Short Text Similarity Detection Based on Graph Attention Networks and Knowledge Augmentation

Text similarity detection is a crucial task within the field of natural language processing Traditional word embedding methods consider about either semantics or structure of the text, and rarely consider knowledge-based information. Therefore, these methods have insufficient information for text representation and are ineffective in dealing with word polysemy and deep interactions among words. In this paper, firstly HowNet, as a knowledge base, is introduced in the graph attention networks to encode the sememes in HowNet. Secondly, we design the self-attention mechanism to strengthen the correlation among words obtained from the graph attention networks. Finally, the enhanced word vectors are fused with the word vectors generated by the pre-trained BERT model to obtain a more comprehensive text representation. The experiments conducted on three Chinese datasets demonstrate the superior performance of our model outperforms other conventional text similarity detection models.