Beyond True Label: Label-Assumed Evidence Extraction for Personality Prediction

Most existing methods enhance the personality prediction performance of LLMs by integrating few-shot and Chain-of-Thought learning strategies. However, the related studies may introduce biased reasoning in LLMs by providing true labels during CoT construction. Furthermore, they normally overlook the discriminative contributions of different label-specific evidence when selecting demonstration examples. In this paper, we propose a Label-Assumed Evidence Extraction (LAEE) method to classify user personality. Concretely, we assume the user’s personality labels to extract supporting evidence for each label. The role of the evidence in our LAEE method is twofold. On the one hand, we perform a weighted fusion of the label-specific evidence to construct sample representations that emphasize discriminative cues, enabling the selection of highly relevant demonstration samples for few-shot learning. On the other hand, we guide the LLM to synthesize evidence from different label assumptions without access to the ground-truth label, thereby producing unbiased and comprehensive CoTs that further support the few-shot prediction process. The experimental results demonstrate that our LAEE method not only achieves the highest classification performance on four personality traits but also offers more comprehensive reasoning that considers both label-consistent and label-inconsistent evidence.