PsyCoT: A Window into the Future of AI That Can Perceive Our Personalities and Moods
Personality detection involves identifying an individual's personality traits purely from their written text. This capability has many potential applications, like improving human-computer interaction. However, traditional machine learning models struggle to capture the nuances required for accurate personality detection from texts.
Researchers from Sun Yat-sen University propose a new method called PsyCoT that takes inspiration from psychological questionnaires to significantly boost the performance of large language models (LLMs) like GPT-3.5 for zero-shot personality detection.
The PsyCoT Approach
PsyCoT treats personality detection as a multi-step reasoning task, similar to a person completing a personality questionnaire. It has the LLM take on the role of an AI assistant specializing in text analysis.
Over multiple dialogue turns, the assistant is prompted to rate relevant questionnaire items based on the input text, with access to previous ratings. After rating all items, the assistant makes an overall personality judgment. This mimics the human process of reasoning through a questionnaire to evaluate someone's personality. The items act as an effective chain-of-thought, enhancing the LLM's capability for nuanced personality inference.
- PsyCoT prompts the LLM to rate questionnaire items in a multi-turn dialogue, mimicking how humans complete personality tests. This acts as an effective chain-of-thought.
- Experiments use two personality taxonomies - Big Five and MBTI. PsyCoT improves performance over standard prompting by 4.23 and 10.63 average F1 points respectively.
- On the Essays dataset, PsyCoT achieves better results than standard prompting on 4 out of 5 Big Five traits, even outperforming fine-tuned BERT and RoBERTa on some.
- Ablations show the multi-turn dialogue structure is important for consistency and accuracy compared to single-turn rating of all items together.
- PsyCoT exhibits greater robustness to input variations like changed option order compared to other prompting methods.
- The questionnaire used significantly impacts performance, so selecting the right one to cover relevant traits is crucial.
Overall, PsyCoT provides an effective way to incorporate rigorous human reasoning processes into LLMs for improved complex judgment capabilities.
By incorporating the rigorous reasoning process of questionnaires, PsyCoT unlocks stronger personality detection abilities in LLMs without any training. It provides a promising new prompting technique for complex human judgments.
Future work can explore PsyCoT's effectiveness for different LLMs, languages, and questionnaires. Overall, it offers an exciting way to elicit nuanced reasoning from the knowledge and capabilities of LLMs.
Full credit to Yang, Tao, Tianyuan Shi, Fanqi Wan, Xiaojun Quan, Qifan Wang, Bingzhe Wu, and Jiaxiang Wu. "PsyCoT: Psychological Questionnaire as Powerful Chain-of-Thought for Personality Detection." arXiv preprint arXiv:2310.20256 (2023).
The paper is available on arXiv at this link: https://arxiv.org/abs/2310.20256