A Compare-and-contrast Multistage Pipeline for Uncovering Financial Signals in Financial Reports

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 14307–14321 July 9-14, 2023 ©2023 Association for Computational Linguistics A Compare-and-contrast Multistage Pipeline for Uncovering Financial Signals in Financial Reports Jia-Huei Ju1, Yu-Shiang Huang1,2, Cheng-Wei Lin1, Che Lin2,3,4, and Chuan-Ju Wang1,2 1Research Center for Information Technology Innovation, Academia Sinica, 2Graduate Program of Data Science, National Taiwan University and Academia Sinica, 3Graduate Institute of Communication Engineering, National Taiwan University, 4Department of Electrical Engineering, National Taiwan University, {jhjoo, yushuang, lcw.1997}@citi.sinica.edu.tw, chelin@ntu.edu.tw, cjwang@citi.sinica.edu.tw Abstract In this paper, we address the challenge of dis- covering financial signals in narrative finan- cial reports. As these documents are often lengthy and tend to blend routine informa- tion with new information, it is challenging for professionals to discern critical financial signals. To this end, we leverage the inher- ent nature of the year-to-year structure of re- ports to define a novel signal-highlighting task; more importantly, we propose a compare-and- contrast multistage pipeline that recognizes dif- ferent relationships between the reports and lo- cates relevant rationales for these relationships. We also create and publicly release a human- annotated dataset for our task. Our experiments on the dataset validate the effectiveness of our pipeline, and we provide detailed analyses and ablation studies to support our findings. 1 Introduction With the rapid growth of information, many tasks in the field of natural language processing (NLP) involve streamlining information comprehension. One such task is summarization, which selects a subset of sentences or generates new content that best represents the given document (Hermann et al., 2015; See et al., 2017; Cohan et al., 2018). This task helps humans save time and effort by iden- tifying important information in a text. In the fi- nance context, comprehending regulatory narrative reports is a classic example of efficiently mining signals from a large amount of text. As these re- ports often contain rich information concerning spe- cific financial entities, discovering valuable insights is crucial for academia and the finance industry. Much research has shown that textual features from financial reports contain valuable financial sig- nals about future firm performance and market reac- tions (e.g., Badertscher et al., 2018; Ertugrul et al., 2017; You and Zhang, 2009). However, authorities such as the Securities and Exchange Commission (SEC) require that companies provide comprehen- sive and detailed information about their current status in these reports, which often contain much unimportant and already-known information. For example, the token overlap ratio between annual 10-K reports of the same company between adja- cent years is often high,1 making it a challenging and tedious task to acquire important signals in new reports (termed as the overlapping characteristic hereafter). Recent advances in NLP technology have in- cluded attempts to efficiently and effectively com- prehend lengthy financial documents. One ap- proach to address this problem is through summa- rization (e.g., Zmandar et al., 2021b; Orzhenovskii, 2021; Gokhan et al., 2021). Other approaches addi- tionally leverage numerical metrics, such as stock return volatility and abnormal trading volumes, to …

Content truncated to 500 words