Comparative Opinion Summarization via Collaborative Decoding

Opinion summarization focuses on generating summaries that reflect popular opinions of multiple reviews for a single entity (e.g., a hotel or a product.) While generated summaries offer general and concise information about a particular entity, the information may be insufficient to help the user compare multiple entities. Thus, the user may still struggle with the question "Which one should I pick?" In this paper, we propose the comparative opinion summarization task, which aims at generating two contrastive summaries and one common summary from two given sets of reviews of different entities. We develop a comparative summarization framework COCOSUM, which consists of two few-shot summarization models that jointly generate contrastive and common summaries. Experimental results on a newly created benchmark COCOTRIP show that COCOSUM can produce higher-quality contrastive and common summaries than stateof-the-art opinion summarization models. represents salient opinions in input reviews. How-042 ever, existing opinion summarization techniques 043 are designed to generate a single-entity opinion 044 summary that reflects popular opinions for each 045 entity, without taking into account contrastive and 046 common opinions that are uniquely (commonly) 047 mentioned in each entity (both entities) as depicted 048 in Figure 1. Therefore, the user still needs to fig-049 ure out which opinions are distinctive or common 050 between the entities by carefully reading and com-051 paring summaries generated by existing opinion 052 summarization solutions. 053 To this end, we take one step beyond the cur-054 rent scope of opinion summarization and propose 055 a novel task of generating contrastive and common 056 2.2 The COCOTRIP Corpus 131 As the task requires three types of reference sum-132 maries for each entity pair, none of the existing 133 benchmarks for single-entity opinion summariza-134 tion can be used for evaluation. Therefore, we 135 create a comparative opinion summarization cor-136 pus COCOTRIP that contains human-written con-137 trastive and common summaries for 50 pairs of 138 entities. We sampled the entity pairs and reviews 139 from the TripAdvisor corpus (Wang et al., 2010). 140 We sampled 16 reviews for every pair (i.e., 8 141 reviews for each entity.) For every entity pair, we 142 collected 3 gold-standard summaries written by 143 different annotators for two contrastive summaries 144 and one common summary. Details of the corpus 145 creation process are described in Appendix. 146 We summarize the COCOTRIP dataset and 147 compare it with existing opinion summarization 148 datasets in Table 1. Our dataset contains a similar 149 scale of summaries to existing abstractive opinion 150 summarization datasets, and the input reviews are 151 about three times longer than others. 152 3 COCOSUM 153 For single-entity opinion summarization, input re-154 views can be used as pseudo summaries for training 155 summarization models in a self-supervised fashion. 156 This approach is not suitable for comparative opin-157 ion summarization as the task takes two sets of 158 reviews for different entities to generate contrastive 159 and common summaries, which have significantly 160 different characteristics from the original review as 161 supported by Table 1. In addition, recent studies 162 583