The Research on Abstractive Summarization Method Based on Topic Model and Contrastive Learning

In the field of text summarization, generative models based on Transformer have made significant progress. However, they suffer from exposure bias and loss of global semantic information when processing text. This paper proposes a text summarization method based on topic models and contrastive learning. By introducing topic models to obtain the relationship between words and latent topics, and designing an alignment function to fuse the representation vectors generated by the topic model with the input text representation, it ensures that topic information is effectively utilized by the model. In addition, contrastive learning is used to adjust the model's loss function, bridging the gap between the training and inference stages, effectively alleviating the exposure bias problem. Experimental results on the NLPCC dataset show that this method effectively alleviates the exposure bias problem and retains key information in the text, achieving higher ROUGE scores.