Finding Deceptive Opinion Spam by Correcting the Mislabeled Instances

Assessing the trustworthiness of reviews is a key in natural language processing and computational linguistics. Previous work mainly focuses on some heuristic strategies or simple supervised learning methods, which limit the performance of this task. This paper presents a new approach, from the viewpoint of correcting the mislabeled instances, to find deceptive opinion spam. Partition a dataset into several subsets, construct a classifier set for each subset and select the best one to evaluate the whole dataset. Error variables are defined to compute the probability that the instances have been mislabeled. The mislabeled instances are corrected based on two threshold schemes, majority and non-objection. The results display significant improvements in our method in contrast to the existing baselines.