Opinion Summarization by Weak-Supervision from Mix-structured Data

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3086 - 3096 December 7-11, 2022 ©2022 Association for Computational Linguistics Opinion Summarization by Weak-Supervision from Mix-structured Data Yizhu Liu1 Qi Jia2 Kenny Q. Zhu3∗ Shanghai Jiao Tong University Shanghai, China {1liuyizhu, 2Jia_qi}@sjtu.edu.cn, 3kzhu@cs.sjtu.edu.cn Abstract Opinion summarization of multiple reviews suf- fers from the lack of reference summaries for training. Most previous approaches construct multiple reviews and their summary based on textual similarities between reviews, resulting in information mismatch between the review in- put and the summary. In this paper, we convert each review into a mix of structured and un- structured data, which we call opinion-aspect pairs (OAs) and implicit sentences (ISs). We propose a new method to synthesize training pairs of such mix-structured data as input and the textual summary as output, and design a summarization model with OA encoder and IS encoder. Experiments show that our approach outperforms previous methods on Yelp, Ama- zon and RottenTomatos datasets. 1 Introduction Opinion summarization aims at producing a sum- mary for a set of subjective user reviews about an entity (i.e., a product, a service or a movie), which can help users quickly understand the en- tity. Opinion summarization is a specialized multi- document summarization (Fabbri et al., 2019; Liu et al., 2018a; Ma, 2021). Unlike traditional multi- document summarization, opinion summarization focuses on the aspects and opinions of entities in the input documents. Deep learning techniques have made great successes in summarization (See et al., 2017; Liu et al., 2018b; Lewis et al., 2020; Liu et al., 2021), which require training on a large number of document-summary pairs. Unfortu- nately, opinion summarization generally lacks the training pairs with reviews as input and summary as output, as it is difficult and costly for annotators to write summaries for multiple reviews (known as multi-reviews in this paper) on a large scale. ∗Kenny Q. Zhu is the corresponding author, and is par- tially supported by SJTU-CMBCC Joint Research Scheme and SJTU-Meituan Joint Research Scheme. Synthetic Summary (sampled from the reviews) very disappointed in food and service . the beef was burned . the staff dismissed my comments . Synthetic Input (to be summarized) Textual (unstructured) Random R1 very average spanish food . R2 definitely not a peruvian restaurant . R3 love this place! location is great . Similarity R1 bad food and service . disappointed R2 the food and service was not good . very bad expe- rience . R3 awful food and service . the staff was unfriendly . Structured OpiDig OA disappointed, food; disappointed, service; burned, beef Mix-structured Ours OA not good, food; great, location; bad, food; disap- pointed, service; not fresh, beef; unfriendly, staff; awful, service IS not recommended .; my questions were dismissed .; the staff ignored our comments . Table 1: A summary and various inputs synthesized by different methods. Matched words in both input and output are bolded. R denotes textual review. OA indicates explicit opinion-aspect pairs. The italicized sentence doesn’t contain explicit …

Content truncated to 500 words