Abstract— Recent years have marked the beginning and rapid expansion of the social web, where people can freely express their opinion on different objects such as products, persons, topics etc on blogs, forums or e-commerce sites and opinion analysis is one emerging research field. As e-commerce is fast growing, product reviews on the Web have become an important information source for customers’ decision making when they plan to buy products online. Classifying the reviews automatically into different semantic orientations has become a major problem for customers as the reviews are too many for the customers to go through. In this paper we propose a different approach which performs the sentence level classification even the reviews contains mixed opinions. In this approach, a typical feature selection method based on sentence tagging is employed and a naive bayes classifier is used to create a base classification model, which is then combined with certain heuristic rules for review sentence classification. Experiments show that this approach achieves better results than using general naive bayes classifiers. Keywords— Sentence level classification, naive bayes classifier, sentence tagging
I. INTRODUCTION Web contains product reviews in variety of forms such as some particular sites dedicated to a specific type of product like sites for magazines and sites for movie reviews are different. Recent years have marked the strong influence of the “participative, social web” on the lives of both consumers and producer companies. This phenomenon encouraged the development of specialized sites, blogs, forums, as well as the inclusion of a review component in the already existing ecommerce sites, where people can write and read opinions and comments on their “objects” of interest – products, people, and topics. Product reviews, which contain customers’ feelings or opinions, available from many e-commerce web sites and professional product review portals, have become an important information source for a customer’s decision making when he/she plans to purchase products online. However, as the quantities of product reviews are often large, it is difficult for customers to read all of them before they make a decision. Sentiment classification, or semantic orientation classification is a way to automatically classify product reviews into two classes: recommended and not recommended, thus helping customers read them. This
classification approach is usually used to classify a customer’s review in a whole to determine its class. However, in reality, a customer often expresses mixed feelings in one review by pointing out some aspects are excellent but others are not so satisfactory. In this case, it is not reasonable to make an overall classification on it. In this paper, we propose an approach to determine customers’ semantic orientations in product reviews at a smaller granularity level (i.e. sentence level).
This sentence-level semantic classification (SLSC) approach employs a naive bayes (NB) classifier, which is used widely in text classification tasks, as its base classification model. It performs part-of-speech tagging to review sentences and uses certain types of words (adjectives, adverbs) as its features, which is different from common feature selection methods used in building naive bayes classification models. In the approach proposed, we concentrated on two main problems that had not been addressed so far by research in the field.
The first one was that of discovering the features that will be quantified. The second problem we addressed was that of quantifying the features in a product-dependent manner. During classification, heuristic rules are combined with the naive byes classifier to predict semantic orientations of review sentences. Experiments show that this approach performs better than the naive bayes classifier without the enhancement of the typical feature selection method and the heuristic rules. II. RELATED WORKS Sentiment classification has been initially studied as a cognitive linguistic problem. Early study as the work of Hearst [5] proposes a metaphoric model to determine the directionality of texts.
This approach requires a manuallyconstructed lexicon to derive such directionality information. Recently, machine learning methods and information retrieval techniques have been employed to address this problem. Pang et al [6] investigate several supervised machine learning methods to semantically classify movie reviews. Turney [2] employs a specific unsupervised learning method for the review semantic orientation classification. The employed method relies on the computation of mutual information between review phrases and the words “excellent” and “poor”. Dave, et al [7] develops a method for automatically classifying positive and negative reviews and experiment several methods related to feature selections and scoring. The above works focus mainly on the sentiment classification of a whole document or review, whereas our study realizes that mixed feelings or opinions may exist in one review. Besides the machine learning and related methods, there are other various ways to classify sentiment.
Subasic and Huettner [8] use fuzzy logic to manually construct a lexicon, based on which fuzzy techniques applicable to fuzzy sets are used to analyze the affect of documents. Liu, et al [9] build linguistic affect models for six basic emotions by utilizing relationships from the Open Mind Common Sense (OMCS) knowledge base and manually specified ground truth. An affect sensing engine is then built to judge the affect of given passages. Hu and Liu [10] use the adjective synonym sets and antonym sets in WordNet [11] to judge semantic orientations of adjectives. They extend a seed set of adjectives by searching synonyms and antonyms in WordNet. The above methods either require certain amounts of manual constructions or rely on external structured information sources (i.e. OMCS, WordNet), which is what this study tries to reduce or avoid. The method we propose is language and customer-review independent. It extracts a set of general product features, finds product specific features and feature attributes and is thus applicable to all possible reviews in a product class.
We describe the steps performed to obtain the features for each product class and the manner in which input text is processed to obtain the opinion expressed by customers. III. THE SLSC APPROACH The following several subsections will discuss the SLSC approach in detail A. Data collection and pre-processing Most product reviews are posted on some e-commerce web sites or professional review web portals. Their formats vary from site to site. Some web sites provide a numeric rating mechanism. Besides this, some web sites also allow users to post reviews to freely express their feelings for the products.
Certain web sites even let users put their comments under different titles such as “Pros” / “Cons”. For example, CNet allows numeric rating, separated comments, and free comments. Web pages that may contain costumer reviews are crawled from those web sites. Matching rules is used to ensure high “harvest rate” of web pages containing real costumer reviews. For example, words like “review”, “user opinion” and etc. in the web pages are good indicators of pages containing costumer reviews. Special review content wrappers are designed for pre-processing web pages from particular web sites.
Figure1: Pre Processing As depicted in Figure 1, in our approach, we start from the following scenario: a user enters a query about a product that he/she is interested to buy. The search engine will retrieve a series of documents containing the product name, in different languages. Further on, two parallel operations are performed: the first one uses language identifier software to filter and obtain two categories – one containing the reviews in English and the other the reviews in Spanish. The second operation implies a modified version of the system for the classification of person names. We use this system in order to determine the category the product queried belongs to. B. Review sentence tagging Not all the words in review sentences are useful for classifying semantic orientations or related tasks. As Hu and Liu [10] point out, nouns and noun phrases in the sentences are likely to be the features that customers comment on, while adjectives are often used to express opinions and feelings.
The following review sentences are typical examples from reviews of a digital camera: “The picture quality is nice and the pictures are huge”; “The buttons are very simple and nice”. Provided this phenomenon, we use part-of-speech (POS) tagging to distinguish adjectives and adverbs in the sentences as candidate features that indicate semantic orientations. The following is tagged examples by applying Stanford POS tagger [12], which is used in this study: “The/DT picture/NN quality/NN is/VBZ nice/JJ and/CC the/DT pictures/NNS are/VBP huge./JJ”. The slashes and the followed up case letters in the sentence indicates the types of words before them. For example, “nice” is tagged with “/JJ”, indicating it’s an adjective in the sentence and it should be noted that POS taggers are not always perfect and review sentences may be quite complex or irregular. Therefore, tagging errors cannot be avoided. In order to perform review sentence tagging first we need to do the word classification.
Trying to understand attributes of a subjective element such as whether it is positive or negative (polarity or semantic orientation) or has different intensities (gradability) is even more difficult. Hatzivassiloglou and McKeown [5] used textual conjunctions such as “fair and legitimate” or “simplistic but well-received” to separate similarly and oppositely connoted words. Other studies showed that restricting features used for classification to those adjectives that come through as strongly dynamic, gradable, or oriented improved performance in the genre-classification task. The next step is sentiment classification. Feature selection: tokenizin g, initial states, thresholdi ng, language processin g, substituti on, collocatio n Score words from stats smoothi ng
2) Sentiment classification: Using fuzzy logic was one interesting approach to classifying sentiment. Subasic and Huettner [20] manually constructed a lexicon associating words with affect categories, specifying an intensity and centrality (degree of relatedness to the category).
For example, “mayhem” would belong, among others, to the category violence with certain levels of intensity and centrality. Fuzzy sets are then used to classify documents. Another technique uses a manually-constructed lexicon to derive global directionality information (e.g. “Is the agent in favor of, neutral or opposed to the event?”) Which converts linguistic pieces into roles in a metaphoric model of motion, with labels like BLOCK or ENABLE [7]? Recently, Liu et al. [10] used relationships from the Open Mind Commonsense database and manually-specified ground truth to assign scalar affect values to linguistic features.
These corresponded to six basic emotions (happy, sad, anger, fear, disgust, surprise).
Several techniques were applied to classify passages using this knowledge, and user studies were conducted with an email composer that presented face icons corresponding to the inferred emotion. Recommendations: At a more applied level, Das and Chen used a classifier on investor bulletin boards to see if apparently positive postings were correlated with stock price. Several scoring methods were employed in conjunction with a manually crafted lexicon, but the best performance came from a combination of techniques. Another project, using Usenet as a corpus, managed to accurately determine when posters were recommending a URL in their message.
Recently, Pang et al. [15] attempted to classify movie reviews posted to Usenet, using accompanying numerical ratings as ground truth. A variety of features and learning methods were employed, but the best results came from unigrams in a presence-based frequency model run through a Support Vector Machine (SVM), with 82.9 percent accuracy. Limited tests on this corpus1 using our own classifier yielded only 80.6 percent accuracy using our baseline bigrams method. As we discuss comparisons are less clear-cut on our own corpus. We believe this is due to differences in the problems we are studying; among other things, the messages in our corpus are smaller and cover a different domain.
2012 model for those words is essential in classifying customer’s opinions. Several approaches have been proposed to build such a model. However, they either require certain amounts of manual constructions or rely on some external structured information sources. In this study, models of semantic orientations of words are built by a NB classifier. As a supervised learning method, NB classifier requires tagged/labeled examples for building the classification model. Fortunately, as some of reviews that have been posted on the web sites already have their separate titles. These titles can be treated as their labels. It is then not necessary to create labeled training examples manually. The words used to build the model are the candidate features extracted during the previous tagging process.
The probability of word wt for different semantic orientations is estimated by the following formula by naive bayes can yield reasonably good results for determining the semantic orientations of words. D. Classification By formula (2), we can predict the probability of each semantic orientation for a given review sentence. Therefore, the given review sentence can be assigned with the semantic orientation that has the highest semantic orientation. As in this case, only two semantic orientations are available, the following derived scoring formula can be used
where N(wt, dk) is the count of the number of times word wt occurs in sentence dk, V is the vocabulary or the feature word set for this
classification task, Dt is the labeled sentence set, ci ∈{c+, c-} denoting positive or negative orientations, and Pr(ci|dk) ∈{0, 1} as given by the label for the sentence dk. After Pr (wt|ci) is estimated from the labeled examples, the probability of semantic orientations for a given review sentence dj can be simply predicted by the following formula
Where Pr (ci) is the prior probability of the semantic orientation ci. The detailed classification process will be discussed in the next section. Positive orientation easy 6 quick great 7 nice excellent 8 fast good 9 perfect manual 10 compact Negative orientation 1 heavy 6 hard 2 blurry 7 expensive 3 poor 8 plastic 4 larger 9 slow 5 large 10 optical
Table2: Most predictive Words Table 2 shows the most predictive words for different orientations after training on labeled reviews sentences from 3 digital cameras. Each column presents the ordered list of words that the model indicates are the most “predictive” for that orientation. Here “predictive” is judged by a weighted log likelihood ratio, which is calculated through rainbow, a text classification toolkit. This table indicates that the model built
During the implementation, we do not use the prior probabilities of semantic orientation (Pr(c+), Pr(c-)) estimated from the labeled data and assume they are equal, as the words, if available in the sentence, are indicative enough to determine the semantic orientation. Therefore, a review sentence with a positive score calculated by formula (3) indicates it is classified to have a positive semantic orientation, and vice versa. However, in practice, the performance of naive bayes classifier may be degraded by two situations. One situation is that sometimes customers tend to use words like “not”, “should” to express opposite feelings that the following words indicate. For example, sentence like “Auto modes are not that great, so stay away point and shooters” may be misclassified if only word “great” is used. Another situation is that review sentences may not always contain feature words stored in the naive bayes model, especially when the labeled examples are small. Therefore, some undetermined review sentences may exist due to the failure of naive bayes classification.
We combine heuristic rules with naive bayes classifier to reduce the effects caused by the above situations. For the situation involving negative words, if a feature word is detected closely (e.g. within 3 word length) after a negative word, its probabilities for the different semantic orientations are to be swapped. This rule can also be applied when building naive bayes model with labeled examples… For the situation when naive bayes classification fails with no sign indicating semantic orientations, context information of review sentences is used for prediction. As have been observed, customers usually express their feelings consistently with several consecutive sentences, as long as no negative transition words (e.g. “but”, “however”, and etc) between these sentences.
Therefore, the classification process for a costumer review can be illustrated by Figure 2 by considering the context information. Though this may still not work when the initial naive bayes classification on the whole review sentences fails to provide adequate information, it increases the classification accuracy, which is demonstrated in our experiments. Input: Review from a customer Output: semantic orientations of each sentence begin split review into sentences forming a list S with original order; for each sentence s in S s.nb=score(s); while exist s with s.nb=0 begin select si whose neighbor sn has the higher |sn.nb|; if negative transition exists between si and sn si.nb= -sn.nb; else si.nb= sn.nb; end end 0.8 0.7 0.6 0.5 0.4 0.3 NB-ALL NB-FS
Figure3: Classification process by considering context information IV. EXPERIMENTS . To validate the effectiveness of the proposed approach, experiments are conducted on costumer reviews of seven products (three laptop computers and four digital cameras).
Web pages containing those reviews are crawled from the popular review web site – CNet and cleaned to extract review contents including review title, Pros/Cons , and full user opinion (if available).
As the selected products are fairly popular, a lot of costumers have posted their reviews for them. The quantity of collected reviews for each product ranges from 58 to 190. After pre-processing these reviews, we also manually checked them to eliminate some noises. For example, some customers put sentences like “None so far” under the “Cons” title, which means they actually don’t have any negative opinion. Such sentences should be removed as well as those irrelevant or advertisement sentences and the like. The experiments are conducted by the style of “leaveone-out”. For example, reviews of three digital cameras are used for building classification model and the reviews of the remaining camera are used for test with this model.
Figure 4: Average classification accuracy for Different feature selection methods Figure 4 provides the comparison of classification accuracies generated by our customized naive bayes using selected features (indicated as “NB-FS” in the figure) and the general naive bayes using all the features (“NB-All”).
“DCAVG” and “COMAVG” means that the results are averaged from tested digital cameras and laptop computers respectively. This demonstrates the effectiveness of the typical feature selection method. Using all the features in the review has smaller classification accuracy than using selected features (adjectives and adverbs) according to the experiment. Table 3 below shows the classification accuracy achieved by the naive bayes classifier with and without heuristic rules.
Table 4: Results of tests using naive bayes We obtain more consistent performance across tests with less computation when we use the various calculated frequencies and techniques from information retrieval. V. CONCLUSION
This paper proposes an approach for semantic classification of online product reviews. The approach performs semantic classification at the sentence level by realizing reviews often contain mixed feelings or opinions. A typical feature selection method based on sentence tagging is employed and a naive bayes classifier is used to create the base classification model. Combined with provided heuristic rules, the model yields better results than the general naive bayes classifier. This approach, which can be further improved by the future work, can be used to classify customer’s opinions into different semantic orientations at a smaller granularity level.
Extend tasks such as identifying frequent similar opinions can then be performed to provide more valuable information for customers decision making when they plan to purchase products. Future work includes the development of a method to extend the list of product-dependent features and feature attributes, alternate methodologies for polarity assignation to product dependent feature attributes and finally, the application of a textual entailment system to verify the quality of the feature extracted and the assigned polarity.