Reasoning Factual Knowledge in Structured Data with Large Language Models

TL;DR

StructFact introduces a large, structure-aware benchmark (8,340 questions across five factual tasks) to evaluate how well large language models reason with factual knowledge stored in structured data. By analyzing zero-shot, few-shot, and chain-of-thought prompting across ten diverse LLMs, the study highlights persistent weaknesses in arithmetic and spatiotemporal reasoning, limited gains from prompt engineering for instruction-tuned models, and notable resilience to evidence but greater vulnerability when evidence is absent. Fine-grained task analyses reveal that heterogeneity, sparsity, and topological constraints in tables and lists impede reasoning, while unstructured context can help when integrated carefully. The findings motivate future research on structure-aware modules and retrieval-augmented strategies to improve reliable reasoning over structured data in real-world, knowledge-sensitive tasks.

Abstract

Large language models (LLMs) have made remarkable progress in various natural language processing tasks as a benefit of their capability to comprehend and reason with factual knowledge. However, a significant amount of factual knowledge is stored in structured data, which possesses unique characteristics that differ from the unstructured texts used for pretraining. This difference can introduce imperceptible inference parameter deviations, posing challenges for LLMs in effectively utilizing and reasoning with structured data to accurately infer factual knowledge. To this end, we propose a benchmark named StructFact, to evaluate the structural reasoning capabilities of LLMs in inferring factual knowledge. StructFact comprises 8,340 factual questions encompassing various tasks, domains, timelines, and regions. This benchmark allows us to investigate the capability of LLMs across five factual tasks derived from the unique characteristics of structural facts. Extensive experiments on a set of LLMs with different training strategies reveal the limitations of current LLMs in inferring factual knowledge from structured data. We present this benchmark as a compass to navigate the strengths and weaknesses of LLMs in reasoning with structured data for knowledge-sensitive tasks, and to encourage advancements in related real-world applications. Please find our code at https://github.com/EganGu/StructFact.

Figures (20)

• Figure 1: StructFact aims at evaluating the reasoning ability of LLMs over structured factual knowledge across five tasks.
• Figure 2: Prompts used in different settings (the main differences between each with zero-shot w/o CoT are marked in orange).
• Figure 3: The distribution of three types of responses across five tasks, averaged across 10 LLMs.
• Figure 4: Performance of GPT-4o-mini under different settings of structured evidence.
• Figure 5: Confusion matrices of GPT-4o-mini's performance under the settings w/ and w/o structured data as evidence.