Abstract: Document-level relation extraction entails sifting through extensive document data to pinpoint relationships and pertinent event details among various entities. This process aids intelligence analysts in swiftly grasping the essence of the content while revealing potential connections and emerging trends, thus proving invaluable for research purposes. This paper puts forward a method for document-level relation extraction that leverages an interaction attention mechanism. Initially, building on an evidence-based approach for extracting relations at the document level, the interaction attention mechanism is introduced, extracting the final layer of hidden states containing rich semantic information from the document encoder. Subsequently, these concealed states are fed into a self-attention layer informed by dependency parsing. The outputs from both elements serve as distinct supervisory signals for the interactive input. By pooling these output results, it can derive context embeddings that possess enhanced representational power. Preliminarily, relation triples are extracted using the relation classifier. In conclusion, building on the preliminary relationship results, the process of relationship inference is carried out independently using pseudo-documents created from the source material and pertinent evidence. Only those relationships with a cumulative inference score that surpasses a certain threshold are regarded as the final outcomes. Experimental findings from the publicly accessible datasets indicate commendable performance.

Zhang Mei, Zhao Zhongyuan and Xu Zhitong, “An Interactive Attention-Based Approach to Document-Level Relationship Extraction” International Journal of Advanced Computer Science and Applications(IJACSA), 15(10), 2024. http://dx.doi.org/10.14569/IJACSA.2024.0151032

@article{Mei2024,
title = {An Interactive Attention-Based Approach to Document-Level Relationship Extraction},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2024.0151032},
url = {http://dx.doi.org/10.14569/IJACSA.2024.0151032},
year = {2024},
publisher = {The Science and Information Organization},
volume = {15},
number = {10},
author = {Zhang Mei and Zhao Zhongyuan and Xu Zhitong}
}