| Resumen |
Our motivation is to assess the effectiveness of support vector networks (SVN) on the task of detecting deception in texts, as well as to investigate to which degree it is possible to build a domain-independent detector of deception in text using SVN. We experimented with different feature sets for training the SVN: a continuous semantic space model source represented by the latent Dirichlet allocation topics, a word-space model, and dictionary-based features. In this way, a comparison of performance between semantic information and behavioral information is made. We tested several combinations of these features on different datasets designed to identify deception. The datasets used include the DeRev dataset (a corpus of deceptive and truthful opinions about books obtained from Amazon), OpSpam (a corpus of fake and truthful opinions about hotels), and three corpora on controversial topics (abortion, death penalty, and a best friend) on which the subjects were asked to write an idea contrary to what they really believed. We experimented with one-domain setting by training and testing our models separately on each dataset (with fivefold cross-validation), with mixed-domain setting by merging all datasets into one large corpus (again, with fivefold cross-validation), and with cross-domain setting: using one dataset for testing and a concatenation of all other datasets for training. We obtained an average accuracy of 86% in one-domain setting, 75% in mixed-domain setting, and 52 to 64% in cross-domain setting. |
