Structural feature analysis of chat communication based on high-dimensional relational data

Abstract

Chat communication often exhibits behaviors that hinder smooth interaction, such as prolonged discussions, topic drift, and branching interactions. This study proposes an analytical framework that represents dialogue structures as high-dimensional relational data, rather than summarizing them using predefined network indicators. We construct a content tensor based on lexical distributions (5 dimensions) and a reply tensor based on reply relations (6 dimensions), and apply tensor decomposition to extract content factors and reply factors. Abnormality is treated not as a supervised label or a qualitative phenomenon, but as an analytical axis representing statistical deviations in thread length and lexical distributions. A classification task is used as an analytical probe for interpreting the extracted factors. Comparative analyses using the Ubuntu Dialogue Corpus and the Nico Nico Pedia dataset show that reply factors are consistently more strongly associated with statistical abnormality clusters than content factors.