Opinion Summarization via Submodular Information Measures

This paper focuses on opinion summarization for constructing subjective and concise summaries representing essential opinions of online text reviews. As previous works rarely focus on the relationship between opinions, topics, and sentences, we propose a set of new requirements for Opinion-Topic-Sentence, which are essential for performing opinion summarization. We prove that Opinion-Topic-Sentence can be theoretically analyzed by submodular information measures. Thus, our proposed method can reduce redundant information, strengthen the relevance to given topics, and informatively represent the underlying emotional variations. While conventional methods require human-labeled topics for extractive summarization, we use unsupervised topic modeling methods to generate topic features. We propose four submodular functions and two optimization algorithms with proven performance bounds that can maximize opinion summarization's utility. An automatic evaluation metric, Topic-based Opinion Variance, is also derived to compensate for ROUGE-based metrics of opinion summarization evaluation. Four large, diversified, and representative corpora, OPOSUM, Opinosis, Yelp, and Amazon reviews, are used in our study. The results on these online review texts corroborate the efficacy of our proposed metric and framework.