Mining opinion summarizations using convolutional neural networks in Chinese microblogging systems

详细信息    查看全文

• 作者：Qiudan Li^a ; ^{qiudan.li@ia.ac.cn" class="auth_mail" title="E-mail the corresponding author} ; Zhipeng Jin^a ; ^b ; ^{jinzhipeng2013@ia.ac.cn" class="auth_mail" title="E-mail the corresponding author} ; Can Wang^a ; ^b ; ^{wangcan2015@ia.ac.cn" class="auth_mail" title="E-mail the corresponding author} ; Daniel Dajun Zeng^a ; ^b ; ^c ; ^{zeng@email.arizona.edu" class="auth_mail" title="E-mail the corresponding author}
• 关键词：Chinese microblogging systems ; Hot topics ; Convolutional neural network ; Opinion summarization ; Maximal marginal relevance
• 刊名：Knowledge-Based Systems
• 出版年：2016
• 出版时间：1 September 2016
• 年：2016
• 卷：107
• 期：Complete
• 页码：289-300
• 全文大小：1540 K

文摘

Chinese microblogging is an increasingly popular social media platform. Accurately summarizing representative opinions from microblogs can increase understanding of the semantics of opinions. The unique challenges of Chinese opinion summarization in microblogging systems are automatic learning of important features and selection of representative sentences. Deep-learning methods can automatically discover multiple levels of representations from raw data instead of requiring manual engineering. However, there have been very few systematic studies on sentiment analysis of Chinese hot topics using deep-learning methods. Based on the latest deep-learning research, in this paper, we propose a convolutional neural network (CNN)-based opinion summarization method for Chinese microblogging systems. The model first applies CNN to automatically mine useful features and perform sentiment analysis; then, by making good use of the obtained sentiment features, the semantic relationships among features are computed according to a hybrid ranking function; and finally, representative opinion sentences that are semantically related to the features are extracted using Maximal Marginal Relevance, which meets “relevant novelty” requirements. Experimental results on two real-world datasets verify the efficacy of the proposed model.