ALLPCB
Overview
This article surveys inference tasks on tabular data using large language models (LLMs). Tabular inference tasks require a model to produce answers or decisions based on one or more tables, for example table-based question answering and fact verification.
1. Introduction
Inference on tabular data plays an important role in computer science, especially in natural language processing (NLP). The task asks a model to generate required outputs given structured table data and an associated user query or instruction.
Past research on table inference progressed through rule-based methods, neural network approaches, and pretraining-based models. Recent work shows that large language models (LLMs) achieve notable performance across many NLP tasks. In particular, their in-context learning capability allows adaptation to new tasks without extensive task-specific fine-tuning, reducing the need for labeled data.
Consequently, building systems that can accurately and efficiently infer required information from complex structured data has become increasingly important, underscoring the relevance of tabular inference tasks. For brevity, this article refers to "inference tasks on tabular data" as "table inference tasks".
2. Background
2.1 Task definition
In table inference tasks, the model input consists of a user query or requirement, structured table data, and optionally accompanying textual table descriptions. The output is an answer or decision that satisfies the user query. Typical tasks include table-based question answering, fact verification, table-to-text generation, and text-to-SQL generation. A summary of related table inference tasks is commonly used to categorize benchmarks and approaches.
2.2 Relevant datasets
Although many recent works apply LLMs to table inference using in-context learning without additional training, labeled datasets remain necessary to evaluate LLM performance on these tasks. Such datasets provide standardized benchmarks for table QA, verification, and semantic parsing to SQL, enabling comparison across methods.
3. Recent progress
To help researchers identify avenues for improving LLM capabilities on tables, existing LLM-based table inference work can be grouped into five categories: supervised fine-tuning, modular decomposition, in-context learning, tool usage, and robustness improvement.
4. Future directions
This section outlines ways to improve each class of methods to further enhance LLM performance on table inference tasks, aiming to inspire future research directions.
5. Conclusion
This article provides a concise summary and outlook on research concerning LLMs for table inference, intended to help researchers navigate current methods and open problems.
