Detection of opinion spam with character n-grams

Abstract

In this paper we consider the detection of opinion spam asa stylistic classification task because, given a particular domain, the deceptiveand truthful opinions are similar in content but differ in the wayopinions are written (style). Particularly, we propose using character ngramsas features since they have shown to capture lexical content aswell as stylistic information. We evaluated our approach on a standardcorpus composed of 1600 hotel reviews, considering positive and negativereviews. We compared the results obtained with character n-gramsagainst the ones with word n-grams. Moreover, we evaluated the effectivenessof character n-grams decreasing the training set size in order tosimulate real training conditions. The results obtained show that charactern-grams are good features for the detection of opinion spam; theyseem to be able to capture better than word n-grams the content ofdeceptive opinions and the writing style of the deceiver. In particular,results show an improvement of 2.3% and 2.1% over the word-based representationsin the detection of positive and negative deceptive opinionsrespectively. Furthermore, character n-grams allow to obtain a good performancealso with a very small training corpus. Using only 25% of thetraining set, a Na¨ıve Bayes classifier showed F1 values up to 0.80 for bothopinion polarities.