Document-Level Mathematical Relation Extraction Using Pre-Training Models

Mathematical relation extraction is pivotal for identifying relationships between entities in texts, a crucial step in aiding automated systems to comprehend the semantics of mathematical content. While prior studies have concentrated on sentence-level relation extraction (RE), their applicability in real-world scenarios remains limited. Document-level RE (DocRE), in contrast, poses a more complex challenge, necessitating multi-sentence reasoning and the prediction of relationships across entire documents. The scarcity of high- quality datasets, attributed to the constraints of manual data annotation, further complicates this task. In our study, we focus on extracting relationships from mathematical documents through document-level RE techniques. To effectively tackle the complexities of multi-label and multi-entity scenarios in document-level RE, we introduce two innovative methods: adaptive thresholding and evidence context pooling. These approaches allow document-level RE systems to focus on relevant text, significantly enhancing relation extraction capabilities. Furthermore, we implement a self-training strategy to learn evidence retrieval knowledge from a large amount of distantly-supervised data that lacks specific evidence annotations. Our experimental findings affirm the success of these methods, showcasing an impressive accuracy rate reaching up to 90.14% in the domain of mathematical relation extraction.