Sentiment polarity identification using machine learning techniques

The paper proposes an improved approach to the problem of sentiment polarity identification. Its main focus is on identifying and extracting the relevant information from natural language texts in order to obtain a set of best predictive features to be used for the classification task. Our approach of determining the polarity of a text consists of a combination of several processing techniques that obtains an efficient set of appropriate information for the underlying text. Among techniques, we have considered pruning the feature set to discard features without polarity or with less discriminative power, since their presence tend to mislead the learning process. Moreover, using word co-occurrence techniques, new composed bi-grams with high discriminative power are added which enhances the classification process. The best results are obtained using different combinations of techniques, depending on the dataset's homogeneity. On a homogeneous dataset, the performance in terms of precision is approximately 88% and, in terms of recall, a value of 93% is reached. In the case of a diverse dataset, the performance attained is 100%.