Semantic Similarity Evaluation Method Based on Text Generation Data Augmentation

The similarity evaluation method based on neural network has achieved good results, but it has higher requirements on the scale and quality of the corpus. Based on this problem, this paper proposes a semantic similarity evaluation method based on text generation data augmentation. This method combines Seq2Seq with a masked language model for data augmentation, and uses the expanded data to fine-tune the pre-trained language model. The pre-trained language model and the Siamese network are combined to build a semantic similarity evaluation model. Finally, experiments were carried out on the standard sentence similarity evaluation data set SentEva12012-2016. Compared with the benchmark model, the Spearman correlation coefficient improved by 3.11%. Experiments show that the semantic similarity evaluation method based on data augmentation can effectively solve the problem of low accuracy due to lack of data.