Research On Sentiment Orientation Analysis Of Chinese Product

Nowadays the e-business is sweeping the world, how to quickly and ac-curately collect the information on product reviews on the Internet has great significance both to merchants'mastering consumer preferences and potential users'acquiring product trends. It's obvious inefficient and expensive if sim-ply collecting, analyzing and organizing these review information by manual labor. Text semantic orientation analysis, which is capable of processing them automatically and effectively has become one research hot spot with high ap-plication value in the Natural language processing area.Sentiment orientation analysis of Chinese product reviews, which involves 3 key technologies:Chinese sentence classification, extraction of product fea-tures with sentiment words and polarity analysis of sentiment words, are studied in this paper.(1) Chinese sentence classificationWe propose a cross-lingual sentiment classification method using adaptive algorithm and multiple classifiers integration framework.Firstly, utilize several translation engine services to translate English train-ing corpora into several Chinese training corpora, as to eliminate the gap be-tween English training corpora and Chinese testing corpora. Secondly, learn several classifiers with adaptive algorithm to improve the accuracy of classi-fiers. Lastly, integrate these classifiers to improve the stability of classification results.We choose NTCIR English corpora as training corpora, NTCIR Chinese corpora as testing corpora in the experiment, the F-measure of sentence opinion analysis achieves 65.55%, close to the preset upper limit 67.35%.(2) Extraction of product features with sentiment wordsWe propose a product features and sentiment words extraction algorithm based on POS(part of speech) dependency templates. Meanwhile we annotate a corpus based on our own annotation specification to test this algorithm's va-lidity.Firstly, extract possible POS dependency templates from training corpora with sequence mining algorithm. Secondly, utilize template with high confi-dence to extract product features with sentiment words from testing corpora. We choose 4 areas to experiment, the average precision achieves 54.18% and the recall is 29.45%.(3) Polarity analysis of sentiment wordsWe propose a method which is based on PageRank algorithm for polar-ity analysis of sentiment words. Implement the polarity analysis of sentiment words by using polarity of seed sentiment words and pseudo polarity of senti-ment words. The experiment results indicate that the precision of word's polar-ity analysis could achieve above 90%.Finally, we design and implement a product reviews system for automotive area.