MSIF: Multi-source Information Fusion for Financial Question Answering

Man Lin, De-Long Zeng, Jiarui Ouyang, Ying Shen

Abstract: Research on hybrid data, which combines tabular and textual content, has garnered significant interest in financial question answering. Recent approaches mainly focus on the encoding of tables and texts to facilitate model generation. However, there is still room for improvement in the question reasoning type and question semantic topic. Therefore, we propose MSIF, a novel model based on Multi-Source Information Fusion which integrates additional information, encompassing question reasoning type and question semantic topic, from multiple sources for the original question. Specifically, we first identify the question reasoning type that helps discrete reasoning from the existing corpus TAT-QA. We then obtain the question semantic topic that helps analyze the question by leveraging the large-scale generative language model ChatGPT. Finally, we introduce an information fusion module to integrate additional information into the context representation. Experiments on the FinQA dataset show the effectiveness of our model in financial question answering, which outperforms most baselines on Execution Accuracy and Program Accuracy.