ALS-MRS: Incorporating aspect-level sentiment for abstractive multi-review summarization

Abstract

Multi-review summarization, the process of automatically generating a summary for a set of reviews, is of value to customers when they are making informed decisions. Incorporating sentiments can improve the performance of abstractive review summarization, as shown in some prior studies. However, these studies assume that each review contains a single sentiment, which does not reflect reality in many real conditions. In practice, a review always contains many aspects with conflicting sentiments. In this paper, we present ALS-MRS, a novel abstractive multi-review summarization model that combines multi-review representations and aspect-level sentiment. We propose an aspect-level sentiment consistency function to keep the sentiments of the various aspects in the generated summaries the same as those in the reference summaries. Specifically, ALS-MRS constructs aspect-sentiment tuples via an aspect extractor and a sentiment analysis model. In the aspect extractor, the aspects are identified according to the aspect terms obtained by an unsupervised neural attention model, and the sentiment polarity of a sentence about the aspect is detected in the sentiment analysis model. The experimental results, evaluated by automatic and human metrics on two public datasets, show that ALS-MRS performs favorably when compared against many state-of-the-art approaches.