From Lexicons to Large Language Models: A Holistic Evaluation of Psychometric Text Analysis in Social Science Research
研究发现大语言模型在文本心理分析中准确度媲美专业AI，成本更低，且将种族和性别偏见降低60%，并提出了“认知-情感提示”技术以提升准确率。
Research Spotlight Abstract Extracting psychological insights from text is vital for modern analytics, yet organizations often rely on analysis tools that are either biased and simplistic or prohibitively expensive to build. Our research demonstrates that Large Language Models (LLMs) offer a superior alternative. They match the accuracy of specialized artificial intelligence (AI), while significantly reducing costs and technical barriers. Crucially for policy considerations, we find LLMs are statistically fairer than traditional methods. In our tests, they reduced racial and gender bias by up to 60%. Beyond assessing performance, we introduce a practical technique called “cognitive-affective prompting.” By instructing the AI to adopt specific human strengths, such as using “superior reasoning” for complex tasks or “emotional intelligence” for sentiment analysis, practitioners can boost accuracy by over 10%. To facilitate adoption, we provide a user-friendly “cookbook” to help nonexperts apply these findings immediately. For policymakers and business leaders, this research validates LLMs as a robust, consistent, and equitable standard for analyzing human behavior at scale.