PoS tag-based Attention for Feature Selection in Sentiment Analysis

The growth in the usage of the Internet and Web-based applications has led to an exponential increase in text data, including customer reviews, social media comments, and blogs. Sentiment analysis is an essential technique to analyze this data and determine the polarity of opinions. However, traditional sentiment analysis methods face challenges in feature selection and representation due to the unstructured nature of the text and associated noise. We propose a novel framework called PoPoSBert that uses BERT to capture syntax, semantic, and contextual features effectively. We leverage part-of-speech (PoS) tags which assign a unique identifier to every single token in a text corpus to indicate the part of speech, to capture polarity words that reveal the sentiment behind the text. We use the transformer architecture's attention mechanism to improve the weights for polarity words, resulting in high-quality feature selection for sentiment classification. We conducted an extensive experimental analysis to evaluate the suggested framework against the existing methods. The suggested framework is assessed using several standard evaluation measures such as accuracy, precision, recall, and F1-score against the traditional baseline methods such as TF-IDF and existing models like BERT, Word2Vec, and GloVe techniques. The findings indicate that the PoS tag-based attention concept improves the accuracy of sentiment classification, achieving superior performance over existing methods on several benchmark datasets. The suggested framework can be utilized in various applications, including market research, monitoring social media and analyzing customer feedback to make informed decisions.