Impact of polarity in deception detection

Usually, most works use and combine different methods for generating features in order to improve deception detection; nevertheless, they do not take into account the fact that features may change depending on the nature of text. In this research, a study on the effect of the polarity over the set of features generated for deception detection task was carried out. We implemented a polarity classifier to generate subsets of positive and negative opinions. Next, a semantic and lexical method were used over the subsets to generate features and construct vectors. It was proven that adding polarity information did not positively impacted on deception detection. However, partitioning datasets improved classification results. To classify subsets, attribute selection was implemented and a Bayesian classifier was fed with the resulting vectors. Research findings show that cues to deception are affected by the opinion polarity. In addition, this approach registered up to 86% f-measure.