GPT is an effective tool for multilingual psychological text analysis.

Area of Science:

• Social and behavioral sciences
• Computational linguistics
• Psychological science

Background:

• Automated text analysis is increasingly used in social and behavioral sciences to measure psychological constructs.
• Existing methods often lack cross-linguistic capabilities and require significant expertise.

Purpose of the Study:

• To evaluate the efficacy of Generative Pre-trained Transformer (GPT) models for automated psychological text analysis in multiple languages.
• To compare GPT's performance against traditional dictionary-based methods and fine-tuned machine learning models.

Main Methods:

• Utilized 15 datasets comprising 47,925 annotated tweets and news headlines.
• Tested GPT versions (3.5 Turbo, 4, 4 Turbo) for detecting psychological constructs (sentiment, emotions, offensiveness, moral foundations) across 12 languages.
• Compared GPT performance with English-language dictionary analysis and established machine learning models.

Main Results:

• GPT models (r = 0.59 to 0.77) significantly outperformed English-language dictionary analysis (r = 0.20 to 0.30).
• GPT performance was comparable to, and sometimes exceeded, top-performing fine-tuned machine learning models.
• Performance improved with newer GPT versions, especially for less common languages, and costs decreased.

Conclusions:

• GPT models provide a highly accurate, versatile, and user-friendly alternative for automated psychological text analysis.
• GPT's ability to function without training data and with simple prompts democratizes advanced NLP for researchers, including those studying under-resourced languages.
• LLMs facilitate cross-linguistic research and make sophisticated text analysis more accessible to a broader scientific community.