Extracting Evaluative Conditions from Online Reviews: Toward Enhancing Opinion Mining

International Joint Conference on Natural Language Processing, pages 878–882, Nagoya, Japan, 14-18 October 2013. Extracting Evaluative Conditions from Online Reviews: Toward Enhancing Opinion Mining Yuki Nakayama Department of Computer Science Graduate School of Information Science and Engineering Tokyo Institute of Technology {nakayama.y.aj@m,fujii@cs}.titech.ac.jp Atsushi Fujii Abstract A fundamental issue in opinion mining is to search a corpus for opinion units, which typically comprise the evaluation by an au- thor for a target object from an aspect, such as “This hotel is in a good location”. However, no attempt has been made to ad- dress cases where the validity of an eval- uation is restricted on a condition in the source text, such as “for traveling with small kids”. In this paper, we propose a method to extract such conditions, namely evaluative conditions, from sentences in- cluding opinion units. Our method uses supervised machine learning to determine whether each phrase is a constituent of an evaluative condition. We propose several features associated with lexical and syn- tactic information, and show their effec- tiveness experimentally. 1 Introduction Reflecting the rapid growth in the use of opinion- ated texts on the Web, such as customer reviews, opinion mining has been explored to facilitate uti- lizing opinions mainly for improving products and decision-making purposes. While in a broad sense opinion mining refers to a process to discover use- ful knowledge latent in a corpus of opinionated texts, in a narrow sense its purpose is to extract opinions from a corpus. In either case, fundamen- tal issues involve modeling a unit of opinions and searching the corpus for those units, which typi- cally comprise the evaluation by an author for a target object from an aspect. We take the following review sentence as an ex- ample opinionated description. “I think hotel A is in a good location for traveling with small kids”. From the above example, existing methods (Pang and Lee, 2008; Seki et al., 2009; Jin et al., 2009; Zhao et al., 2010; He et al., 2011; Liu and Zhang, 2012) for opinion mining extract the following quintuple as an opinion unit. Target = “hotel A”, Aspect = “location”, Evaluation (Polarity) = “good” (posi- tive), Holder = “I (author)”, Time = N/A Depending on the application, “Evaluation” can be any of a literal evaluation expression (e.g., “good”), a polarity (positive/negative), or a value for multipoint scale rating. However, because this difference is not important in our research, we usu- ally use the term “evaluation”. Given those structured items extracted from a corpus, it is easy to overview the distribution of values for each element or a combination of ele- ments. Those who intend to improve the quality of hotel A may investigate the distribution of val- ues for “Aspect” in the reviews with “Target=hotel A & Polarity=negative”, while those who look for accommodation may compare the distribution of values for “Aspect & Polarity” in reviews for more than one hotel. However, no attempt has been made to address cases where the validity of an evaluation …

Content truncated to 500 words