Negation Handling and Domain Generalization in Sentiment Analysis using Machine Learning Models

Conventional lexicon-based approaches to sentiment analysis typically lack the necessary methods to properly identify the negation window, making it impossible to model negation. An enormous increase in sentiment-rich electronic and social media has been observed daily. Negation modifiers cause problems for Sentiment Classification techniques and have the power to entirely change the discourse’s meaning. Therefore, it becomes essential to manage them well. Opinion mining or sentiment analysis is the study of people’s attitudes, feelings, and views as they are expressed in written language. It is one of the busiest text mining and natural language processing research projects. Even though sentiment analysis research has gained popularity in the field of natural language processing, for this problem, the state-of-the-art machine learning approach is based on Bag of Words. But the BOW model pays little attention to polarity shift, which could have a distinct overall effect. One of the main issues with doing sentimental analysis on any given text or sentence is handling polarity shift, which is what this study attempts to address. Sentiment analysis use Natural Language Processing principles to identify negation in the text. Our goal is to identify the negation effect on customer reviews that, although appearing good, are actually negative. The suggested modified negation methodology helps to increase classification accuracy by providing a method for computing negation identification. In terms of review classification by accuracy, precision, and recall, this approach yielded a noteworthy outcome. When test and training data are from distinct domains, machine learning faces the challenge of domain generalization. Despite the large body of research on cross-domain text classification, the majority of current methods concentrate on one-to-one or many-to-one domain adaptation. Our domain generalization method regularly outperforms state-of-the-art domain adaption methods, according to experimental results. In-depth research on domain-independent sentiment analysis classification models that only need to be trained once is suggested in this publication.