分析顧客評論之情景及其情緒分析之研究

Abstract

In this “Big Data Era”, mining valuable information from customer reviews can benefit both companies and customers. Sentiment analysis on customer reviews remains to be challenging due to the variety of human language usages. This thesis proposes a new concept of context in sentiment analysis by exploring real data and referencing works in linguistics. In spite of the popular definition of context which refers to the surroundings of a word and is often used in word sense disambiguation, our context refers to a fragment of texts which only contain ambiguous opinion words or even no opinion words. To conduct the sentiment analysis on customer reviews, a new Chinese customer review dataset in the hotel domain has is on both the snippet level and the clause level. For clause segmentation, two strategies, i.e., punctuation-based method and parsing-based method, have been proposed, and the polarity shift caused by discourse markers and negation operators is also considered in the parsing-based method. To extract the context, two types of the opinion word lexicons are first built and then checked and filtered manually. Thus, four types of lexicons are constructed. We then apply these four lexicons to extracting context on both the snippet level and the context level. In the sentiment classification experiments, Support Vector Machine (SVM) model with bag-of-words features and Convolutional Neuron Network (CNN) model with word embedding vectors are employed. The experimental results show that CNN model can always achieve the highest performance on both levels. Besides, the results also indicate that considering context in the process of sentiment analysis on customer reviews is necessary. In addition, detail error analysis and case study are also performed to understand the experimental results.