrance to tool learning. The code and data are available at https://github.com/Junjie-Ye/ToolEyes. |
liu-etal-2025-federated | 2,025 | Federated Incremental Named Entity Recognition | Federated learning-based Named Entity Recognition (FNER) has attracted widespread attention through decentralized training on local clients. However, most FNER models assume that entity types are pre-fixed, so in practical applications, local clients constantly receive new entity types without enough storage to access old entity types, resulting in severe forgetting on previously learned knowledge. In addition, new clients collecting only new entity types may join the global training of FNER irregularly, further exacerbating catastrophic forgetting. To overcome the above challenges, we propose a Forgetting-Subdued Learning (FSL) model which solves the forgetting problem on old entity types from both intra-client and inter-client two aspects. Specifically, for intra-client aspect, we propose a prototype-guided adaptive pseudo labeling and a prototypical relation distillation loss to surmount catastrophic forgetting of old entity types with semantic shift. Furthermore, for inter-client aspect, we propose a task transfer detector. It can identify the arrival of new entity types that are protected by privacy and store the latest old global model for relation distillation. Qualitative experiments have shown that our model has made significant improvements compared to several baseline methods. |
Subsets and Splits