HELPFULSUMM: Helpful Personalized Opinion Summarization via Reinforcement Learning from Review Helpfulness Votes

ACL ARR 2026 January Submission7371 Authors

Keywords: personalized opinion summarization, helpful review summarization, opinion summarization

Abstract: Personalized opinion summarization (POS) aims to generate a targeted summary of product reviews that is tailored to an individual user's needs and interests. Few existing studies mainly rely on persona representations derived from user-written reviews for personalization, which may not fully capture user interests and fail on cold-start users who have not authored reviews. In real-life social platform, helpfulness votes on reviews represent opinions helpful and of interest to users. In this paper, we propose HELPFULSUMM, a reinforcement-learning-based model that utilizes user historical helpfulness votes for alignment with user preference in both Knowledge Consistency and Persona Consistency, using dual rewards: (i) Helpful Opinion and (ii) Persona Alignment. Experimental results show that HELPFULSUMM outperforms existing persona-based and general opinion summarization approaches and provide more helpful opinions and at higher information and personalization quality. Our source code is available at: https://anonymous.4open.science/r/HELPFULSUMM-A233

Paper Type: Long

Research Area: Summarization

Research Area Keywords: Summarization

Contribution Types: Data analysis

Languages Studied: English

Submission Number: 7371