Spot the Bot: Motifs in Semantic Spaces

Fast growth and introduction of social media to everyday activities have led to an over-exposure of information that is currently presented as reviews or comments, mainly on social networks. As these media, nowadays, are the most important source for decision-making, they have been used to over-promote and discredit products and services, a task that in most cases are carried out through bots that generate comments automatically. This paper addresses the problem of identifying semantic characteristics of sincere-comments and computer-generated for the Spanish language. For its development in the first stage, a vocabulary was generated from a set of literary pieces (from the Spanish language literature) and vector representations of the words, were obtained computing the SVD (order-R) decomposition through the Implicitly Restarted Lanczos Method; with these representations were built the semantic spaces of literary pieces, sincere-comments, and computer-generated comments. To identify the inherent characteristics of these texts two approaches were followed, the first approach was based on subsequence time series clustering using the DBSCAN algorithm, then together with cluster validly measures such as SD, S-Dbw, Dunn, Davies-Bouldin, Silhouette Indexes, were extracted cluster properties to characterize each of the types of texts. In the second approach was applied to our semantic spaces the Bandt and Pompe approach together with the Shannon entropy measure to compute a Statistical Complexity measure and build the so-called Complexity – Entropy causality plane, also, a methodology was proposed to construct the Complexity – Entropy causality plane for the n-dimensional case. The obtained results indicated that the computer-generated comments present more compact clusters than the sincere comments and that the literary pieces differ from the comments in the Complexity – Entropy causality plane, but unfortunately the latter  are indistinguishable on the same plane.