Construction of a Dataset for Extracting the Relationship between Text and Tables for Securities Reports

Abstract

While securities reports contain valuable information to investors, the non-financial information contained therein is not well-structured, and it is not easy for investors to access desired information. Some necessary information cannot only be interpreted from the text but from tables as well, but since the format of such tables is not standardized, it requires considerable effort to find the rationale from the tables. In this paper, we propose a TTRE (Text-to-Table Relationship Extraction) task regarding the numerical values contained in securities reports, assuming that by linking the text with the relevant cells in the tables, we can assist investors in finding such information. We propose the task, build a dataset, which is then published, and host a shared task using it. By analyzing the evaluation results of the baseline methods we implemented, we discussed the difficulty level of the TTRE task and the future challenges to improve its performance. The results showed that the TTRE task requires consideration of not only the text of sentences and cells but also the structure of tables, making the task very challenging. It was also found that regarding the dataset construction, annotators’ judgements contained fluctuations, suggesting that the annotation method and the dataset size need to be improved.