Evaluating the Impact of Feature Selection on Overall Performance of Sentiment Analysis

Now a days the importance of analyzing the hidden sentiments from user reviews playing a prominent role towards increasing profitability in any organization. To address the challenges being faced in analyzing the text information and transforming the same in to polarities values with an objective of saving time in understanding the public opinion on particular product or service. Traditionally, there are different approaches carried out in transforming text data in to values based on different features of Text. In our research we make use of Stanford CoreNLP, Alias-i's Lingpipe (uses Logistic regression for document classification), Senti WordNet and synthesize libraries from different sources to include several other techniques that are used for text mining to evaluate the impact of feature selection on overall sentiment analysis by scoring a sentences in a review using different scoring Techniques. we also included NTU Lib Linear to make use of linear SVM for document classification. The Features considered on our experiments are Term Frequency and N-Gram (1Gram & 2Gram) with Decision Tree as Prediction model to evaluate the Accuracy, Area under ROC Curve and Kappa value. Finally, Compared the polarities of the reviews obtained using three different sentiment scoring approaches. The findings in our research is, Term Frequency have good impact of (0.932) on classifying the sentiment, In contrast, 2Gram have an impact of (0.8505).