, pp.55-59 http://x.oi.org/10.14257/astl.2014.45.11 Dimension-Base Sentiment Polarity Detection for E-Commerce Reviews Yajun Deng, Lizhou Zheng, Peiquan Jin School of Computer Science an Technology, University of Science an Technology of China, 230027, Hefei, China Abstract. E-commerce reviews are very helpful for customers to know other people s opinions on intereste proucts. Meanwhile, proucers are able to learn the public on their proucts being sol in E-commerce platforms. Generally, E-commerce reviews involve many aspects of proucts, e.g., appearance, quality, price, logistics, an so on. In this paper, we efine each of those aspects as a imension of prouct, an present a imension-base analysis approach for E-commerce reviews. In particular, we employ a lexicon expansion mechanism to remove the wor ambiguity among ifferent imensions, an propose a rules an lexicon base algorithm for analysis on E-commerce reviews. We conuct experiments on a large-scale prouct reviews ataset involving over 28 million reviews, an compare our imension-base analysis approach with the traitional way that oes not consier imensions of reviews. The results show that the multi- approach outperforms the traitional way in terms of F-measure. Keywors: Sentiment; E-commerce reviews; Dimension 1 Introuction Online E-commerce reviews contain a lot of customers opinions on certain proucts, an are helpful for ecision making. A piece of E-commerce review generally involves many aspects of prouct, such as appearance, quality, price, logistics, an so on. Therefore, it is not appropriate to give only one positive or negative jugment on an E-commerce review by aopting traitional ocument-base or sentence-base analyzing approaches [1]. In general, a customer may be satisfie with some imensions of a prouct but islike some other features. For example, a review saying 宝贝外观不错, 质量很好, 就是太贵了!(Nice appearance, goo quality, but too expensive!) inicates that the customer likes the appearance an quality of the prouct but has a negative opinion on the price. Therefore, it is not suitable to conuct ocument-level or sentence-level analysis for E-commerce reviews. A better an more suitable way is to perform analysis on separate ISSN: 2287-1233 ASTL Copyright 2014 SERSC
imensions of proucts, i.e., to conuct a imension-base analysis on E-commerce reviews. The focus of imension-base analysis iffers from traitional aspect-oriente analysis. Most of existing aspect-oriente stuies concentrate on fining the explicit or implicit aspects from reviews [2,3]. However, in our work, the imensions are provie by E-commerce company, an we focus on the expansion of lexicons. This paper is aime at presenting a framework for imension-base analysis on E-commerce reviews. Our stuy is base on a large scale of E-commerce reviews ataset consisting of 28 million of reviews. The challenges in imension-base analysis lie in imensions mapping an wor isambiguation. The imension mapping problem refers to mapping opinione text blocks with right imensions. For instance, the text block Nice look! will be mappe with the imension appearance. The wor isambiguation problem refers that a wor may be connecte with two or more imensions. In this paper, we focus on those two problems an propose an effective framework to finally extract imension-oriente opinions from E-commerce reviews. In particular, we propose a new approach to removing the wor ambiguity by introucing a lexicon expansion mechanism to construct the imension-base lexicon. We conuct experiments on a real ataset incluing 28 million E-commerce reviews, an compare our algorithm with the traitional algorithm without consieration. The results show that our algorithm is superior to the competitor algorithm. 2 Our Approach Figure 1 shows the general framework of the imension-base polarity etection for E-commerce reviews. It consists of five moules, as shown in Fig.1. The Text Pre-Processing an Sentence Segmentation steps are very similar with existing works, so we will not iscuss them in etail. E-commerce Reviews Textual Pre-Processing sentences Sentence Segmentation Dimension Mapping text blocks <text block, imension> mapping Dimensional Sentiment Lexicon Expansion expane wors Dimension Sentiment Lexicon wors Sentiment Analysis polarity Fig. 1. The framework of our approach Dimension Mapping. Dimensions are typically omain-epenent. Since a imension may appear in many prouct categories, we have to etermine the right 56 Copyright 2014 SERSC
imension for a given review about a certain prouct category. Given a review sentence s, we can obtain its category c an the imension set D c of category c irectly. For each wor w i in the review sentence s an a imension j D c, along with j s initial keywor set KW j, we assign a probability score wprob i,j which escribe the probability of wor w i belongs to category j (as shown in Formula 2.1). Here freq(w i, kw k, j ) means the frequency of wor w in short sentence s appearing together with key wor kw k in imension j. w p ro b freq ( w, ) * freq ( w, kw, ) i k j (2.1) i, j i j freq ( w, kw, ) kw k K W i k j Then, the probability of a short sentence s belonging to a imension j, sprob s,j, is calculate as the prouct of the probability of all wors in sentence s belonging to imension j (as shown in Formula 2.2). sp ro b w p ro b (2.2) s, j i, j wi s Finally we map a review sentence s with the imension j that has the maximum probability score with s. Dimensional Sentiment Lexicon Expansion. Sentiment lexicons, such as HowNet Lexicon an NTUSD Lexicon, consist of the polarity of many popular wors an are wiely use in analysis [4,5]. However, those lexicons i not istinguish the polarities of one wor on ifferent imensions about a prouct. For example, " 硬 /har, soli" is a positive wor when escribing builing materials but is a negative wor when escribing foo. Besies, they only cover a small part of wors that are not enough for the analysis on E-commerce reviews. We use a lexicon expansion metho to solve the problems mentione above. For each imension, we perform two steps to extract its corresponing lexicon: (1) Step 1: see-base lexicon expansion. First, we use some see wors to enlarge the lexicon. This see-wor-base proceure starts with some manually collecte see wors which coul be of the same polarity uner ifferent imensions (e.g., positive wors like 好 /goo an 不错 /nice, negative wors like 糟糕 /terrible an 惨不忍睹 /too horrible to look at, etc.). Base on the co-occurrence information with see wors, we coul enlarge the original see wors iteratively, in each roun of iteration we put top-3 wors with the highest co-occurrence score into the see wor set. The proceure ens when the size of see wor set is larger than 20 or the co-occurrence score of each Top-K wors are below a threshol. (2) Step2: filtering. Another proceure of our metho is to filter wors of high frequency. The filtering step is base on the the DF*ICF score (similar to the classical TF*IDF measurement, here DF means imension frequency, an ICF means inverse corpus frequency) uner a specific imension for each wor in. We efine the DF*ICF score of a wor w i uner in Formula 2.3. D sen ten ce sen ten ce D F * IC F * lo g i, sen ten ce sen ten ce i, co rp u s (2.3) i, co rp u s Copyright 2014 SERSC 57
Here sentence i, is the count of sentence contains wor w i in imension, an sentence is the count of sentence uner imension. sentence corpus is the count of sentence uner the whole corpus, an sentence i,corpus is the count of sentence contains wor w i in the whole corpus. For each imension, we put the Top-20 wors with highest DF*ICF scores into the caniate wor set. Sentiment Analysis. Our analysis algorithm on prouct reviews is base on lexicon an rules. The ifference of our work compare to previous works is that we combine both imension lexicon an public lexicon together. Besies, we use multiple rules to improve the performance of analysis. Particularly, we first efine some strong rules, which can be applie to recognize the polarity with high precision. For example, 与 想象的一样 (it s the same as imagine) is recognize as a positive strong rule, which infers that sentences matching with this rule will be labelle as positive without consiering other factors. Next, we calculate the linear weighte sum of all the positive, negative, an neutral polarities of the wors appearing in each short sentence s. This sum is finally outputte as the polarity of the sentence. As a result, the polarity of our algorithm will output three ifferent scores, namely the positive score, negative score, an neutral score. 3 Experimental Results Our experiment is base on a real ataset containing 28 million E-commerce prouct reviews. Each of the review belongs to a prouct category, which has been provie by the E-commerce company (totally 8000+ categories). For each category, there are multiple imensions which escribe ifferent aspects of the proucts in that category. For each imension, there are several keywors which are regare as see wors in our algorithm. F-measure 100.00% 80.00% 60.00% 40.00% 20.00% 0.00% With lexicon Without lexicon Fig.2 F-measure of our metho an the baseline algorithm We ranomly choose 9,000 outputte results to show the performance of our metho. We compare our metho with the approach which oes not contain the wor lexicon (only utilize the HowNet an NTUSD 58 Copyright 2014 SERSC
lexicon), an the results show that our algorithm can improve the precision, recall, an F-measure for each type of polarities (positive, negative an neutral). Due to the space limitation, we only give the F-measure comparison in Fig.2. The positive case, negative case, an the neutral case in Fig.2 represent the positive scores, negative scores, an the neutral scores outputte by the analysis step in Fig.1. To conclue, in this work, we propose to expan an construct lexicons for imension-base analysis on E-commerce reviews. Our preliminary results show that this approach is helpful to improve the performance of analysis on E-commerce reviews. 5 Conclusion In this paper, we propose a framework for analysis over E-commerce reviews. The major feature of our approach is that we introuce a ictionary into analysis, an the experimental results on a real ata set showe that the propose metho is effective in polarity etection for E-commerce. Acknowlegement. This paper is supporte by the National Science Founation of China (No. 71273010), an the National Science Founation of Anhui Province (no. 1208085MG117). References 1. R. Felman. Techniques an applications for analysis. Communications of the ACM, 2013, 56(4): 82-89. 2. A. Popescu, O. Etzioni, Extracting prouct features an opinions from reviews, in Proc. Of EMNLP, 2005 3. B. Liu, Sentiment analysis an subjectivity. in Hanbook of natural language processing, 2n eition, CRC press, 2010. 4. Q. Zhang, P. Jin, L. Yue, Extracting Focuse Locations for Web Pages, First International Workshop on Web-base Geographic Information Management (WGIM) (in conjunction with WAIM'11), LNCS 7142, Springer, 2011, pp.76-89 5. P. Jin, X. Zhang, Q. Zhang, S. Lin, L. Yue, Ranking Web Pages by Associating Keywors with Locations, In Proc. Of WAIM, LNCS 7923, Springer, Beiaihe, China, 2013 Copyright 2014 SERSC 59