Semantic similarity detection based on knowledge augmentation for short text

The task of semantic similarity detection is crucial to natural language processing. Deep learning-based word embedding with text-matching models has improved the accuracy of spotting semantic similarities in recent years. However, traditional word embedding techniques cannot deal with polysemy sufficiently. In order to improve the quality of the semantic similarity detection for short text, this paper proposed a method based on knowledge augmentation. Here HowNet, as the tool for knowledge augmentation, is introduced integrating pre-trained BERT with fine-tuning and attention mechanisms. Experiments on two Chinese datasets show that the proposed method outperforms a variety of typical text similarity detection methods.