Using Cohesion-Based and Sentiment-Based Attributes to Classify Spoilers in Movie Reviews

Abstract

Spoiler reviews have different narrative patterns from non-spoiler reviews. Their narrative is more precise about what happened in the movies, while that of non-spoiler reviews is more obscure due to the omission of specific details. Our research extracted 108 cohesion-based and 6 sentiment-based attributes from movie reviews, which captured these patterns. The classification was done using Naive Bayes and a support vector machine (SVM) with a linear kernel. SVM achieved the best performance of 78% accuracy and 0.78 F -measure of class spoiler. Most contributing attributes were also determined from the weight vector given by the SVM. They supported our initial observation about the differences in narrative patterns between spoilers and non-spoilers.