Existing Chinese paraphrase generation models often lose proper nouns in the original sentence when generating paraphrased sentences,which results in semantic deviation and reduces the usability of the paraphrased sentence,decreasing performance on downstream tasks.To solve these problems,this study proposes proper noun-enhanced method for paraphrase generation.Specifically,a placeholder-based paraphrase generation model is proposed for an original sentence containing a single proper noun.The model retains the placeholder by replacing the proper noun in the training sentence pair with a placeholder,to train the moeel's ability to retain placeholder.A lexically constrained paraphrase generation model is proposed for an original sentence containing multiple proper nouns.By concatenating and distinguishing the list of proper nouns from the original sentence,the model is trained to recognize and reproduce multiple proper nouns,improving the appropriate noun retention rate in the paraphrased sentences.In addition,a reference-free metric is proposed to evaluate the quality of the generated paraphrased sentences by considering both semantic consistency and expression diversity.This study considers the intent recognition cold-start task in a real dialogue business system as the downstream task.By comparing the quality of the paraphrased sentences generated by different models and the accuracy of the intent recognition task,the experimental results show that the lexical-constrained paraphrase generation model can generate a high-quality paraphrase corpus,and the related model has the highest accuracy rate.Compared to the paraphrase model which does not consider proper nouns,the accuracy is increased by 5.38%.
国家科技期刊平台
NSTL
万方数据
