Abstractive Summarization of Product Reviews Using Discourse Structure

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1602–1613, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Abstractive Summarization of Product Reviews Using Discourse Structure Shima Gerani‡† ∗ Yashar Mehdad‡ ∗ Giuseppe Carenini‡ Raymond T. Ng‡ Bita Nejat‡ †University of Lugano ‡University of British Columbia Switzerland Vancouver, BC, Canada {gerani,mehdad,carenini,rng,nejatb}@cs.ubc.ca Abstract We propose a novel abstractive summa- rization system for product reviews by tak- ing advantage of their discourse structure. First, we apply a discourse parser to each review and obtain a discourse tree repre- sentation for every review. We then mod- ify the discourse trees such that every leaf node only contains the aspect words. Sec- ond, we aggregate the aspect discourse trees and generate a graph. We then select a subgraph representing the most impor- tant aspects and the rhetorical relations be- tween them using a PageRank algorithm, and transform the selected subgraph into an aspect tree. Finally, we generate a natural language summary by applying a template-based NLG framework. Quan- titative and qualitative analysis of the re- sults, based on two user studies, show that our approach significantly outperforms ex- tractive and abstractive baselines. 1 Introduction Most existing works on sentiment summarization focus on predicting the overall rating on an en- tity (Pang et al., 2002; Pang and Lee, 2004) or estimating ratings for product features (Lu et al., 2009; Lerman et al., 2009; Snyder and Barzilay, 2007; Titov and McDonald, 2008)). However, the opinion summaries in such systems are extractive, meaning that they generate a summary by concate- nating extracts that are representative of opinion on the entity or its aspects. Comparing extractive and abstractive sum- maries for evaluative texts has shown that an ab- stractive approach is more appropriate for sum- marizing evaluative text (Carenini et al., 2013; ∗The contribution of the first two authors to this paper was equal. Di Fabbrizio et al., 2014). This finding is also supported by a previous study in the context of summarizing news articles (Barzilay et al., 1999). To the best of our knowledge, there are only three previous works on abstractive opinion summariza- tion (Ganesan et al., 2010; Carenini et al., 2013; Di Fabbrizio et al., 2014). The first work (Gane- san et al., 2010) proposes a graph-based method for generating ultra concise opinion summaries that are more suitable for viewing on devices with small screens. This method does not provide a well-formed grammatical abstract and the gener- ated summary only contains words that occur in the original texts. Therefore, this approach is more extractive than abstractive. Another limitation is that the generated summaries do not contain any information about the distribution of opinions. In the second work, (Carenini et al., 2013) ad- dresses some of the aforementioned problems and generates well-formed grammatical abstracts that describe the distribution of opinion over the en- tity and its features. However, for each product, this approach requires a feature taxonomy hand- crafted by humans as an input, which is not scal- able. …

Content truncated to 500 words