Abstract:
Document-Level Relation Extraction (Document-Level RE) was designed to automatically recognise entities and their relationships within text, representing them as structur...Show MoreMetadata
Abstract:
Document-Level Relation Extraction (Document-Level RE) was designed to automatically recognise entities and their relationships within text, representing them as structured data. It requires aggregating information from multiple sentences throughout the entire document for inference. Existing methods primarily rely on syntactic trees built upon the relationships between words in unstructured text to infer relationships between entities. Previous work may have focused more on local, shallow reasoning or external dependencies, neglecting deep, multidimensional complex reasoning. In response, we propose a double staircase interactive model. Initially, we construct a dependency matrix modeling various mention dependencies to assist entity reasoning in acquiring rich contextual information. Building upon this foundation, we integrate a structured attention network to obtain non-local interactions for entities, executing contextual reasoning and structural inference interactively, thereby learning deeper structural-aware document representations. Our model outperforms the SGR [28] model on the public dataset DocRED, improving its performance by 0.55. Furthermore, through discussion analysis and auxiliary experiments, we validate the validity of the model in cross-sentence reasoning.
Published in: 2024 36th Chinese Control and Decision Conference (CCDC)
Date of Conference: 25-27 May 2024
Date Added to IEEE Xplore: 17 July 2024
ISBN Information: