Abstractive Aspect-Based Comparative Summarization

Hyeon Jin, Chaewon Yoon, Yurim Oh, Hyun-Je Song

Abstract: Comparative summarization aims to generate a set of summaries that highlight the relevant differences and commonalities between two comparable entities. While these summaries provide users with general comparative information, they may not fully meet users' specific information needs, particularly when users seek detailed information about specific aspects of the entities. In this paper, we introduce the task of abstractive aspect-based comparative summarization, which identifies the aspects of entities from a set of two reviews and then generates abstractive contrastive and common summaries for each aspect. To support this task, we construct two new datasets and propose a simple large language model-based summarization model that generates both aspects and the corresponding contrastive and common summaries. Experimental results on the newly constructed datasets demonstrate that the proposed summarization model can generate higher-quality aspect-based comparative summaries compared to the baselines.