Characterization of the Discrepancies between Scores and Texts of Movie Reviews

• Karen S. Martins UFMG
• Pedro O. S. Vaz de Melo UFMG

Resumo

Review websites have changed the business world, influencing the sale and purchase of products, music and movies. They allow users to evaluate a particular subject from a score and a text. Since the reviews are based on personal opinions, it is very common to find large score variations for the same movie. A reasonable score for one critic can be considered low for another. In addition, it is also possible to find variations in the texts. Some critics use negative words to describe a movie and choose a high score. While, other critics praise and choose a lower rating, compared to the other critic who used only negative words. In this work, we formulate a sentiment analysis task to analyze and quantify such discrepancies. To do that, we applied a state of the art deep learning architecture on a large collection of movie reviews posted on Metacritic. Our results reveal that the score and text are usually not compatible. Finally, we show that, while some users consistently give discrepant reviews, some movies are more likely to receive reviews of this nature.