Türkçe metinlerde duygu analizi için nitelik seçimi

Sentiment analysis is the classification of sentiments expressed in review documents. Like other classification tasks, it involves data preprocessing, feature selection, and classification steps. One aim of this study is to determine which preprocessing combinations and feature selection methods are effective for the sentiment analysis of Turkish reviews. Another aim is to propose a new feature selection method that helps identify the most valuable features for sentiment analysis. We consider several major feature selection methods, including Chi-square, Information Gain, Document Frequency Difference, and Optimal Orthogonal Centroid so that we can improve both the accuracy and efficiency of the sentiment analysis process and compare the performance of our new proposal. Experiments are conducted using four commonly used classifiers: Naïve Bayes Multinomial, Support Vector Machines, Logistic Regression, and Decision Trees. We find that keeping certain punctuation marks and stop words is helpful for Turkish reviews, and using feature selection methods of Chi-square, Information Gain, and Document Frequency Difference with Naïve Bayes Multinomial classifier tends to give us better results. Our proposed method achieves better classification performance with respect to the other methods. We further consider four common term weighting methods and investigate their effects on the sentiment analysis. We also try these weighting methods with different feature selection methods and examine how these term weighting methods respond to the reduced text representation. Finally, similar experiments are conducted on English reviews in order to compare their differences with Turkish reviews.