| Keyphrase generation is a fine-grained task in the field of natural language processing.It can be applied to many research tasks and business scenarios such as text classification,text clustering and information retrieval.In recent years,it has gradually attracted more attention of researchers,and has been greatly developed in terms of data sets,evaluation criteria and model structure.This paper focuses on the keyphrase generation task.Through the analysis of the dataset,we found that keyphrases are mostly composed of adjectives or nouns.However,the existing research on keyphrase generation ignores part of speech.Especially,when the copy network module copies words from the original text,it treats all equally regardless of their part of speech.In addition,by analyzing the keyphrases generated by the existing models,we found that the existing models have wrong biases,and it is easy to generate substrings of standard keyphrases or phrases frequently appearing in the original text as keyphrases,which is obviously incorrect.To deal with the above issues,we propose two novel methods as follows.(1)Use part of speech to enhance keyphrase generation.Specifically,in addition to using part of speech to enhance word embedding,we also use part of speech to enhance copy network to copy words from the original text.By using part of speech to enhance word embedding representation,word embedding can contain more abundant semantic information;part of speech guides the copy network to copy nouns and adjectives as much as possible.And they complement each other to guide the model to generate more nouns or adjectives.(2)Propose a training method based on contrast learning for keyphrase generation.Based on comparative learning,the model can learn the differences between the wrong keyphrases and the standard keyphrases.Thus,the model can reduce the probability of generating the wrong keyphrases and generate higher-quality keyphrases.In this paper,sufficient experiments are carried out on existing benchmark datasets and compared with existing keyphrase generation models.The experimental results show the effectiveness of our method. |
