Unsupervised Classification of Sentiment and Objectivity in Chinese Text

Unsupervised Classification of Sentiment and Objectivity in Chinese Text Taras Zagibalov John Carroll University of Sussex Department of Informatics Brighton BN1 9QH, UK {T.Zagibalov,J.A.Carroll}@sussex.ac.uk Abstract We address the problem of sentiment and objectivity classification of product re- views in Chinese. Our approach is distinct- ive in that it treats both positive / negative sentiment and subjectivity / objectivity not as distinct classes but rather as a con- tinuum; we argue that this is desirable from the perspective of would-be customers who read the reviews. We use novel unsuper- vised techniques, including a one-word 'seed' vocabulary and iterative retraining for sentiment processing, and a criterion of 'sentiment density' for determining the ex- tent to which a document is opinionated. The classifier achieves up to 87% F-meas- ure for sentiment polarity detection. 1 Introduction Automatic classification of sentiment has been a focus of a number of recent research efforts (e.g. (Turney, 2002; Pang et al., 2002; Dave at al., 2003). An important potential application of such work is in business intelligence: brands and com- pany image are valuable property, so organizations want to know how they are viewed by the media (what the 'spin' is on news stories, and editorials), business analysts (as expressed in stock market re- ports), customers (for example on product review sites) and their own employees. Another important application is to help people find out others' views about products they have purchased (e.g. consumer electronics), services and entertainment (e.g. movies), stocks and shares (from investor bulletin boards), and so on. In the work reported in this pa- per we focus on product reviews, with the intended users of the processing being would-be customers. Our approach is based on the insight that posi- tive and negative sentiments are extreme points in a continuum of sentiment, and that intermediate points in this continuum are of potential interest. For instance, in one scenario, someone might want to get an idea of the types of things people are say- ing about a particular product through reading a sample of reviews covering the spectrum from highly positive, through balanced, to highly nega- tive. (We call a review balanced if it is an opinion- ated text with an undecided or weak sentiment di- rection). In another scenario, a would-be customer might only be interested in reading balanced re- views, since they often present more reasoned ar- guments with fewer unsupported claims. Such a person might therefore want to avoid reviews such as Example (1) – written by a Chinese purchaser of a mobile phone (our English gloss). (1) 软件不行，发送短信时有时对方接收不 到；兼容性也不行，有的手机收到的短 信是乱码！还有死机现象！拍照效果次！ 不是循环或自定义式闹铃，每次都要调， 太麻烦了！后盖不够严密！原装配件中 无座充！ The software is bad, some sent SMS are nev- er received by the addressee; compatibility is also bad, on some mobile phones the re- ceived messages are in a scrambled encod- ing! And sometimes the phone 'dies'! Photos are horrible! It doesn't have a cyclic or pro- 304 grammable alarm-clock, you have to set it every time, how cumbersome! The back cov- er does not fit! The original software has …

Content truncated to 500 words