Multi-Gate Mixture-of-Experts with Explanation for Predictive Computational Personality Analysis

Abstract Personality recognition has emerged as a rapidly expanding research area with significant applications in human–computer interaction, psychological assessment, and social robotics, driven by advances in computational modeling and the increasing availability of linguistic and behavioral data. Despite notable advances, accurately predicting personality traits from text remains challenging due to the opaque nature of deep learning models and their limited explainability across linguistic contexts. However, because personality prediction is underexplored in most languages, existing research has primarily focused on English texts, resulting in limited studies on multilingual tasks. In this paper, we propose MMoECP, a Multi-Gate Mixture of Experts for Computational Personality Analysis framework that integrates a diverse set of computational models. We apply local and global model-agnostic explanation techniques to enhance explainability across models. We conduct our experiments on five distinct datasets, and the results show that the proposed MMoECP framework achieves superior performance compared to multi-label learning models. This framework offers a promising approach to building explainable personality-aware systems in multilingual and culturally diverse contexts. Furthermore, we employ chain-of-thought prompting to generate explainable reasoning from large language models, contributing to a deeper qualitative understanding of personality trait analysis.