Structured Information Extraction and Applications in Complex Reasoning

Abstract

Structured information is generally easier to store, understand, and utilize compared to unstructured text. Extracting structured information from unstructured text and applying this extracted information to various tasks and use cases has been a longstanding research area. In this dissertation, we address both the extraction of structured information and its application to complex reasoning tasks. For structured information extraction, we focus on temporal information extraction, particularly on temporal normalization. Through systematic comparative experiments, we propose a strategy that enhances the generalization capability of time expression recognition systems—the crucial first step in the temporal normalization task—in new domains. Additionally, we introduce a semantic parsing framework based on large language models for end-to-end temporal expression recognition. Regarding the application of structured information, we focus on two tasks: temporal reasoning and open-domain multi-hop reasoning. In temporal reasoning, we combine extracted temporal graphs with Transformer-based question-answering systems, significantly improving their temporal reasoning capabilities. For the open-domain multi-hop reasoning task, we propose a semi-structured chain-of-thought approach that effectively integrates structured knowledge graphs, unstructured text, and parametric knowledge within large language models to answer knowledge-intensive questions.