Multi-feature Fusion for Relation Extraction using Entity Types and Word Dependencies

Most existing methods do not make full use of different types of information sources to extract effective features for relation extraction. This paper proposes a multi-feature fusion model based on raw input sentences and external knowledge sources, which deeply integrates diverse lexical, semantic, and syntactic features into deep neural network models. Specifically, our model extracts lexical features of different granularity from the original input text representation, entity type features from the entity annotation information of the corpus, and dependency features from the dependency trees. Meanwhile, the dimension-based attention mechanism is proposed to enrich the diversity of entity type features and enhance their discriminability. Different features enable the model to comprehensively utilize various types of information, so this paper fuses these features and train a classifier for relation extraction. The experimental results show that the proposed model outperforms the existing state-of-the-art baselines on the TACRED Revisited, Re-TACRED, and SemEval datasets, with macro-average F1 scores of 81.2%, 90.2%, and 89.4%, respectively, improving the performance by 1.4%, 4.4%, and 2% on average, which indicates the effectiveness of multi-feature fusion modeling.

@article{Zhang2023,
  title     = {Multi-feature Fusion for Relation Extraction using Entity Types and Word Dependencies},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {14},
  number    = {7},
  year      = {2023},
  publisher = {The Science and Information Organization},
  author    = {Pu Zhang and Junwei Li and Sixing Chen and Jingyu Zhang and Libo Tang},
  doi       = {10.14569/IJACSA.2023.0140731},
  url       = {https://doi.org/10.14569/IJACSA.2023.0140731}
}