Generating Recommendation Status of Electronic Products from Online Reviews ()
1. Introduction
The rapid growth of the web has led to rapid expansion of e-commerce among other things. More customers are turning towards online shopping because it is convenient, reliable, and fast. In order to enhance customer shopping experience, it has become a common practice for online merchants to enable their customers to write reviews on products that they have purchased. Customer reviews of a product are generally considered more honest, unbiased and comprehensive than descriptions provided by the seller. In fact, review comments are one of the most powerful and expressive sources of user preferences. Furthermore, reviews written by other customers describe the usage experience and perspective of (non-expert) customers with similar needs. They give customers a voice, increase consumer confidence, enhance product visibility, and can dramatically increase sales [1,2].
With more and more users becoming comfortable with the Web, an increasing number of people are writing reviews. As a result, the number of reviews that a product receives grows rapidly [3]. Moreover, the consumer reviews are in free form text and consumers prefer to use natural language to express their opinion. It is difficult for a program to “understand” the text information and use these data. Many reviews are long, and as such, it is not an easy task for a potential customer to make a decision whether or not to purchase a product based on the reviews he reads. It is therefore a very important and challenging problem to mine these reviews and produce a summary of them and also propose a recommendation decision to the potential customer.
A host of research works have been proposed to profer solution to problems related to mining and summarizing customer reviews [4,5], which is called opinion mining or sentiment analysis. Opinion Mining has two main research directions, document level opinion mining and feature level opinion mining [6]. Document level mining involves determining the document’s polarity by calculating the average semantic orientation of extracted phrases. In feature level opinion mining, reviews are summarized and classified by extracting high frequency feature and opinion keywords. Feature-opinion pairs are identified by using a statistical approach or labeling approach and dependency grammar rules to handle different kinds of sentence structures. Generally, feature level opinion mining has greater precision over the document level and uses basically product features in analysis and evaluation. The main tasks present in most past and current research works are: to find product features that have been commented on by reviewers [7] and to decide whether the comments are positive, negative or neutral [3]. Nevertheless, it is important to discover how positive or negative the comments are and further, make a concise decision about the product (recommended or not) to the potential customer. This work makes a contribution by classifying the opinions about each feature, hence showing how negative or positive it is. It also proposes a technique to include comments on infrequent features into the recommendation decision for electronic products.
The remainder of this paper is organized as follows. Section 2 contains a summary of related works. We describe the proposed technique in Section 3. Section 4 contains the experiments and evaluation, while Section 5 concludes the work and presents proposed future research work.
2. Related Work
Opinion mining has been studied by many researchers in past years. The earliest research on opinion mining was on identifying opinion (or sentiment) bearing words. Hatzivassiloglou and McKeown [8] identified several linguistic rules that can be exploited to identify opinion words and their orientations from a large corpus. The work was applied, extended and improved in [9].
Another major development in the area of opinion mining is sentiment classification of product reviews at the document level [10]. The objective of this task is to classify each review document as expressing a positive or a negative sentiment about an object (such as a movie, camera, car).
Feature-based opinion mining and summarization have been proposed by a number of researchers. Hu & Liu [7] proposed a technique based on association rule mining to extract product features. The main idea is that people often use the same words when they comment on the same product features. Then frequent itemsets of nouns in reviews are likely to be product features while the infrequent ones are less likely to be product features. This work also introduced the idea of using opinion words to find additional (often infrequent) features. Liu et al. [9] improved upon Hu’s work by proposing a technique based on language pattern mining to identify product features from pros and cons in reviews in the form of short sentences. They also made an effort to extract implicit features. Moreover, [11] proposed feature extraction for capturing knowledge from product reviews. The output of Hu and Liu’s system served as input to their system, and the input was mapped to the user-defined taxonomy features hierarchy thereby eliminating redundancy and providing conceptual organization. To identify the expressions of opinions associated with features, Hu & Liu focused on adjacent adjectives that modify feature nouns or noun phrases. They used adjacent adjectives as opinion words that are associated with features.
Popescu and Etzioni [12] investigated the same problem. Their algorithm requires that the product class is known. The algorithm determines whether a noun/noun phrase is a feature by computing the pointwise mutual information (PMI) score between the phrase and class specific discrimination. This work first used part-whole patterns for feature mining, but it finds part-whole based features by searching the Web and querying the Web is time-consuming. Qiu et al. [13] proposed a double propagation method, which exploits certain syntactic relations of opinion words and features, and propagates through both opinion words and features iteratively. The extraction rules are designed based on different relations between opinion words and features, and among opinion words and features themselves. Dependency grammar was adopted to describe these relations.
Another related work, but using a bit different approach is that of [5]. They proposed a method to extract product features from user reviews and generate a review summary, using product specifications rather than resources like segmenter, POS tagger or parser. At the feature extraction stage, multiple specifications are clustered to extend the vocabulary of product features. Hierarchy structure information and unit of measurement information are mined from the specification to improve the accuracy of feature extraction. At summary generation stage, hierarchy information in specifications is used to provide a natural conceptual view of product features.
Another area is mining of comparative sentences. It basically consists of identifying what features and objects are compared and which objects are preferred by their authors (opinion holders) [14]. Another related work involving comparison is that of [15] that compare reviews of different products in one category to find the reputition of the target product. However, it does not summarize reviews, and it does not mine product features on which the reviewers have expressed their opinions. Although they do find some frequent phrases indicating reputations, these phrases may not be product features (for example, “doesn’t work”, “benchmark result” and “no problem(s)”).
Feng et al. [16] and Ojokoh and Kayode [17] also proposed relateds method to solve the problem of opinion mining and applied this for some electronic products from amazon. There are a few differences between the works and ours. The method for determing the opinion polarity and aggregation is different from ours. Moreover, [16] only stopped at classifying the aggregated opinion as negative or positive and adopted that for recommendation. Also, their work did not consider infrequent feature identification. This works proposes the adoption of a threshold value for recommendation based on experiments with amazon rated views.
3. The Proposed Method
Figure 1 gives the architectural overview of the proposed system. In summary, the system executes the following steps: