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

Abstract

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.

Introduction

With the rapid development of social media platforms, people can access and analyze much user-generated content (UGC) to automatically identify authors personality traits. Many studies have shown that automatic personality detection systems play an essential role in various applications, such as user interest mining [1], information dissemination [2], recommendation systems [3], [4], [5], and intelligent machine design [6]. Therefore, analyzing and detecting users’ personality traits is significant for grasping users’ current and future psychologies and predicting their reactions and behaviors.

Personality detection research based on user-generated text is mainly divided into psycholinguistic lexicon-based, neural language model-based, and interpretability research. Earlier researchers captured psycholinguistic lexicon statistics features such as Linguistic Inquiry and Word Count (LIWC) [7] and Medical Research Council (MRC) [8] features in texts for personality detection [9], [10]. However, the obtaining artificial features are a costly operation, and a statistical analysis cannot effectively represent the original semantics. To avoid feature engineering, deep neural models are employed to learn text-distributed representations from end to end, and the resulting detection accuracy is greatly improved [11], [12], [13]. However, neural language model embeddings lack the ability to explain personality. Recently, some researchers combined common knowledge to detect personality [14], [15], providing some ability to explain personality and contributing to the analysis of personality traits. The latest researchers employed interpretable machine learning to clearly quantify the impacts of various psycholinguistic statistical features [16], [17]. However, these methods do not deeply exploit psycholinguistic domain knowledge and fail to effectively integrate psycholinguistic knowledge and text semantics into the associated neural models.

In the psychology field, personality traits are defined as attribute combinations of individual thoughts and emotions to explain the differences in human behaviors [18]. The generally used measurement metric are the Big Five personality, including openness, conscientiousness, extroversion, agreeableness, and neuroticism [19]. The relationship between personality and language has been studied for a long time. Psycholinguistics found an interesting phenomenon in empirical research: personality traits affect people’s use of language, which refers to their choice of vocabulary. Specifically, the LIWC lexicon [20], [21] and some personality adjectives [22] (Personality Adjectives Check List)¹ have linear correlations with each personality trait. In addition, people with the same personality traits usually have the same fixed emotional polarities [23]. The details regarding this topic are described in Appendix. Fig. 1 shows a visual example of a neurotic user’s psycholinguistic knowledge. From the words “hate”, “murder”, and “hell”, we can roughly infer that he/she is a neurotic user. Based on the relationship between the synonym “damn”, emotional polarity, and personality traits, this inference is more confident to be confirmed. It can be seen that conducting personality detection research from the lexical psycholinguistic knowledge perspective can bring rich domain structure knowledge rather than superficial psycholinguistic statistical information. Although research on personality detection has achieved remarkable results, some challenges still remain.

• Fusion of text semantics and psycholinguistic knowledge: It is a challenge to fully fuse lexical psycholinguistic knowledge and text semantics while accurately representing the personality traits derived from the user’s language.

• Interpretability of personality detection: It is a challenge to utilize personality psychology knowledge to realize explainable personality detection in neural models.

To meet the above challenges, we propose a novel lexical psycholinguistic knowledge-guided graph neural network model for interpretable personality detection. Our model enriches personality document representations by incorporating heterogeneous external knowledge through the use of personality lexicons as intermediaries. In particular, instead of directly using previous pretrained word embeddings, we first refine a kind of personality-aware word embedding via position encoding and an attention mechanism. Second, to fully fuse knowledge and semantics, we align the personality lexicons with the constructed personality knowledge graph and automatically build a heterogeneous personality word graph for each user. Then, we develop a Heterogeneous Personality Message-passing graph neural Network (HPMN) and perform interactions among the word nodes, emotion and personality heterogeneous nodes in directed edges. Finally, regarding the interpretability of personality traits, we design a graph-level readout function, which delicately selects all heterogeneous nodes for incorporation as knowledge-guided document embeddings to achieve user-generated text personality understanding and interpretation. Therefore, personality detection is transformed into a heterogeneous word graph classification problem. After conducting a verification on 4 public personality datasets, the results show that our model can effectively improve the accuracy of personality detection and pay more attention to critical knowledge.

In summary, our contributions can be summarized as follows.

• To the best of our knowledge, this is the first work that integrates lexical psycholinguistic knowledge and text semantics information into a neural model to achieve interpretable personality detection. Moreover, it provides support for lexical hypotheses in psycholinguistic research from a computational linguistic perspective.

• Our model incorporates the distribution representations of words and the lexical knowledge by learning personality-aware word embeddings. In addition, we construct a heterogeneous personality word graph and develop a message-passing network, which extracts explicit lexicon and knowledge relations via the interactions among heterogeneous graph nodes. All heterogeneous nodes are selectively incorporated as knowledge-guided document embeddings for personality understanding and interpretation through a carefully designed graph readout layer.

• Experiment results on four public datasets demonstrate that our model outperforms the state-of-the-art techniques in terms of personality detection. Our model can help various types of social software mine user information and help psychologists study and analyze personality traits in depth.

The rest of this paper is organized as follows. Section 2 introduces the work related to personality detection. Section 3 provides the problem formulation and Section 4 describes the proposed method. Further, Section 5 presents and analyzes the experimental results obtained on 4 public datasets. Finally, Section 6 outlines the conclusion and future research.