Incorporating Rating: Abstractive Summarization of Review Sentences Considering Rating

Abstract

Document summarization is very effective when a review on a review site is lengthy. Abstractive summarization is flexible because it can include words and expressions not included in the original document. On the other hand, because this method is based on deep learning, if the dataset to be trained has a lot of positive content, the generated summary will be biased toward positive content. In this study, we propose an abstractive summarization model that takes into account user evaluations unique to web reviews. Specifically, we propose a method for improving the accuracy of the model by incorporating the rating of reviews into the model. The abstractive summarization model is based on a specialized model for abstractive summarization called PEGASUS. As a result, the learning model using the proposed method improves the RO U G E value compared with existing methods. In addition, the proposed method obtained a better human evaluation than existing methods, denending on the domain and method.