A subjectivity detection-based approach to sentiment analysis

Abstract

With the rise of Web 2.0 where loads of complex data are generated every day, effective subjectivity classification has become a difficult task in these days. Subjectivity classification refers to classifying information into subjective (expressing feelings) or objective (expressing facts). In this paper, we use Yelp reviews dataset. Our aim is to prove that a dataset with the objective sentences removed from each review gives better results than the dataset containing both subjective and objective sentences. To achieve this, we have used two approaches, each divided into two phases. The first phase of both the approaches is mainly the subjectivity classification phase where we filter out the objective sentences and keep the subjective sentences in the reviews, thus creating a new dataset with purely subjective reviews. The second phase of the first approach uses CountVectorizer which creates word vectors, and we fit the model to the classifiers. The second phase of first approach is repeated for both the datasets, and we get better results for the newly created dataset which contains purely subjective reviews. The second phase of the second approach uses Word2Vec, an implementation of neural network which creates distributed word vectors. We fit this Word2Vec model to the classifier, and we analyze the results. Again, the newly created dataset gives better results after we repeat this phase of the second approach for both the datasets.