An Inner Table Retriever for Robust Table Question Answering

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9909–9926 July 9-14, 2023 ©2023 Association for Computational Linguistics An Inner Table Retriever for Robust Table Question Answering Weizhe Lin∗2, Rexhina Blloshmi1, Bill Byrne1,2, Adrià de Gispert1, and Gonzalo Iglesias1 1Amazon Alexa AI 2University of Cambridge wl356@cam.ac.uk {blloshmi, willbyrn, agispert, gjii}@amazon.com Abstract Recent years have witnessed the thriving of pre- trained Transformer-based language models for understanding semi-structured tables, with sev- eral applications, such as Table Question An- swering (TableQA). These models are typically trained on joint tables and surrounding natural language text, by linearizing table content into sequences comprising special tokens and cell information. This yields very long sequences which increase system inefficiency, and more- over, simply truncating long sequences results in information loss for downstream tasks. We propose Inner Table Retriever (ITR),1 a general- purpose approach for handling long tables in TableQA that extracts sub-tables to preserve the most relevant information for a question. We show that ITR can be easily integrated into existing systems to improve their accuracy with up to 1.3-4.8% and achieve state-of-the-art re- sults in two benchmarks, i.e., 63.4% in Wik- iTableQuestions and 92.1% in WikiSQL. Ad- ditionally, we show that ITR makes TableQA systems more robust to reduced model capacity and to different ordering of columns and rows. 1 Introduction Tables offer a systematic way of storing informa- tion in the Web. Extracting information from Web tables poses different challenges than extracting information from relational databases with logi- cal queries, especially when queried via Natural Language (NL) user questions. Table Question Answering (TableQA) is the task of answering such questions with factoid answers extracted from table content. This requires developing models with the ability to reason over and understand ta- bles. Following the success of the pre-training paradigm for understanding NL text (Devlin et al., 2019), some recent research has focused on pre- training Transformer models (Vaswani et al., 2017) ∗Work done as an intern at Amazon Alexa AI. 1We make our code available at: https://github.com/ amazon-science/robust-tableqa Rank Mountain Peak Mountain Range Elevation 0 Mount Whitney Sierra Nevada 14,505 ft 1 Mount Williamson Sierra Nevada 14,379 ft 2 Red Slate Mountain Sierra Nevada 13,162 ft 3 Mount Ritter Sierra Nevada 13,149 ft Question: which mountain peak is no higher than 13 , 149 ft ? Table (a): [HEAD] rank | mountain peak | mountain range | elevation [ROW] 1 : 0 | mount whit ney | s ierra ne v ada | 14 , 505 ft [ROW] 2 : 1 | mount will iam Sub-table (b): [HEAD] mountain peak | elevation [ROW] 1 : red slate mountain | 13 , 162 ft [ROW] 2 : mount r itter | 13 , 149 ft c1 c2 c3 c4 r1 r2 r3 r4 Mountain Peak Elevation Red Slate Mountain 13,162 ft Mount Ritter 13,149 ft c2 c4 r3 r4 (a) (b) Answer: mount whitney Answer: mount ritter Figure 1: TableQA example with the model input length budget set to …

Content truncated to 500 words