Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SeQwen at the Financial Misinformation Detection Challenge Task: Sequential Learning for Claim Verification and Explanation Generation in Financial Domains

Jebish Purbey, Siddhant Gupta, Nikhil Manali, Siddartha Pullakhandam, Drishti Sharma, Ashay Srivastava, Ram Mohan Rao Kadiyala

TL;DR

This work develops SeQwen, a sequential fine-tuning framework for financial misinformation detection that first specializes a lightweight LLM for claim classification and then adapts it to jointly produce explanations. Using the FIN-Fact-derived dataset and LoRA-based low-rank adaptation, the authors demonstrate that sequential learning yields higher classification accuracy and richer explanations than single-stage training, with Qwen2.5 7B achieving strong results (test F1 0.8283; ROUGE-1 0.7253; ROUGE-2 0.6763; ROUGE-L 0.6911). The study highlights the value of task-specific adaptation and interpretability in high-stakes financial contexts, while acknowledging computational constraints and the need for further robustness and domain adaptation research.

Abstract

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.

SeQwen at the Financial Misinformation Detection Challenge Task: Sequential Learning for Claim Verification and Explanation Generation in Financial Domains

TL;DR

This work develops SeQwen, a sequential fine-tuning framework for financial misinformation detection that first specializes a lightweight LLM for claim classification and then adapts it to jointly produce explanations. Using the FIN-Fact-derived dataset and LoRA-based low-rank adaptation, the authors demonstrate that sequential learning yields higher classification accuracy and richer explanations than single-stage training, with Qwen2.5 7B achieving strong results (test F1 0.8283; ROUGE-1 0.7253; ROUGE-2 0.6763; ROUGE-L 0.6911). The study highlights the value of task-specific adaptation and interpretability in high-stakes financial contexts, while acknowledging computational constraints and the need for further robustness and domain adaptation research.

Abstract

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.

Paper Structure

This paper contains 7 sections, 9 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Dataset & Task
  3. Methodology
  4. Results
  5. Conclusion
  6. Appendix
  7. Confusion Matrix

Figures (9)

  • Figure 1: System design workflow. The development set is initially used to select the best-performing model, which is then fine-tuned on the train set using the sequential learning approach. The final model is then used for inference on the test set.
  • Figure 2: Distribution of financial claims across different sectors. Adapted from rangapur2024finfactbenchmarkdatasetmultimodal.
  • Figure 3: Comparison of prompts used for classification and classification & explanation generation.
  • Figure 4: Llama3 8B's Confusion Matrix for classification on the dev set
  • Figure 5: Mistral 7B's Confusion Matrix for classification on the dev set
  • ...and 4 more figures