1) Download online product reviews;

2) Categorize review(s) into segments based on product features;

3) Classify review segments based on polarity;

4) Score the obtained polarized review segments;

5) Incorporate infrequent feature segments;

6) Make recommendations to potential customers in an online shop.

Given a product class, C with instances I and reviews R, the goal of the system is to find the set of (feature, opinions) tuples s.t ƒ ε F and o₁,∙∙∙, o_j ε O, where:

F is a set of product class features in R;

O is a set of opinion phrases in R;

ƒ is a feature of a particular product instance;

o is an opinion about ƒ in a particular sentence.

These steps as reflected in the architectural overview are described in details as follows.

3.1. Preprocessing

The HTML files containing online reviews downloaded from a product review website [18] were analyzed and parsed to extract useful information and review(s) about every product instance present. Java HTML Parser was used for analysis and extraction of the reviews. Java HTML Parser is a java library used to parse HTML files in either a linear or nested fashion, and it is primarily used for extraction. In this process, the HTML tags in the review documents are removed.

The needed review(s) are then extracted and stored in the database. A list of product features for a particular product class is maintained in the database to enable categorization of the extracted reviews into feature segments.

Each review is then disintegrated into simple sentence segments using punctuations such as “, : ; .”. The segmented reviews are analysed and categorized based on the feature set found in the feature database. Each segment category for every product instance is passed to the POS Tagger.

3.2. Part of Speech (POS) Tagging

Every word in each segment category is parsed and tagged by part of speech using Stanford parser. Stanford parser [19] is a piece of software that reads text in some languages and assigns parts of speech to each word (and other tokens), such as noun (NN), verb (VBZ), adjective (JJ), noun phrase (NP), verb phrase (VP), adjective phrase (ADJP), adverb (RB) etc and the root sentence (S).

To parse a sentence like “battery life is extremely poor”, the outcome is given in Figure 2.

The parts of speech we are interested in are nouns (NN) and adjectives (JJ), since they correspond to feature and opinion pairs that are to be extracted. However, some words could not be extracted independently; for example, considering tagging the sentence like:

“This product is not too bad”, where customer’s opinion is negated by the adverb “not”. Therefore, the idea proposed in [20] was employed. Two consecutive words are extracted from the review segments if their tags conform to any of the patterns in Table 1.

In Table 1, JJ tags indicate adjectives, the NN tags are nouns, the RB tags are adverbs, and the VB tags are verbs. The first pattern from the table, means that two consecutive words are extracted if the first word is an adjective and the second is a noun; the second pattern means that two consecutive words are extracted if the first word is an adverb and the second word is adjective, but the third word—which is not extracted—cannot be a noun; the third pattern means that two consecutive words are extracted if they are both adjectives but the following word is not noun. Singular and plural proper nouns are avoided, so that the names of items in the review cannot

Figure 2. Hierarchical tree output from Stanford parser.

Table 1. Feature-opinion extraction pattern.

influence the classification.

3.3. Opinion Words Extraction, Polarization and Aggregation

Opinion words are the words people use to express a positive or negative comment about a subject. In this work, a lexicon—based approach is used (based on previous work on the correlation of subjectivity and the presence of adjectives in sentences [21,22] where opinion words are assumed to be adjectives). The list, containing 2006 positively and 4783 negatively oriented adjectives, created by [7], is used for this purpose.

An aggregated score [3] is obtained for each segment. The summation of the scores obtained for segments in this group forms the grade for the feature. The features grade is used in turn to summarize the instance.

The opinions obtained are identified and classified based on the review phrase whose key word has a negative, positive or neutral semantic orientation. A value, feature-opinion word score, is assigned to the feature for the segment where feature and opinion word is found. The summation of this value determines the score for the feature of an instance.

3.4. Infrequent Feature Segments Aggregation

Segments that are referred to as “Infrequent feature segments” by our system are the ones with comments that are associated with silent features, or non-identifiable ones such as “This is an excellent laptop”. In this case, there are no specific features that such opinion words are associated with. We discover in this work that they still need some consideration rather than being considered as too significant. Hence, before a final recommendation decision is made for any product, our system proposes the inclusion of some values for the infrequent feature segments to form the final summation value that determines if a final recommendation will be made or not. Therefore, the aggregation function proposed by [3] is modified to accomodate infrequent feature segments as follows.

Given lists of positive, negative and context dependent opinion words, including phrases and idioms, the opinion aggregation function for frequent features is calculated using Equation (1).

(1)

where w_j is an opinion word, V is the set of all opinions in our list, s_k is the sentence segment that contains the feature f_i, and dis(w_j, f_i) is the distance between feature f_i and opinion word w_j in s_k∙w_j∙SO is the semantic orientation of w_i. If the final score is positive, then the opinion on the feature in the sentence s is positive. If the final score is negative, then the opinion on the feature is negative, otherwise it is considered to be neutral. The multiplicative inverse in the formula is used to give low weights to opinion words that are far away from feature f_i. A positive word is assigned the semantic orientation score of 1, and a negative word is assigned the semantic orientation score of −1. In Hu & Liu (2004), a simple summation is used. The new function is better because far away opinion words may not modify the current feature.

Further, a distinction is made between the positive and negative segments scores as shown in Equations (2) and (3):

(2)

represents positive feature score for a given segment s_k that contains the feature f_i. The sum of all the positive scores obtained from segment s_k for a particular feature is represented by and n is the total number of positive score segments.

(3)

represents negative feature score for a given segment s_k that contains the feature f_i. The sum of all the negative scores obtained from segment S_k for a particular feature is represented by, where n is the total number of negative score segments.

An aggregated score is obtained for every feature, f_i, as shown in Equation (4):

(4)

This work also proposes a percentage value to be used to determine the quality of product feature f_i from reviews S. The higher the positive percentage value the better the product feature, and the higher the negative percentage value, the lower the quality. These are described in Equations (5) and (6).

(5)

is the aggregated positive feature score in percentage

(6)

is the aggregated negative feature score in percentage.

This work proposes a method for the inclusion of infrequent feature segments, with some explanations. It was noticed that with many of the datasets, the infrequent feature segments have some effect on the results, especially when they are negative segments. With some experiments, it was also discovered, that as the number increases (if is positive) or decreases (if it is negative), the effect reduces. So, the technique proposed was to reduce the effect of the inclusion of these segments with increase in the number for negative segments and a threshold was set for positive segments.

The proposed technique is proposed in the algorithm in Table 2.

Table 2. Opinion aggregation for infrequent feature segments.

The algorithm proposed in [3] was modified to incurporate our idea of review classification and recommendation decision inclusion. It is described in the algorithm in Table 3.

This work also proposes a ranking system for the feature scores. This is reasonable, because many comments on some of the popular electronic commerce are rated. For instance, amazon [23] awards stars (1 to 5; the higher, the better) to reviews of different products. From this ranking, the features scores and remarks are obtained. The method used is shown in Table 4. So from the aggregated value obtained previously, for feature, a

Table 3. Predicting the orientations of opinion on product features.

Table 4. Feature score value and remark.

feature score value and corresponding remark are generated.

3.5. Product Instance Result and Summarization

The recommendation status is derived, based on the threshold (60%) fixed for recommendation. If the positive sentiment score is greater than or equal to the threshold, it means the authors recommend the product. Otherwise, it means the authors do not recommend this product. For instance, if the recommendation summary generated is:

Positive: 90% Negative: 10% The Recommendation Status is “Recommended”.

The threshold of 60% was arrived at after performing experiments with some rated downloaded reviews from amazon, in addition to the classification decision adopted by [16]. It is reasonable to assume that the 4 star and 5 star reviews should lead to a positive recommendation, while 1, 2 or 3 star should lead to negative recommendation (not recommended).

4. Experiments and Evaluation

Figures 3 and 4 show the screen shots of the developed online shopping mall, where our opinion mining application was implemented. Figure 3 is the home page, while

Figure 4 is the web page that provides the means to display the reviews of the different electronic products used in the system. The product classes used for the experiments are laptop, printer, scanner, DVD player and television set, with 59 product instances in all. Several user reviews downloaded from an online review website [18] were used for the experiments. The characteristics of these data are shown in Table 5.

Experiments were carried out with online reviews of different brands (instances) of DVD players, laptops, scanners, printers and television sets. Some sample experimental results are shown in Tables 6-9.

Experiments were carried out to finally arrive at the threshold of 60%. It was discovered with the trend of the reviews downloaded from amazon and from Feng et al. (2010)’s adoption of 4 and 5 star reviews of amazon as positive reviews and the others (1 to 3) as negative, that it is reasonable to set the threshold for recommendation as 60% (as our evaluation results affirm).

Evaluation

In order to evaluate the performance of the proposed system, 499 laptops rated reviews were downloaded from amazon to test the accuracy of the classification of the reviews and the final recommendation decision made by the system. Table 10 gives an analysis of the number of

Table 5. Characteristics of the review data.

the reviews used for the evaluation experiments in terms of the number of ratings by amazon. We adopted the method described above in order to rate our reviews into five ranks like amazon, and tried to evaluate how the system performs. The set of reviews that are eventually given the same rank as amazon does is termed “correctly ranked” and “not correctly ranked” otherwise. Table 11 shows the results of this evaluation. It can be seen from the results that incorporating infrequent feature segments results contributes significantly to the improvement of the accuracy of the system for review classification.

Table 12 shows the result of the final recommendation result, which is the actual goal of the proposed system. The evaluation conducted here follows the method adopted by [16]. For all the product instances, any 4 or 5 star review that receives a recommendation decision by our system is termed correct, while any 1, 2 or 3 star review that receives a recommendation decision is termed incorrect. The proposed system performs significantly better than some related systems in making recommendation decision. For instance, our proposed system makes 100% accuracy for three products out of five and 60% and 80% for the others without incorporating infrequent feature segments results. The recommendation process achieves 100% accuracy with the tested data sets for all the electronic products after implementing our proposed technique for incorporating infrequent feature segments. The accuracy of the ranking done by our system is shown in Table 13. It achieves 100% accuracy for ranks 1, 3 and 5. The accuracy for ranks 2 and 4 are relatively low. However, for the missing places in rank four, the system classifies most of the results under rank 5, which will lead to a correct recommendation decision at the end of the day. The low results for rank two are mostly for the Acer and Toshiba Sattelite datasets. The reasons are not clearly known. Many existing systems used their methods on clearly different datasets and different products, from the ones used in the system. However, according to the published results of one of the nearest systems [16] (shown in Table 14), for the same decision (but for different products), 85%, 84%, 70%, 78% accuracy were obtained for the four different products they evaluated on. Although the products are different, this still shows that our system is effective, especially at making a final recommendation decision.

Table 6. Product: DVD player instance name: Bush DVD2023.

Table 7. Product: Laptop instance name: Apple iBook G4.

Table 8. Product: Scanner instance name: HP 3200C.

Table 9. Product: Television set instance name: Sharp Aquos LC42XD1E LCD.

Table 10. Analysis of reviews used for evaluation experiments.

5. Conclusion and Future Work

In this paper, we propose a set of techniques for mining and classifying opinions expressed on the features of products as recommended or not recommended. It provides a method for scoring review segments containing “infrequent features”. It also proposes the inclusion of a rank or rating for each feature like other electronic com

Table 11. The evaluation result of review rating.

Table 12. The evaluation result of recommendation decision.

Table 13. The accuracy of the ranks/ratings.

Table 14. The evaluation result of review classification [16].

merce sites do. It finally aggregates all these to make a recommendation decision using an experimental threshold value for potential buyers of electronic products in an online shopping mall. The experimental results indicate that the proposed techniques are effective at performing this task. In the future, the work can be extended to test our proposed algorithm on other data sets, and also improve on the methods adopted in this work for automatic tag identification on the review web pages, to handle more complex sentences, especially those written in improperly structured English language.

REFERENCES

NOTES

^*Corresponding author.