A lexical psycholinguistic knowledge-guided graph neural network for interpretable personality detection

With the blossoming of online social media, personality detection based on user-generated content has a significant impact on information scientific and industrial applications. Most existing approaches rely heavily on semantic features or superficial psycholinguistic statistical features calculated by existing tools and fail to effectively exploit psycholinguistic knowledge that can help determine and interpret peoples personality traits. In this paper, we propose a novel lexical psycholinguistic knowledge-guided graph neural model for interpretable personality detection, which leverages the personality lexicons as a bridge for injecting relevant external knowledge to enrich the semantics of a document. Specifically, we learn a kind of personality-aware word embedding, that encodes psycholinguistic information in the continuous representations of words. Then, a Heterogeneous Personality word graph is constructed by aligning the personality lexicons with the personality knowledge graph, which is fed into a Message-passing graph Network (HPMN) to extract explicit lexicon and knowledge relations through the interactions among heterogeneous graph nodes. Finally, through a carefully designed readout function, all heterogeneous nodes are selectively incorporated as knowledge-guided document embeddings for user-generated text personality understanding and interpretation. Experiments show that our model effectively detects personality traits. Moreover, it provides a certain level of support for lexical hypotheses in psycholinguistic research from a computational linguistics perspective.