Joint Extraction of Entities and Relations Based on Multi-feature Fusion

Abstract

Joint extraction of entities and relations is essential for understanding massive text corpora. In recent years, the span-based joint models have achieved excellent results in the entity and relation extraction task. However, the previous literature and experimental results suggest that the usage of span-based method in entity and relation extraction may produce more redundant entities, although it can solve the overlapping problem of entities. In order to solve the problem of entity redundancy, this paper proposes a joint extraction model based on multi-feature fusion. The overall network follows the framework as SpERT, which is the state-of-the-art model for joint entity and relation extraction. In addition to the word embedding features in SpERT, the proposed model also considers the part-of-speech features. We believe that the part-of-speech features in entities are helpful for entity recognition and can effectively alleviate the entity redundancy problem. The proposed model is evaluated on two public data sets, CoNLL04 and ADE. The experimental results show that the proposed joint extraction model based on multi-feature fusion significantly outperforms current state-of-the-art methods.