AN ABLATIVE STUDY OF THE RELATIONAL TRIPLE EXTRACTION MODEL RESC

Kuzmenko Andrey V., Kireev Vasiliy S.

Knowledge graphs are an important tool for improving recommendation systems, information retrieval systems, and question-and-answer systems. One of the ways to build them based on the corpus of natural language texts is to extract relational triples – entities and relationships between them. This article explores the RISC model previously proposed by the authors, based on a BERT-like encoder and allowing to solve the problem of extracting triples with high performance. The authors conducted a study of the ablation of the RESC model on a key set of NYT 11 texts. The optimal architecture parameters were determined and the most effective learning strategy was identified.

Keywords:relational triple, neural network, natural language processing, transformers, knowledge graphs