Extractive Negative Opinion Summarization of Consumer Electronics Reviews

Online consumers create enormous reviews of electronic devices or services daily. Extracting negative opinions from such an amount of data is a crucial task for improving products and developing new features. Opinion summarization can help public consumers and businesses understand and extract the proper amount of negative information from large-scale data. However, automatically and concisely summarizing opinions with negative emotions and sentiments has yet to be explored. This paper proposes an extractive summarization framework that automatically detects fine-grained negative opinions. While the conventional opinion summarization only considers a general full affective coverage, our proposed method exploits submodular diversity, relevance, and opinion functions focusing on summarizing reviews with negative emotional variations. At the same time, an algorithm with 1-1/e-ϵ -approximation is applied to optimize the proposed functions. Most of the existing datasets cannot provide golden summaries with negative opinions. Our experiment explores reference-free metrics for evaluation, which requires neither reference nor human-created golden summaries. According to the metric scores, the proposed framework outperforms all baselines at summarizing negative opinions of consumer electronics on eight popular online shopping platforms. We analyze the generated summaries in detail and provide a possible application example in electronic product development.