Care more about customers: Unsupervised domain-independent aspect detection for sentiment analysis of customer reviews

Highlights

• We proposed a model used for detecting explicit and implicit aspects in sentiment analysis.

• The model performs the detection with finding single and multi-word aspects, filtering by A-score metric and pruning.

• Experimental results show considerable improvements of the proposed model over conventional techniques.

Abstract

With the rapid growth of user-generated content on the internet, automatic sentiment analysis of online customer reviews has become a hot research topic recently, but due to variety and wide range of products and services being reviewed on the internet, the supervised and domain-specific models are often not practical. As the number of reviews expands, it is essential to develop an efficient sentiment analysis model that is capable of extracting product aspects and determining the sentiments for these aspects. In this paper, we propose a novel unsupervised and domain-independent model for detecting explicit and implicit aspects in reviews for sentiment analysis. In the model, first a generalized method is proposed to learn multi-word aspects and then a set of heuristic rules is employed to take into account the influence of an opinion word on detecting the aspect. Second a new metric based on mutual information and aspect frequency is proposed to score aspects with a new bootstrapping iterative algorithm. The presented bootstrapping algorithm works with an unsupervised seed set. Third, two pruning methods based on the relations between aspects in reviews are presented to remove incorrect aspects. Finally the model employs an approach which uses explicit aspects and opinion words to identify implicit aspects. Utilizing extracted polarity lexicon, the approach maps each opinion word in the lexicon to the set of pre-extracted explicit aspects with a co-occurrence metric. The proposed model was evaluated on a collection of English product review datasets. The model does not require any labeled training data and it can be easily applied to other languages or other domains such as movie reviews. Experimental results show considerable improvements of our model over conventional techniques including unsupervised and supervised approaches.

Introduction

With the rapid growth of user-generated content on the internet, the number of customer reviews that a product or service receives grows rapidly. A significant number of websites, blogs and forums (e.g., www.amazon.com, rottentomatoes.com, epinions.com) allow customers to post opinions about a variety of products or services. This online word of mouth behavior introduces a new and important source of information for business intelligence and marketing. In the other words customer reviews are essential to other potential customers, retailers and product manufacturers (potential users) in their efforts to understand the general opinions of customers and help them to make better decisions. As the number of customer reviews expands, it becomes very hard for users to obtain a comprehensive view of opinions of previous customers about various aspects of products through a manual analysis. Consequently proper analysis and summarization of customer reviews can further enable potential users to visualize previous positive and negative opinions about specific features or aspects of products. Therefore it is highly desirable to produce an automatic analysis or summary of customer reviews.

For the past few years, sentiment analysis (or opinion mining) for online customer reviews has attracted a great deal of attentions from researchers of data mining and natural language processing [1], [3], [5], [7], [8], [11], [9], [24], [25], [27], [33].

Sentiment analysis is a type of text analysis under the broad area of text mining and computational intelligence. Three fundamental problems in sentiment analysis are: aspect detection, opinion word detection and sentiment orientation identification [24], [27], [33].

Aspects are topics on which opinions are expressed. In the field of sentiment analysis, other names for aspect are: features, product features or opinion targets [3], [5], [7], [8], [6], [12], [24], [27], [33]. Aspects are important because without knowing them, the opinions expressed in a sentence or a review are of limited use. For example, in the review sentence “after using iPod, I found the size to be perfect for carrying in a pocket”, “size” is the aspect for which an opinion is expressed. Likewise aspect detection is critical to sentiment analysis, because its effectiveness dramatically affects the performance of opinion word detection and sentiment orientation identification. Therefore, in this study we concentrate on aspect detection for sentiment analysis.

Existing aspect detection methods can broadly be classified into two major approaches: supervised and unsupervised. Supervised aspect detection approaches require a set of pre-labeled training data. Although the supervised approaches can achieve reasonable effectiveness, building sufficient labeled data is often expensive and needs much human labor. Since unlabeled data are generally publicly available, it is desirable to develop models that work with unlabeled data. Additionally, due to variety and wide range of products and services being reviewed on the internet, supervised, domain-specific or language-dependent models are often hard to apply. Therefore we conclude the framework for the aspect detection must be robust and easily transferable between domains or languages.

In this paper, we present a novel unsupervised model which addresses the core tasks necessary to detect explicit and implicit aspects from review sentences in a sentiment analysis system. Our model differs from existing techniques in that it requires no labeled training data or additional information, not even for the initial seed information. Therefore the model can easily be transferred between domains or languages. The proposed model is based on the observation that there is inter-relation information between the aspects in reviews. Inter-relation information is the probability of the co-occurrence of two aspects in a review. Therefore the model explores review dataset by using both frequency-based and inter-relation information to find the aspects. Furthermore we have found that opinion words and aspects themselves have relations in opinionated sentences. Finally the model uses explicit extracted aspects and opinion words to detect implicit aspects.

In the remainder of this paper, Section 2 gives a definition of the aspect-level sentiment analysis, detailed discussions of existing works on aspect detection will be given in Section 3. Section 4 describes the proposed aspect detection model for sentiment analysis, including the overall process and specific aspects of the design of the workflow. Subsequently we describe our empirical evaluation and discuss the major experimental results in Section 5. Finally we conclude with a summary and some future research directions in Section 6.