Prompted Aspect Key Point Analysis for Quantitative Review Summarization

Key Point Analysis (KPA) aims for quantitative summarization that provide key points (KPs) as succinct textual summaries and quantities measuring their prevalence. KPA studies for argument and reviews have been reported in the literature. Majority of KPA studies for reviews adopt supervised learning to extract short sentences as KPs and matching KPs to review comments for quantification of KP prevalence. Recent abstractive approaches still generate KPs based on sentences, often leading to KPs with overlapping and hallucinated opinions, and inaccurate quantification. In this paper, we propose Prompted Aspect Key Point Analysis (PAKPA) for quantitative review summarization. PAKPA employs aspect sentiment analysis and prompt in-context learning with Large Language Models (LLMs) to generate and quantify KPs grounded in aspects for business entities, which achieves faithful KPs with accurate quantification, and remove the need for large amounts of annotated data for supervised training. Experiments on the popular review dataset Yelp and the aspect-oriented review summarization dataset SPACE show that our framework achieves state-of-the-art performance. Source code and data are available at: https://anonymous.4open.science/r/ PAKPA-A233 1 Introduction With the sheer volume of reviews, it is impossible for humans to read all reviews. Although the star ratings aggregated from customer reviews are widely used by E-commerce platforms as indicators of quality of service for business entities (Mc-Glohon et al., 2010; Tay et al., 2020), they can not explain specific details for informed decision making. Early studies on review comment (text) summarization focused to capture important points with high consensus (Dash et al., 2019; Shandilya et al., 2018), yet overlooked to include minor ones and also unable to measure the opinion prevalence. Key Point Analysis (KPA), is proposed to sum-043 marize opinions in review comments into con-044 cise textual summaries called key points (KPs), 045 and quantify the prevalence of KPs. KPA studies 046 were initially developed for argument summariza-047 tion (Bar-Haim et al., 2020a), and then adapted to 048 business reviews (Bar-Haim et al., 2020b, 2021). 049 Most KPA studies adopt the extractive approach, 050 which employs supervised learning to identify in-051 formative short sentences as Key Points (KPs), 052 which often leads to non-readable adn incoher-053 ent KPs. Recently, KPA studies apply abstrac-054 tive summarization methods to paraphrase and gen-055 erate KPs from comments (sentences) (Kapadnis 056 et al., 2021; Li et al., 2023). In summary, existing 057 sentence-based KPA systems, whether extractive 058 or abstractive, often generate KPs containing over-059 lapping opinions, and inaccurate quantity for their 060 prevalence. 061 In this paper we propose Prompted Aspect Key 062 Point Analysis (PAKPA). Different from previous 063 sentence-based KPA studies, we propose to employ 064 aspet sentiment analysis to identify aspects in com-065 ments as the opinion target and then generate and 066 quantify KPs grounded in aspects and their senti-067 ment. Importantly, we employ prompt in-context 068 learning with LLMs for aspect sentiment analysis 069 of comments and KP generation, deviating from the 070 supervised learning approach in most KPA studies. 071 Our contribution are two-fold. To our best 072 knowledge, we are the first to employ prompt con-073 text learning for abstractive KPA summarization of 074 reviews, which removes supervised training using 075 large amount of annotated data. Secondly, our ap-076 proach of integration of aspect sentiment analysis 077 (ABSA) into KPA for fine-grained opinion anal-078 ysis of review comments ensures generating KPs 079 grounded in aspects for business entities and more 080 accurate matching of comments to KPs, resulting 081 in faithful KPs for distinct aspects as well as more 082 accurate quantification of KP prevalence. 083 133 (ABSA) to extract, aggregate, and organize review 134 sentences into a hierarchy based on features (i.e. 135 aspects) such as food, price, service, and their senti-136