On the Use of Homogenous Sets of Subjects in Deceptive Language Analysis

Recent studies on deceptive language suggest that machine learning algorithms can be employed with good results for classification of texts as truthful or untruthful. However, the models presented so far do not attempt to take advantage of the differences between subjects. In this paper, models have been trained in order to classify statements issued in Court as false or not-false, not only taking into consideration the whole corpus, but also by identifying more homogenous subsets of producers of deceptive language. The results suggest that the models are effective in recognizing false statements, and their performance can be improved if subsets of homogeneous data are provided.