Source-linked AI summary
A Survey on Table Question Answering: Recent Advances
Nengzheng Jin, Joanna Siebert, Dongfang Li, Qingcai Chen
TL;DR
Table QA has many studies but lacks a comprehensive survey, despite its role in answering questions from tables. This paper surveys datasets and methods, then outlines challenges and potential future directions; it notes that flattening tables for open-domain QA may lose important information.
Problem
Despite many studies, table QA lacks a comprehensive, systematic survey of its datasets, methods, and challenges.
Source: §1 · p. 2Method
The survey reviews available datasets, maps datasets to methods, summarizes five method categories, and discusses challenges and future directions.
Source: §1 · p. 2Results
The survey concludes that some methods perform desirably on relatively easier benchmarks, while key challenges remain, especially for non-database table QA.
Source: §6 · p. 10Takeaways & Limitations
The survey identifies table encoding and table reasoning as main challenges and discusses potential future directions.
Source: §1 · p. 2Open-domain methods that flatten tables may lose important information in semi-structured data, leaving table encoding open for further exploration.
Source: §5.4 · p. 10
Abstract
from arXiv · showhide
Table Question Answering (Table QA) refers to providing precise answers from tables to answer a user's question. In recent years, there have been a lot of works on table QA, but there is a lack of comprehensive surveys on this research topic. Hence, we aim to provide an overview of available datasets and representative methods in table QA. We classify existing methods for table QA into five categories according to their techniques, which include semantic-parsing-based, generative, extractive, matching-based, and retriever-reader-based methods. Moreover, as table QA is still a challenging task for existing methods, we also identify and outline several key challenges and discuss the potential future directions of table QA.
1 Introduction
Table QA seeks precise answers by understanding and reasoning over tables, but the field lacks a systematic survey. This survey reviews datasets and methods and identifies challenges and potential future directions.
1 Introduction
- It classifies table QA methods into five categories and introduces each category: semantic-parsing-based, generative, extractive, matching-based, and retriever-reader-based.The introduction places semantic parsing among methods that transform questions into structured queries, while other approaches may extract or generate answers or retrieve tables before reading them.Source: §1 · p. 2
- The survey addresses the lack of a systematic overview of table QA by covering datasets, methods, challenges, and future directions.It aims to help researchers grasp recent advances in the field.Source: §1 · p. 2
- The survey maps available datasets to existing methods to show which methods can be applied to each dataset.Source: §1 · p. 2
- The survey identifies table encoding and table reasoning as two main challenges and discusses potential directions for future research.Source: §1 · p. 2
2 Background
This background section introduces table composition, semantic parsing, and task distinctions used to frame table QA. It distinguishes tables by structure, domain, and answer form.
2.1 The Composition of Tables
- Tables are grids of cells in rows and columns, while web tables may also include titles, captions, and headers.Related sentences before or after a table are called pre-annotation and post-annotation and treated as supplement parts.Source: §2.1 · pp. 2–3
2.2 Semantic Parsing
- Semantic parsing transforms a natural-language utterance into an executable logical form, such as an SQL query.Source: §2.2 · p. 3
2.3 Table Question Answering
- Table QA uses database tables and non-database tables as reference sources, unlike KBQA's knowledge bases and text QA's unstructured text.Source: §2.3 · p. 3
- For web tables, open-domain QA searches large-scale table documents, whereas closed-domain QA uses a limited number of tables, usually one.Source: §2.3 · p. 3
- Free-form table QA generates dialog-like answers, while non-free-form QA returns factual answers usually consisting of a few words.Source: §2.3 · p. 3
3 Overview of Datasets
The survey reviews table QA datasets and relates them to representative methods. It distinguishes table-only datasets from datasets that pair tables with surrounding text.
3 Overview of Datasets
- Most surveyed datasets are closed-domain and use factoid questions.The survey extends question types to include free form and multiple choice alongside factoid, list, definition, and complex.Source: §3 · p. 3
- Table 1 aligns representative methods with dataset groups and marks methods currently adopted on particular datasets.Source: Table 1 · p. 4
- Table-only datasets provide either logical-form supervision, as in WikiSQL and Spider [47], or final-answer supervision, as in WTQ and SQA.Most contain regularly structured relational tables; HiTab and AITQA instead include hierarchical structure and merged cells.Source: §3 · p. 4
- Non-table-only datasets pair tables with pre- or post-annotations; open-domain examples include OTTQA and NQ-tables [17], while TAT-QA and FinQA come from financial reports.TabMCQ and GeoTSQA contain multiple-choice questions, and these datasets require modeling both tables and text.Source: §3 · p. 4
4 Semantic-parsing-based Methods
Semantic-parsing-based table QA converts questions into logical forms and executes them on tables. The survey covers weakly supervised learning strategies and fully supervised generation- and sketch-based approaches.
4 Semantic-parsing-based Methods
- Semantic-parsing-based methods transform questions into logical forms, execute them on tables, and may be trained with weak or full supervision.Weak supervision uses the final answer without gold logical forms; full supervision provides logical forms that execute to correct answers.Source: §4 · pp. 4–5
4.1 Weakly-supervised Table Semantic Parsing
- Weakly supervised methods range from hand-crafted grammar and deduction rules to neural parsers, with MML, reinforcement learning, and MMR used to address training challenges.Floating Parser uses recursive rules and type constraints; Neural Enquirer uses encoders and a neural executor, while Neural Programmer selects operations and columns. The survey also describes MML, policy-gradient approaches, and DynSP.Source: §4.1 · pp. 5–6
- Policy-gradient methods include pointer-based copying, policy shaping with prior knowledge, and MAPO [28], which stores high-reward logical forms to stabilize and accelerate training.Source: §4.1 · p. 6
4.2 Fully-Supervised Table Semantic Parsing
- Fully supervised semantic parsing for database tables is organized into generation-based and sketch-based methods.Generation approaches produce logical forms with Seq2Seq decoders, while sketch methods predict query modules and recompose SQL; schema linking connects question mentions to schema content.Source: §4.2 · pp. 6–7
5 Non-semantic-parsing-based Methods
Non-semantic-parsing approaches answer table questions through direct generation, extraction, fragment matching, or retrieval followed by reading. Their designs vary in how they represent table structure and support open-domain questions.
5.1 Generative Methods
- Generative methods output answers rather than logical forms; handles free-form QA, while ’s graph model represents tables as graphs and is table-only.Zayats et al. extend the graph-based approach to hybrid QA, but the paper presents that method in the extractive subsection.Source: §5.1 · pp. 7–8
5.2 Extractive Methods
- Extractive methods select spans as answers or evidence, with variants adding row/column features, structured attention masks, or attention biases.The multi-view attention approach reorders inputs and uses windowed attention to reduce quadratic complexity to linear; TableFormer uses 13 table-text attention biases.Source: §5.2 · p. 8
5.3 Matching-Based Methods
- Matching-based methods score questions against table fragments, such as rows, columns, or cells, then use the most relevant matches to retrieve answers.The RCI model combines a row-question score with a column-question score to estimate whether a cell is the answer.Source: §5.3 · p. 9
5.4 Retriever-reader-based Methods
- Open-domain systems retrieve table-bearing documents from a corpus, then use generative or extractive readers; retrievers may be sparse, dense, or iterative.Iterative retrieval appends reliable results from an earlier step to inform the next retrieval step.Source: §5.4 · p. 9
- Extractive-reader systems combine retrieval with cross-block reading, reasoning-chain extraction, or table-oriented pretraining.The fusion retriever pre-aligns table segments with related passages; another system appends flattened reasoning chains to retrieved tables for the reader.Source: §5.4 · pp. 9–10
- Generative-reader systems use FiD to generate answers or SQL from retrieved candidates, with BM25 or dense retrieval and, in UniK-QA, multiple knowledge sources [35].UniK-QA adapts to text, tables, and knowledge bases by flattening semi-structured data [35].Source: §5.4 · p. 10
- Flattening tables for open-domain QA may lose important semi-structured information, leaving table encoding as an area for further exploration [35].Source: §5.4 · p. 10
6 Challenges and Future Directions
The survey identifies numerical representation and complex reasoning in non-database table QA as challenges. It notes that existing methods often support only limited reasoning and calls for more capable representations and logical forms.
6 Challenges and Future Directions
- Some methods perform desirably on easier benchmarks, but the survey identifies further challenges, especially for non-database table QA.Source: §6 · p. 10
- About 55% of TAGOP errors in the RoBERTa-based model stem from incorrect evidence extraction, most often because of premature numerical representation.The survey identifies better numerical representations as a challenge, particularly for spreadsheet tables.Source: §6 · p. 10
- Non-database table QA methods often support only simple reasoning: TAGOP supports one-step operations, while FinQANet supports nested operations limited to four basic arithmetic operations.The survey calls for a more general logical form that supports complex reasoning across most non-database table QA tasks.Source: §6 · pp. 10–11