The paper proposes a new paradigm for summarizing online reviews, called rationale-based opinion summarization. Conventional opinion summaries often lack supporting details and can be too generic. To address this, the authors introduce rationale-based opinion summaries, which output representative opinions as well as one or more corresponding rationales (supporting details).
The authors define four desirable properties of good rationales: relatedness, specificity, popularity, and diversity. They present an unsupervised extractive system called RATION that has two components: an Opinion Extractor to extract representative opinions, and a Rationales Extractor to extract corresponding rationales.
The Opinion Extractor uses an existing extractive summarization model to identify summarizing review sentences, and then extracts representative opinions of the form "A is B" from these sentences. To remove redundancy, it uses a graph-based approach to cluster similar representative opinions.
The Rationales Extractor first estimates the four desirable properties of rationales using an alignment model fine-tuned on in-domain data. It then uses Gibbs sampling to extract a user-specified number of rationales for each representative opinion, approximating the joint probability of the rationales based on these properties.
Human evaluation shows that rationale-based summaries generated by RATION are less redundant, more coherent, and more useful for decision-making compared to conventional opinion summaries. Automatic evaluation also demonstrates that the rationales extracted by RATION have better quality than strong baselines.
toiselle kielelle
lähdeaineistosta
arxiv.org
Tärkeimmät oivallukset
by Haoyuan Li,S... klo arxiv.org 04-02-2024
https://arxiv.org/pdf/2404.00217.pdfSyvällisempiä Kysymyksiä