Socializing in Interference: Quantum-Inspired Topic Model with Crystal-Like Structure Grid for Microblog Topic Detection

Abstract

Capturing highly coherent topics from large-scale informal social media messages has always been a challenging task. Existing topic models fail to model the nonlinear dependence between text and structure, which is called the interference effect in quantum cognition. Therefore, we propose a Quantum-inspired Topic Model (QTM), which naturally admits a non-linear context composition. Specifically, based on the iterative-deepening random walks, we design a Crystal-like Structure Grid (CSG) to obtain the user structural features. Then, we propose a quantum density operator based network embedding. Such a density operator is essentially a nonlinear joint representation of the user structure and text information. The resulting user sequence embeddings are fed into the Neural Variational Inference (NVI) for topic detection. Extensive experimental results on three real-world microblog datasets demonstrate that QTM outperforms state-of-the-art models.