Aspect and Sentiment Aware Abstractive Review Summarization

Proceedings of the 27th International Conference on Computational Linguistics, pages 1110–1120 Santa Fe, New Mexico, USA, August 20-26, 2018. 1110 Aspect and Sentiment Aware Abstractive Review Summarization Min Yang1, Qiang Qu1∗, Ying Shen2, Qiao Liu3, Wei Zhao1, Jia Zhu4 1 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences 2 Peking University Shenzhen Graduate School 3 University of Electronic Science and Technology of China 4 South China Normal University {min.yang,qiang,wei.zhao}@siat.ac.cn, shenying@pkusz.edu.cn qliu@uestc.edu.cn, jzhu@m.scnu.edu.cn Abstract Review text has been widely studied in traditional tasks such as sentiment analysis and aspect extraction. However, to date, no work is towards the end-to-end abstractive review summa- rization that is essential for business organizations and individual consumers to make informed decisions. This work takes the lead to study the aspect/sentiment-aware abstractive review sum- marization in an end-to-end manner without hand-crafted features and templates by exploring the encoder-decoder framework and multi-factor attentions. Specifically, we propose a mutual atten- tion mechanism to interactively learns the representations of context words, sentiment words and aspect words within the reviews, acted as an encoder. The learned sentiment and aspect represen- tations are incorporated into the decoder to generate aspect/sentiment-aware review summaries via an attention fusion network. In addition, the abstractive summarizer is jointly trained with the text categorization task, which helps learn a category-specific text encoder, locating salient aspect information and exploring the variations of style and wording of content with respect to different text categories. The experimental results on a real-life dataset demonstrate that our model achieves impressive results compared to other strong competitors. 1 Introduction User generated reviews on products are expanding rapidly with the emergence and advancement of e- commerce. These reviews are valuable to business organizations for improving their products and to individual consumers for making informed decisions. Unfortunately, reading though all the product reviews is hard, especially for the reviews that are lengthy and have low readability. It is therefore es- sential to provide coherent and concise summaries of user generated reviews. In this paper, we focus on generating abstractive summaries of product reviews. Abstractive text summarization is the task of generating a short and concise summary that captures the salient ideas of the source text. The gener- ated summaries potentially contain new phrases and sentences that may not appear in the source text. Inspired by recent success of sequence-to-sequence (seq2seq) model in statistical machine translation, most abstractive summarization systems employ seq2seq framework to generate summaries (Nallapati et al., 2016; See et al., 2017; Paulus et al., 2017). In general, the seq2seq model firstly uses an encoder to convert the input text as a vector representation, and then it feeds this representation into a decoder to generate summary. Despite the remarkable progress of previous studies, generating aspect/sentiment-aware summaries of product reviews remains a challenge in real-world for two reasons. (i) First, neural sequence-to- sequence models tend to generate trivial and generic summary, often involving high-frequency phrases. These summaries cannot capture the aspect and sentiment information from the product reviews which play a vital role in …

Content truncated to 500 words