Rationale-based Opinion Summarization

TL;DR

Rationale-based opinion summarization addresses the shortcomings of generic, evidence-poor summaries by producing representative opinions paired with concise rationales. The authors introduce RATION, an unsupervised, extractive system with an Opinion Extractor and a Rationales Extractor that optimize four rationale properties (relatedness, specificity, popularity, diversity) via Gibbs sampling, guided by a domain-adapted alignment model. Through automatic and human evaluations on Space and Yelp, RATION yields more informative, less redundant, and more decision-useful summaries than strong baselines, and its rationales are of higher quality than those produced by comparators such as InstructGPT or SemAE. This work advances explainable, user-centered opinion summarization with practical implications for search, recommendation, and consumer decision support; the code is publicly available for reproducibility and further development.

Abstract

Opinion summarization aims to generate concise summaries that present popular opinions of a large group of reviews. However, these summaries can be too generic and lack supporting details. To address these issues, we propose a new paradigm for summarizing reviews, rationale-based opinion summarization. Rationale-based opinion summaries output the representative opinions as well as one or more corresponding rationales. To extract good rationales, we define four desirable properties: relatedness, specificity, popularity, and diversity and present a Gibbs-sampling-based method to extract rationales. Overall, we propose RATION, an unsupervised extractive system that has two components: an Opinion Extractor (to extract representative opinions) and Rationales Extractor (to extract corresponding rationales). We conduct automatic and human evaluations to show that rationales extracted by RATION have the proposed properties and its summaries are more useful than conventional summaries. The implementation of our work is available at https://github.com/leehaoyuan/RATION.

Figures (10)

• Figure 1: Examples of a conventional and a rationale-based opinion summary (generated by RATION ) for the same entity. In rationale-based summary, each line presents a representative opinion and its rationale.
• Figure 2: Overview of RATION and its two components: the Opinion Extractor and the Rationales Extractor.
• Figure 3: Three sentences and their constituency parsing trees. A orange box denotes one extracted clause.
• Figure 4: Three sample rationale-based summaries. Each line presents a representative opinion and its rationale.
• Figure 5: Example instruction for extracting one rationale for each representative opinion using InstructGPT.