OWRM: Extracting All Entities and Relations with Only One Word-Word Relation Matrix

Extracting named entities and their relations is a crucial task in natural language processing (NLP), and various cleverly designed algorithms have been proposed to tackle it. However, these algorithms still have at least two deficiencies: (1) a lack of clear and unified task objectives, resulting in incomplete extraction and challenges in uniformly evaluating algorithm performance; (2) the inability to extract all entities and relations in a sentence through a single prediction, leading to a significant reduction in algorithm efficiency. To address the above two issues, we define the objectives of the joint extraction task and propose a novel algorithm, OWRM, featuring five carefully designed tags. Our algorithm aims to extract every word of entities, classify them, and predict their relations by utilizing only one word-word relation matrix. To achieve these goals, we develop a neural network architecture with four components: sentence encoding, word-word representation generation, word-word relation prediction, and entity and relation decoding. Extensive experimental results on multiple widely used datasets demonstrate that our algorithm exhibits superior or comparable accuracy compared to state-of-the-art baseline algorithms while significantly reducing training time by 1.8–2.7 times and prediction time by \(9.4\) – \(14.1\) times.