Work

Decoding the Textual Tapestry: Unraveling Structured Knowledge through Relation Extraction and Discovery

Year: 2025
Type: dissertation
Abstract: Relation Extraction (RE) is a critical task in Natural Language Processing (NLP), enabling the transformation of unstructured text into structured knowledge. While recent advancements in deep learning and large pre-trained language models have improved RE performance, challenges persist in handling noisy supervision, discovering novel relation types, and adapting RE to domain-specific tasks. This dissertation addresses these challenges through four key contributions, improving RE across multiple dimensions, including pre-training, open-world discovery, domain adaptation, and inference-based modeling.First, we introduce Fine-grained Contrastive Learning (FineCL), a novel pre-training method that refines contrastive learning with noise-aware weighting. By leveraging learning order denoising to assess the reliability of distantly supervised training instances, FineCL prioritizes high-quality labels, leading to more accurate relation representations and improved downstream performance in relation extraction.Second, we advance Open-world Relation Extraction (OpenRE) by grounding it in a real-world setting that includes both known and novel relation classes, as well as negative instances where no relation exists. We develop a method called Known and Novel Relation Discovery (KNoRD), which leverages intrinsic qualities of explicit and implicit relations, significantly improving the discovery of long-tail relation classes.Third, we introduce MiDRED, a manually annotated corpus for microbe-disease relation extraction, containing 3,116 high-quality annotations. MiDRED serves as a valuable resource for training and evaluating RE models in the biomedical domain, addressing the scarcity of well-labeled data in this field.Finally, we explore the adaptation of RE to Natural Language Inference (NLI) with MetaEntail-RE, a framework that reframes RE as an entailment task. By incorporating meta-class analysis, feasible hypothesis filtering, and group-based prediction selection, MetaEntail-RE enhances performance, achieving significant F1-score improvements on both biomedical and general-domain RE benchmarks.Together, these contributions advance RE by improving robustness to noisy supervision, enhancing open-world discovery, introducing a key domain-specific dataset, and integrating NLI techniques for better extraction performance. These innovations not only enhance RE methodologies but also contribute to broader NLP applications, including knowledge graph construction, automated fact verification, and biomedical text mining.
Author: Hogan, William Patrick
Language: English
Open Access status: green
OpenAlex ID: W7008932483
Cites: 92
Cited by:
Related to: 10