表格问答研究综述

Summary

This survey from Chongqing Jiaotong University reviews table question answering (TableQA), where users query tables in natural language rather than writing SQL. The authors define three task settings by the number of tables and dialogue turns: single-table single-turn, multi-table single-turn, and multi-table multi-turn. For each setting, they describe public datasets, evaluation metrics and representative methods, and they group the methods into families.

• Single-table methods: rule-based, sketch/slot-filling (chunking), and type-information approaches.
• Multi-table methods: SQL-grammar, graph neural network, and intermediate-representation approaches.
• Conversational (multi-turn) methods: context-based and data-augmentation approaches.

A separate part compares table pre-training models by how they build training data, encode inputs and define pre-training objectives.

The review traces the field's path. It began with early natural-language database interfaces, moved to Text-to-SQL semantic parsing after WikiSQL appeared in 2017, and has seen large gains from pre-trained models since 2021. Single-table results now exceed human performance, while multi-table and dialogue settings remain hard. The survey is a useful map for anyone building SQL-free access to relational or document tables, including Chinese-language resources.

Key points

• Single-table single-turn datasets:
  • WikiSQL: 24,241 tables, 80,654 questions, 77,840 SQL queries.
  • WikiTableQuestions (WikiTQ): 22,203 questions over 2,108 tables, answers only, no logical forms.
  • ATIS: a 1990 air-travel set with 27 tables.
  • Domain-specific sets for finance: FinQA and TAT-QA.
• Chinese TableQA dataset: about 64,891 questions, 20,311 SQL queries and more than 6,000 tables, drawn from financial reports and public websites. Roughly 30% of its items require entity linking.
• Multi-table datasets:
  • Spider (Yale, 2018): 10,181 questions, 5,693 complex SQL queries, 200 databases, 138 domains.
  • CSpider: a Chinese translation of Spider.
  • DuSQL: 23,797 question–SQL pairs over 200 databases and 813 tables.
• Metrics: logical form accuracy and execution accuracy are used for single-table tasks; exact set match and execution accuracy for multi-table tasks. The authors note that execution accuracy can wrongly reward a semantically incorrect query, for example when both queries return NULL.
• Model comparison on WikiSQL: the survey covers SQLNet, Coarse2Fine, TypeSQL, TaBERT, GraPPa, SeaD, TAPEX and TaPas. SeaD performs best, with 87.5% logical form accuracy and 93.0% execution accuracy. Models covered on Spider include IRNet, RAT-SQL, GAP, GraPPa and SDCUP.