This paper presents our experimental work on two aspects of sentiment analysis. First, we evaluate the performance of different machine learning as well as lexicon based methods for sentiment analysis of texts obtained from variety of sources. Our performance evaluation results are on six different datasets of different kinds, including movie reviews, blog posts and twitter feeds. To the best of our knowledge no such work on comprehensive evaluative account involving different techniques on variety of datasets have been reported earlier. The second major work that we report here is about the heuristic based scheme that we devised for aspect-level sentiment profile generation of movies. Our algorithmic formulation parses the user reviews for a movie and generates a sentiment polarity profile of the movie based on opinion expressed on various aspects in the user reviews. The results obtained for the aspect-level computation are also compared with the corresponding results obtained from the document-level approach. In summary, the paper makes two important contributions: (a) it presents a detailed evaluative account of both supervised and unsupervised algorithmic formulations on six datasets of different varieties, and (b) it proposes a new heuristic based aspect-level sentiment computation approach for movie reviews, which results in a more focused and useful sentiment profile for the movies.
Performance Evaluation in Arabic Sentiment Analysis during the Covid-19 Pandemic