Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding

Table-based reasoning with large language models (LLMs) is a promising direction to tackle many table understanding tasks, such as table-based question answering and fact verification. Compared with generic reasoning, table-based reasoning requires the extraction of underlying semantics from both free-form questions and semi-structured tabular data. Chain-of-Thought and its similar approaches incorporate the reasoning chain in the form of textual context, but it is still an open question how to effectively leverage tabular data in the reasoning chain. Specifically, we guide LLMs using in-context learning to iteratively generate operations and update the table to represent a tabular reasoning chain. LLMs can therefore dynamically plan the next operation based on the results of the previous ones. This continuous evolution of the table forms a chain, showing the reasoning process for a given tabular problem. The chain carries structured information of the intermediate results, enabling more accurate and reliable predictions. Tables are a popular data format and widely used in daily life (Cafarella et al., 2008). Understanding tabular data with language models can benefit various downstream tasks, such as table-based fact verification (Chen et al., 2019), and table-based question answering (Jin et al., 2022). Distinct from pure text, tables deliver rich information through the interaction between rows and columns in the tabular structure, which enhances the data capacity but also increases the difficulty for language models to understand them. Thus, reasoning over the tabular data is an important direction in natural language processing and attracts increasing attention from both academia and industry. In recent years, several approaches have been suggested to tackle the problem of table understanding by training language models. One common direction is to add specialized embedding layers or attention mechanisms into language models and pre-train the models by recovering table cells or segments (Herzig et al., 2020; Wang et al., 2021; Gu et al., 2022; Andrejczuk et al., 2022).