Simple Yet Effective Synthetic Dataset Construction for Unsupervised Opinion Summarization

Findings of the Association for Computational Linguistics: EACL 2023, pages 1898–1911 May 2-6, 2023 ©2023 Association for Computational Linguistics Simple Yet Effective Synthetic Dataset Construction for Unsupervised Opinion Summarization Ming Shen♥∗ Jie Ma♣ Shuai Wang♣ Yogarshi Vyas♣ Kalpit Dixit♣ Miguel Ballesteros♣ Yassine Benajiba♣ ♥Arizona State University ♣AWS AI Labs mshen16@asu.edu;{jieman,wshui,yogarshi,kddixit,ballemig,benajiy}@amazon.com Abstract Opinion summarization provides an important solution for summarizing opinions expressed among a large number of reviews. However, generating aspect-specific and general sum- maries is challenging due to the lack of an- notated data. In this work, we propose two simple yet effective unsupervised approaches to generate both aspect-specific and general opinion summaries by training on synthetic datasets constructed with aspect-related review contents. Our first approach, Seed Words Based Leave-One-Out (SW-LOO), identifies aspect- related portions of reviews simply by exact- matching aspect seed words and outperforms existing methods by 3.4 ROUGE-L points on SPACE and 0.5 ROUGE-1 point on OPOSUM+ for aspect-specific opinion summarization. Our second approach, Natural Language Inference Based Leave-One-Out (NLI-LOO) identifies aspect-related sentences utilizing an NLI model in a more general setting without using seed words and outperforms existing approaches by 1.2 ROUGE-L points on SPACE for aspect- specific opinion summarization and remains competitive on other metrics. 1 Introduction Customer reviews play a vital role in decision- making for customers and product (or business) providers, as customers usually resort to reviews to guide their purchasing decisions and product providers improve their products based on reviews as feedback. However, it becomes hard for cus- tomers or product providers to read through all reviews before making decisions with the explo- sion of online reviews in recent years. Opinion summarization (Hu and Liu, 2006; Wang and Ling, 2016; Angelidis and Lapata, 2018; Bražinskas et al., 2020; Brazinskas et al., 2022; Angelidis et al., 2021; Amplayo et al., 2021a; Basu Roy Chowdhury ∗Work done during an internship at AWS AI Labs. et al., 2022), the task of generating a general sum- mary of salient opinions expressed among reviews, provides a feasible solution to this problem. Different from summarization in Wikipedia and news domains (Nallapati et al., 2016; Narayan et al., 2018a; See et al., 2017; Narayan et al., 2018b; Liu and Lapata, 2019; Cachola et al., 2020), opinion summarization cannot rely on reference summaries for model training since it is difficult and expen- sive to annotate large scale reviews-summary pairs. Also, customers usually care about specific aspects of a product instead of a general high-level sum- mary. Thus, fine-grained aspect-specific opinion summaries are required, and this makes the annota- tion process even more difficult and expensive. Amplayo et al. (2021a) propose an abstractive approach to generate aspect-specific opinion sum- maries by training on synthetic datasets. They construct synthetic datasets with review elements (words, phrases, or sentences) identified by a mul- tiple instance learning (MIL) module (Keeler and Rumelhart, 1991) learned with silver-standard la- bels obtained using aspect seed words. We first follow this direction to propose a more straightfor- ward and effective method that excludes the com- plex learning module to …

Content truncated to 500 words