CATS: A Pragmatic Chinese Answer-to-Sequence Dataset with Large Scale and High Quality

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2983–3000 July 9-14, 2023 ©2023 Association for Computational Linguistics CATS: A Pragmatic Chinese Answer-to-Sequence Dataset with Large Scale and High Quality Liang Li1,2, Ruiying Geng3, Chengyang Fang1,2, Bing Li1, Can Ma1∗, Rongyu Cao3, Binhua Li3 , Fei Huang3, Yongbin Li3∗ 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China 3DAMO Academy, Alibaba Group {liliang, macan}@iie.ac.cn {ruiying.gry, shuide.lyb}@alibaba-inc.com Abstract There are three problems existing in the popu- lar data-to-text datasets. First, the large-scale datasets either contain noise or lack real appli- cation scenarios. Second, the datasets close to real applications are relatively small in size. Last, current datasets bias in the English lan- guage while leaving other languages underex- plored. To alleviate these limitations, in this paper, we present CATS, a pragmatic Chinese answer-to-sequence dataset with large scale and high quality. The dataset aims to generate tex- tual descriptions for the answer in the practi- cal TableQA system. Further, to bridge the structural gap between the input SQL and table and establish better semantic alignments, we propose a Unified Graph Transformation ap- proach to establish a joint encoding space for the two hybrid knowledge resources and con- vert this task to a graph-to-text problem. The experiment results demonstrate the effective- ness of our proposed method. Further analysis on CATS1 attests to both the high quality and challenges of the dataset. 1 Introduction Data-to-text (D2T) generation (Kukich, 1983; Re- iter and Dale, 1997) aims to generate a natural language description conditioned on structured or semi-structured data, such as graphs (Song et al., 2018; Wang et al., 2020c) or tables (Lebret et al., 2016; Wiseman et al., 2017). It helps people get the key points of the input data and makes the stored information accessible to a broader range of end- users. A large number of datasets have been pro- posed as the testbed for neural D2T models and are driving the domain. However, as shown in Table 1, we note three problems existing in the popular datasets. First, the large-scale datasets either contain noises (e.g., ∗Corresponding authors: Can Ma, Yongbin Li 1CATS is available at https://github.com/ AlibabaResearch/DAMO-ConvAI/tree/main/cats Robot … 08:24 SQL Query Execution Result Text-to-SQL Execute Database ������������������ ��������������� What are the names and fields of the three Nobel Prizes that were first awarded and the maximum age of their winners? SELECT �� , �� , ������� FROM ����� ORDER BY ������� ASC LIMIT 3 SELECT Name, Field, The maximum age of the winner FROM Nobel Prize ORDER BY The first award time ASC LIMIT 3 The names and fields of the first three Nobel Prizes awarded at the earliest time and the maximum age of their winners are: 1. Nobel Prize in Chemistry, the field of chemistry, the maximum age of the winner is 77; 2. Nobel Prize in Physics, the field of physics, the maximum age of the winner is …

Content truncated to 500 words