Attributable and Scalable Opinion Summarization

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8488–8505 July 9-14, 2023 ©2023 Association for Computational Linguistics Attributable and Scalable Opinion Summarization Tom Hosking Hao Tang Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB tom.hosking@ed.ac.uk hao.tang@ed.ac.uk mlap@inf.ed.ac.uk Abstract We propose a method for unsupervised opin- ion summarization that encodes sentences from customer reviews into a hierarchical discrete latent space, then identifies common opinions based on the frequency of their encodings. We are able to generate both abstractive summaries by decoding these frequent encodings, and ex- tractive summaries by selecting the sentences assigned to the same frequent encodings. Our method is attributable, because the model iden- tifies sentences used to generate the summary as part of the summarization process. It scales easily to many hundreds of input reviews, be- cause aggregation is performed in the latent space rather than over long sequences of to- kens. We also demonstrate that our appraoch enables a degree of control, generating aspect- specific summaries by restricting the model to parts of the encoding space that correspond to desired aspects (e.g., location or food). Auto- matic and human evaluation on two datasets from different domains demonstrates that our method generates summaries that are more in- formative than prior work and better grounded in the input reviews. 1 Introduction Online review websites are a useful resource when choosing which hotel to visit or which product to buy, but it is impractical for a user to read hundreds of reviews. There has been significant interest in methods for automatically generating summaries or meta-reviews that aggregate the diverse opinions contained in a set of customer reviews about an entity (e.g., a product, hotel or restaurant) into a single summary. Early work on opinion summarization extracted reviewers’ sentiment about specific features (Hu and Liu, 2004) or selected salient sentences from re- views based on centrality (Erkan and Radev, 2004), while more recent methods based on neural mod- els have used sentence selection in learned feature spaces (Angelidis et al., 2021; Basu Roy Chowd- hury et al., 2022) or abstractive summarizers that generate novel output (Bražinskas et al., 2020, 2021; Amplayo et al., 2021a,b; Iso et al., 2021; Coavoux et al., 2019). Following Ganesan et al. (2010), we define opin- ion summarization, or review aggregation, as the task of generating a textual summary that reflects frequent or popular opinions expressed in a large number of reviews about an entity. Systems are extractive if they select sentences or spans from the input reviews to use as the summary, or ab- stractive if they generate novel output. Review aggregation is challenging for a number of reasons. Firstly, it is difficult to acquire or create reference summaries, so models are almost always trained without access to gold standard references (Ange- lidis et al., 2021; Amplayo et al., 2021b, inter alia.). Secondly, popular entities may have hundreds of reviews, which can cause computational difficul- …

Content truncated to 500 words