Aspect-Oriented Sentiment Analysis of Customer Reviews Using Distant Supervision Techniques

This thesis provides a comprehensive study of how to model and automatically analyze the opinion-rich information contained in customer reviews and examines the utility of distant supervision techniques to reduce the amount of required human supervision.

Abstract

The opinions and experiences of other people constitute an important source of information in our everyday life. For example, we ask our friends which dentist, restaurant, or smartphone they would recommend to us. Nowadays, online customer reviews have become an invaluable resource to answer such questions. Besides helping consumers to make more informed purchase decisions, online reviews are also of great value to vendors, as they represent unsolicited and genuine customer feedback that is conveniently available at virtually no costs. However, for popular products there often exist several thousands of reviews so that manual analysis is not an option. In this thesis, we provide a comprehensive study of how to model and automatically analyze the opinion-rich information contained in customer reviews. In particular, we consider the task of aspect-oriented sentiment analysis. Given a collection of review texts, the task’s goal is to detect the individual product aspects reviewers have commented on and to decide whether the comments are rather positive or negative. Developing text analysis systems often involves the tedious and costly work of creating appropriate resources --- for instance, labeling training corpora for machine learning methods or constructing special-purpose knowledge bases. As an overarching topic of the thesis, we examine the utility of distant supervision techniques to reduce the amount of required human supervision. We focus on the two main subtasks of aspect- oriented review mining: (i) identifying relevant product aspects and (ii) determining and classifying expressions of sentiment. We consider both subtasks at two different levels of granularity, namely expression vs. sentence level. For these different levels of analysis, we experiment with dictionary-based and supervised approaches and examine several distant supervision techniques. For aspect detection at the expression level, we cast the task as a terminology extraction problem. At the sentence level, we cast the task as a multi-label text categorization problem and exploit section headings in review texts for a distant supervision approach. With regard to sentiment analysis, we present detailed studies of sentiment lexicon acquisition and sentiment polarity classification and show how pros and cons summaries of reviews can be exploited to reduce the manual effort in this context. We evaluate our approaches in detail, including insightful mistake analyses. For each of the tasks, we find significant improvements in comparison to relevant state-of-the-art methods. In general, we can show that the presented distant supervision methods successfully reduce the required amount of human supervision. Our approaches allow to gather very large amounts of labeled data --- typically some orders of magnitude more data than possible with traditional annotation. We conclude that customer review mining systems can benefit from the proposed methods.