ILP-based Opinion Sentence Extraction from User Reviews for Question DB Construction

ILP-based Opinion Sentence Extraction from User Reviews for Question DB Construction Masakatsu Hamashita∗and Takashi Inui University of Tsukuba 1-1-1 Tenoudai, Tsukuba, Ibaraki 305-8573, JAPAN {m.hama@mibel.,inui@}cs.tsukuba.ac.jp Koji Murakami and Keiji Shinzato Rakuten Institute of Technology - Boston, Rakuten USA 2 South Station Suite 400, Boston, MA 02110 {koji.murakami,keiji.shinzato}@rakuten.com Abstract Typical systems for analyzing users’ opinions from online product reviews have been re- searched and developed successfully. How- ever, it is still hard to obtain sufficient user opinions when many reviews consist of short messages. This problem can be solved with an active opinion acquisition (AOA) framework that has an interactive interface and can elicit additional opinions from users. In this paper, we propose a method for automatically con- structing a question database (QDB) essential for an AOA. In particular, to eliminate noisy sentences, we discuss a model for extracting opinion sentences that is formulated as a max- imum coverage problem. Our proposed model has two advantages: (1) excluding redundant questions from a QDB while keeping varia- tions of questions and (2) preferring simple sentence structures suitable for the question generation process. Our experimental results show that the proposed method achieved a pre- cision of 0.88. We also give details on the op- timal combination of model parameters. 1 Introduction Typical systems for analyzing users’ opinions from online product reviews have been researched and de- veloped successfully (Liu, 2012; Jo and Oh, 2011; Kouloumpis et al., 2011; Pozzi et al., 2016). How- ever, it is still hard to obtain sufficient user opin- ions when many reviews consist of short messages. In this situation, it would be practical to elicit ad- ditional opinions by actively asking users questions ∗Currently, Gunosy Inc. instead of just waiting for user posts. We define this procedure as an active opinion acquisition (AOA). Suppose an example which is a review post con- sisting of just one sentence below: u1 This wine has a really refreshing aroma! It is possible to capture the user opinion “refresh- ing aroma” from u1. Here, in the case of an AOA- oriented system (AOAS), the system asks a question like s1 after u1. u1 This wine had a really refreshing aroma! s1 How was the aftertaste? u2 The aftertaste was bitter. Then, it is also possible to obtain the additional opin- ion “bitter aftertaste” from u2. This example shows that an AOAS can efficiently collect user opinions by asking users questions. Here, a question database (QDB), that is, a set of large quantities of question examples, is an essential resource for realizing dialogues between a user and an AOAS (Murao et al., 2003) because it would en- able an AOAS to ask users precise questions in vari- ous situations. Nio and Murakami (2018) proposed a question-conversion method for constructing QDBs automatically. This method runs through a machine translation-like architecture and then converts an af- firmative sentence to an interrogative form such as: The aroma was a bouquet. → How was the aroma? (s1) Thank you so much. (s2) The aroma had a nice …

Content truncated to 500 words