X-PERICL: An Explainable Method for Personality Assessment based on Hybrid Linguistic Features and In-Context Learning LLMs

ICLR 2026 Conference Submission12773 Authors

Keywords: in context learning, large language models, explainable AI, interpretability, personality traits assessment

TL;DR: This work proposes a novel eXplainable method for PERsonality assessment by linguistic features with In-Context Learning LLMs – X-PERICL

Abstract: This work proposes a novel eXplainable method for PERsonality assessment by linguistic features with In-Context Learning LLMs – X-PERICL. It allows predicting the Big Five Personality traits (PTs) by text using a hybrid feature fusion combined with off-the-shelf Large Language Model (LLM). The aim is to improve both the performance and interpretability of the Personality Assessment (PA) models. Transformer-based deep features capture local contextual patterns, while hand-crafted features obtained using the Linguistic Inquiry and Word Count (LIWC) dictionary provide global and local insights into PTs reflected in text. These explainable global and local patterns are used as prompts for LLM to generate final predictions of PTs and explanations. Experiments on the ChaLearn First Impressions v2 corpus demonstrate that the integration of hand-crafted features with deep embeddings outperforms standalone representations, achieving a mean accuracy (mACC) of 0.891 and a Concordance Correlation Coefficient (CCC) of 0.333. The interpretability analysis reveals the linguistic patterns associated with each PTs, offering insights for psychological and computational linguistic research, including paralinguistics. In turn, interpretable global and local patterns based on hybrid feature fusion used as prompts for LLM enable a relative increase in Concordance Correlation Coefficient (CCC) to 9.6%. X-PERICL improves both performance and interpretability in PA, with potential applications in psychological profiling, employee selection, and personalized recommendations. The codes will be available after the paper acceptance.

Primary Area: interpretability and explainable AI

Submission Number: 12773