Answering Numerical Reasoning Questions in Table-Text Hybrid Contents with Graph-based Encoder and Tree-based Decoder

Proceedings of the 29th International Conference on Computational Linguistics, pages 1379–1390 October 12–17, 2022. 1379 Answering Numerical Reasoning Questions in Table-Text Hybrid Contents with Graph-based Encoder and Tree-based Decoder Fangyu Lei1,2, Shizhu He1,2, Xiang Li1,2, Jun Zhao1,2, Kang Liu1,2,3 1National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China 2 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 3 Beijing Academy of Artificial Intelligence, Beijing, 100084, China {leifangyu2022, lixiang2022}@ia.ac.cn {shizhu.he, jzhao, kliu}@nlpr.ia.ac.cn Abstract In the real-world question answering scenar- ios, hybrid form combining both tabular and textual contents has attracted more and more attention, among which numerical reasoning problem is one of the most typical and challeng- ing problems. Existing methods usually adopt encoder-decoder framework to represent hy- brid contents and generate answers. However, it can not capture the rich relationship among numerical value, table schema, and text infor- mation on the encoder side. The decoder uses a simple predefined operator classifier which is not flexible enough to handle numerical rea- soning processes with diverse expressions. To address these problems, this paper proposes a Relational Graph enhanced Hybrid table-text Numerical reasoning model with Tree decoder (RegHNT). It models the numerical question answering over table-text hybrid contents as an expression tree generation task. Moreover, we propose a novel relational graph model- ing method, which models alignment between questions, tables, and paragraphs. We vali- dated our model on the publicly available table- text hybrid QA benchmark (TAT-QA). The pro- posed RegHNT significantly outperform the baseline model and achieve state-of-the-art re- sults1 (2022-05-05). 1 Introduction Question Answering (QA) is an important task of natural language processing (NLP), which is often used to assess the intelligence of an agent. QA systems use various types of knowledge to answer natural language questions. Earlier ap- proaches independently utilized structured data such as tables (Pasupat and Liang, 2015; Yu et al., 2018), knowledge bases (Yih et al., 2016; Talmor and Berant, 2018) or unstructured data such as plain texts (Rajpurkar et al., 2016). In fact, real- world QA systems often need to fuse different data 1We openly released the source code and data at https: //github.com/lfy79001/RegHNT resources with diverse types in answering com- plex questions. Therefore, in recent years, the hy- brid form of question answering over tables and texts (TextTableQA) has attracted more and more attention (Chen et al., 2020a,b, 2021). There are two major question types for Text- TableQA. The first is the fact reasoning ques- tion, whose answer is usually a span from the ta- ble or linked paragraphs, such as the contents in Wikipedia (Chen et al., 2020a,b). The second is the numerical reasoning question, which usually aims to use the contents of tables and texts for nu- merical calculation (Zhu et al., 2021; Chen et al., 2021). Most previous work focuses on the first type, while the numerical reasoning questions have been seldom addressed. The existing datasets such as WikiTableQuestions (Pasupat and Liang, 2015) and DROP (Dua et al., 2019) also contain numeri- cal reasoning questions, but …

Content truncated to 500 words