A Deep Learning Model Fused with Word Sense Knowledge for Textual Entailment Recognition
Yalei Liu, Lingling Mu, Hongying Zan · 2021
Textual entailment recognition is an essential research task in the field of natural language processing. The mainstream textual entailment recognition method based on deep learning does not integrate word sense knowledge in training data, so the inference knowledge of model learning is limited. In addition, polysemy has become a significant challenge for the fusion of semantic knowledge. Therefore, we propose a textual entailment recognition model fused with word sense knowledge. It enhances the explicit knowledge of the data by fusing the synonyms in the CiLin. We use word sense disambiguation to fuse the meaning of polysemy. We use the RoBERTa word vector and the word vector representation based on the sememe to initialize sentence encoding. Our model achieves an accuracy of 80.77% on the CNLI dataset and 80.74% on the XNLI dataset.