Augment before You Try: Knowledge-Enhanced Table Question Answering via Table Expansion

TL;DR

The paper tackles table question answering by introducing a simple augment-then-generate pipeline that augments the given table with external knowledge and then derives an answer via a standard SQL query over the original and augmented tables. This approach preserves table structure, avoids embedding LLM queries in SQL, and yields strong performance across open- and closed-domain benchmarks, notably achieving state-of-the-art or competitive exact-match results on WikiTQ and solid results on TATQA/FinQA. Empirical results demonstrate advantages in handling large tables and complex tabular operations, while analyses highlight reduced execution errors compared to baselines that couple LLMs with SQL or rely solely on linearized text. The work suggests a practical and scalable path for knowledge-enhanced table QA with clear interpretability and easier error analysis through conventional SQL querying over augmented data.

Abstract

Table question answering is a popular task that assesses a model's ability to understand and interact with structured data. However, the given table often does not contain sufficient information for answering the question, necessitating the integration of external knowledge. Existing methods either convert both the table and external knowledge into text, which neglects the structured nature of the table; or they embed queries for external sources in the interaction with the table, which complicates the process. In this paper, we propose a simple yet effective method to integrate external information in a given table. Our method first constructs an augmenting table containing the missing information and then generates a SQL query over the two tables to answer the question. Experiments show that our method outperforms strong baselines on three table QA benchmarks. Our code is publicly available at https://github.com/UCSB-NLP-Chang/Augment_tableQA.

Figures (20)

• Figure 1: Comparison between Program-of-Thought, Binder, and our method.
• Figure 2: Performance grouped by table length.
• Figure 3: Comparison between our method and Binder.
• Figure 4: Performance decomposition by the number of table cells needed to answer the question.
• Figure 5: An example question in WikiTQ. Binder generates a SQL statement that queries LLMs for unsolvable parts. However, the statement leads to an execution error. Our method augments the table with an additional column and correctly generates a SQL statement to answer the question.