KBRTE: A Deep Learning Model for Chinese Textual Entailment Recognition Based on Synonym Expansion and Sememe Enhancement

Yalei Liu, Lingling Mu, Hongying Zan

Abstract: The mainstream textual entailment recognition models ignore the existing language knowledge, so the inference knowledge from the training data is limited, and the generalization ability is not strong. Therefore, this paper proposes a model KBRTE (fusing Knowledge Base in RTE) that combines an attention mechanism and a pre-trained model and uses word vectors based on sememe representation in the HowNet. We use the enhanced CNLI and XNLI datasets as the model’s training set. On the basis of these datasets, monosemous and polysemous in the CiLin are integrated to further enhance the knowledge. Experimental results show that this method could bring significant gains.