Representations for Question Answering from Documents with Tables and Text

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 2895–2906 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 2895 Representations for Question Answering from Documents with Tables and Text Vicky Zayats†∗ Kristina Toutanova† Mari Ostendorf§ †Google Research §University of Washington {vzayats, kristout}@google.com ostendor@uw.edu Abstract Tables in Web documents are pervasive and can be directly used to answer many of the queries searched on the Web, motivating their integration in question answering. Very often information presented in tables is succinct and hard to interpret with standard language rep- resentations. On the other hand, tables often appear within textual context, such as an arti- cle describing the table. Using the information from an article as additional context can po- tentially enrich table representations. In this work we aim to improve question answering from tables by refining table representations based on information from surrounding text. We also present an effective method to com- bine text and table-based predictions for ques- tion answering from full documents, obtaining significant improvements on the Natural Ques- tions dataset. 1 Introduction Tables are a common type of information repre- sentation used across the Internet. With billions of search queries a day,1 question answering on tables is an important task that translates into a large number of search queries every second about information present in tables. In general, research on Question Answering (QA) can be categorized in terms of the resources that are used in answer- ing the question: text documents (often referred as unstructured text in the literature), tables, or a structured knowledge base (KB). In our work we are interested in the combination of text-based and structured resources for question answering, par- ticularly articles that contain both tables and text. This is a natural next step for question answering on tables, in that most tables are embedded in doc- uments that discuss them, creating the challenge of ∗*Most of the work was done while the author was at the University of Washington. 1https://www.internetlivestats.com determining whether the answer is in the text or the table (if anywhere). In addition, very often informa- tion presented in tables is compact and abbreviated. The associated text can potentially provide rich con- text that can be used to enhance the representation of the table for more robust question answering. The main focus of this paper is to investigate how to improve question answering on documents that contain both text and tables. While recently there has been a lot of interest in reading compre- hension for both text and tables, little research has been done in combining the two sources of infor- mation. The only prior study we are aware of is by Chen et al. (2020) who introduced a new dataset for multi-hop QA over tabular and textual data. In their work, the authors heavily rely on the assump- tion that the questions would be unanswerable if either text or table information is missing. Here we investigate a more realistic scenario of natu- …

Content truncated to 500 words