Research On Chinese Emotional Polarity Classification Method Based On Machine Learning

With the rapid development of Web 2.0, people left a large number of user-generated content in networks, therefore the study of opinion mining became a hot issue in nature language processing nowadays. As a key of opinion mining, the task of sentiment polarity classification is determining users' emotion is positive or negative. In this paper, in order to enhance the performance of sentiment analysis, we focus on the expansion of sentiment lexicon, paraphrase generation of opinion text and the choosing, representation of sentiment polarity classification features in the machine learning framework. Specifically, this paper launches the research from the following three aspects:1. Expansion of domain-specific sentiment lexicon. Owning to the widely using of dynamic polarity evaluation, we chose the attribute-evaluation pairs as the entries of lexicon. Firstly, we construct a seed lexicon according to attributes, evaluations and polarity in labeled text. Later, we extract opinion elements in the un-labeled text through supervised method, match the attributes, evaluations with co-occurrence frequency and distance between words. Finally, we make use of improved Polarity Rank algorithm to predict polarity of entries. The sentiment analysis results verify the applicability of domain-specific sentiment lexicon.2. Paraphrase generation based on standards assessment. Given the data sparseness of labeled opinion text, we propose to ease sparse data by opinion paraphrase generation. We firstly generate large amount paraphrase candidates by replacing phrases and adjust words order. Then, we set assessments mechanism from different angles to filter the candidates. In experiments, we generate paraphrase for the train data and test data, the results show that our method is better than the method of baseline and prove the effectiveness of our candidate generation, filtering strategies.3. Feature selection and representation for sentiment classification. The selection and representation of feature is always a hot issue in task of classification. In this paper we make comparison of a variety of feature selection methods in words or phrases features and research feature representation methods. We make a summary via sentiment classification experiments of many combinations of feature types, feature representations and classification frameworks. Finally, we get the most suitable combination of n-gram feature and representation for opinion comments in sentiment polarity classification.