Opinions Summarization: Aspect Similarity Recognition Relaxes The Constraint of Predefined Aspects

Proceedings of Recent Advances in Natural Language Processing, pages 487–496, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_058 487 Opinions Summarization: Aspect Similarity Recognition Relaxes the Constraint of Predefined Aspects Huy Tien Nguyen Japan Advanced Institute of Science and Technology (JAIST), Japan ntienhuy@jaist.ac.jp Tung Le Japan Advanced Institute of Science and Technology (JAIST), Japan lttung@jaist.ac.jp Minh Le Nguyen Japan Advanced Institute of Science and Technology (JAIST), Japan nguyenml@jaist.ac.jp Abstract Recently research in opinions summariza- tion focuses on rating expressions by as- pects and/or sentiments they carry. To ex- tract aspects of an expression, most stud- ies require a predefined list of aspects or at least the number of aspects. Instead of extracting aspects, we rate expressions by aspect similarity recognition (ASR), which evaluates whether two expressions share at least one aspect. This subtask re- laxes the limitation of predefining aspects and makes our opinions summarization applicable in domain adaptation. For the ASR subtask, we propose an attention-cell LSTM model, which integrates attention signals into the LSTM gates. According to the experimental results, the attention-cell LSTM works efficiently for learning latent aspects between two sentences in both set- tings of in-domain and cross-domain. In addition, the proposed extractive summa- rization method using ASR shows signif- icant improvements over baselines on the Opinosis corpus. 1 Introduction Opinions Summarization is the collection of typ- ical opinions mentioned in social media, blogs or forums on the web. This task helps customers to absorb better a large number of comments and re- views before making decisions as well as produc- ers to keep track of what customers think about their products (Liu, 2012). Due to the fast growth of data over the Inter- net, automatically opinions summarization has re- ceived a lot of attention in recent years. Most re- search focus on extractive summarization, where the most salient text units are identified and construct a summary. Ranking candidates for generic summarization usually bases on various handcrafted features such as sentence position and length (Radev et al., 2004), word frequency (Nenkova et al., 2006) or using neural networks for learning salient scores (Zhou et al., 2018). In opinions summarization, however, this task is required to consider aspects and/or sentiments of text candidates for generating a concise and infor- mative summary (Hu and Liu, 2006). The popular framework of this problem involves three subtasks (Hu and Liu, 2004): i) aspect discovery which ex- tracts the properties of interested entities (e.g., bat- tery life, design, customer service); ii) sentiment analysis which assigns sentiment polarity (positive and negative) towards the aspects extracted in the first step; and iii) summary generation which se- lects the most salient opinions to build a summary. For the aspect discovery task, there are two main techniques: supervised and unsupervised learning. The former models the aspect extraction as a sequence labeling task. Due to predefining a list of aspect and heavily relying on annotated data, this approach suffers from domain adapta- tion problems. The latter uses a large amount of unlabeled data for abstracting aspects via the …

Content truncated to 500 words