MAFiD: Moving Average Equipped Fusion-in-Decoder for Question Answering over Tabular and Textual Data

Findings of the Association for Computational Linguistics: EACL 2023, pages 2337–2344 May 2-6, 2023 ©2023 Association for Computational Linguistics MAFiD: Moving Average Equipped Fusion-in-Decoder for Question Answering over Tabular and Textual Data Sung-Min Lee†, Eunhwan Park†, Daeryong Seo‡, Donghyeon Jeon‡, Inho Kang‡, Seung-Hoon Na†∗ †Jeonbuk National University ‡NAVER Corporation {cap1232, judepark, nash}@jbnu.ac.kr, {daeryong.seo, donghyeon.jeon, once.ihkang}@navercorp.com Abstract Transformer-based models for question answer- ing (QA) over tables and texts confront a “long” hybrid sequence over tabular and textual ele- ments, causing long-range reasoning problems. To handle long-range reasoning, we extensively employ a fusion-in-decoder (FiD) and expo- nential moving average (EMA), proposing a Moving Average Equipped Fusion-in-Decoder (MAFiD). With FiD as the backbone architec- ture, MAFiD combines various levels of rea- soning: independent encoding of homogeneous data and single-row and multi-row heteroge- neous reasoning, using a gated cross attention layer to effectively aggregate the three types of representations resulting from various reason- ings. Experimental results on HybridQA indi- cate that MAFiD achieves state-of-the-art per- formance by increasing exact matching (EM) and F1 by 1.1 and 1.7, respectively, on the blind test set. 1 Introduction While most studies have focused on text question answering (QA), where unimodal textual passages are provided as a source of evidence for an answer (Joshi et al., 2017; Yang et al., 2018; Rajpurkar et al., 2018; Welbl et al., 2018; Dua et al., 2019; Karpukhin et al., 2020; Zhu et al., 2021b; Pang et al., 2022), realistic questions often need to re- fer to “heterogeneous” evidences based on both tabular and textual contents to generate an answer, motivating researchers to address table-and-text QA (Chen et al., 2020; Wenhu Chen, 2021; Talmor et al., 2021; Zhu et al., 2021a; Nakamura et al., 2022). Among the various tasks for table-and-text QA, we address the multi-hop table-and-text QA de- scribed in HybridQA (Chen et al., 2020), which is a large-scale table-and-text QA dataset focusing on the multi-hop reasoning across tabular and textual contents to extract an answer. *Corresponding author However, a table usually contains a nontrival number of rows and relevant passages; thus lin- earization of all relevant heterogeneous contents easily exceeds the maximum length limit for trans- formers, thereby causing long range reasoning problems. To address long range reasoning, we present a novel encoder-decoder model that deploys fusion- in-decoder (FiD) (Izacard and Grave, 2021) and ex- ponential moving average (EMA) (Ma et al., 2022), the Moving Average Equipped Fusion-in-Decoder (MAFiD). Armed with FiD as the backbone archi- tecture, MAFiD combines various levels of reason- ing: • Independent encoding of homogeneous data, which independently encodes tabular and textual contents separately for each row, without being fused in the encoder step. Inher- ited from FiD, the resulting encoded represen- tations are jointly fused in the decoder, which significantly reduces the computational time required for self-attention, thereby allowing us to use a longer sequence as an input for the encoder. • Single-row heterogeneous reasoning (also referred to as single-row reasoning), which performs in-depth interaction between tab- ular and textual contents per row; it first …

Content truncated to 500 words