by Michael E. W. Varnum, Nicolas Baumard, Mohammad Atari, Kurt Gray https://www.pnas.org/doi/10.1073/pnas.2407639121
Historical Large Language Models (HLLMs)
Background:
- Traditional behavioral science focuses on present day studies
- Limitations: no access to historical data
- Criticism for being too parochial, call for expansion beyond WEIRD societies
- Need to incorporate diverse historical data for generalizable theories
Introducing HLLMs:
- Generative models trained on historical corpora
- Simulate responses of populations no longer living
- Offer opportunities to gather historical psychological data
Advantages of HLLMs:
- Create new opportunities in behavioral science research
- Expand understanding of human nature beyond present day societies
- Provide insights into various historical attitudes and behaviors
Examples:
- Comparing cooperative tendencies: Vikings vs Ancient Romans vs Early Modern Japanese
- Exploring gender roles: Ancient Persians vs Medieval Europeans
- Addressing hard-to-answer questions
Benefits:
- Escape from temporal trap in behavioral science research
- Diversify data beyond present day participants
- Enhance understanding of historical societies and their psychological aspects.
Large Language Models (LLMs)
- Massive neural networks trained on natural language data
- Understand and generate natural language output
- Predict probable words based on sequence of prior words
- Fine-tuned through supervised learning for specific tasks
- Transforming psychology and adjacent fields
- Simulate human responses across domains
- Replicate patterns in moral judgment, economic game behavior, cognitive biases, obedience experiments
- Used as substitutes for human subjects with caution
- Limitations:
- Reflect the cultures on which they are trained
- Limited cross-cultural generalizability due to WEIRD sampling bias
- LLMs reflect the psychology of different cultural groups through training on different corpora.
- Previous techniques for inferring psychological tendencies from text data:
- Google Ngrams, newspapers, movie dialogue, etc.
- Limited to indirect proxies for psychological traits and tendencies
- Proposed use of HLLMs (High-level Language Models) to venture beyond existing techniques.
Historical Language Models (HLLMs)
- Simulate responses of diverse past societies using modern psychological instruments and behavioral measures
- Enable measurement of historical populations' responses without direct access to living individuals
- Based on large corpora of historical text, including fiction, diaries, letters, scholarly texts
- Reproduce psychological responses of historical populations (13)
- Insight into thinking of populations no longer living
- Study trends in psychological tendencies over longer time spans
- Test historical generalizability of contemporary psychological phenomena
- Complement qualitative approaches used by historians
- Encouraging first steps for simulating historical samples for research (MonadGPT, XunziALLM)
- Unclear how accurately these models reflect the true underlying mindset of past populations.
Building a Historical Language Model:
- Acquire sizable amount of historical text from a specific time period
- Convert text to machine-readable format
- Encode vectors and feed them to neural network architecture
- Generate probability distributions for words
- Create an LLM
- Use chat interface to simulate participants and run psychological experiments.
Challenges and Caveats for Historical Language Models (HLLMs)
Acquiring Sufficient Training Data:
- Smaller training corpora compared to current LLMs
- Historical texts may not be representative of the population as a whole
- Elites are overrepresented in historical text, potentially skewing results
- Validation through other archival sources and traditional approaches needed
Benchmarking:
- Limited availability of benchmarks for HLLMs due to lack of contemporary human data
- Use of historical psychology, ethnographic data, and archaeological data to assess accuracy
- Experts can fine-tune models based on socioeconomic status effects in modern populations
Generalizing from Historical Text:
- Substantial challenges in creating representative samples of past populations
- Importance of validating results through multiple sources and approaches
Potential Solutions and Future Directions:
- Combining HLLMs with other research methods and approaches
- Continued development of computational tools for working with historical data
- Increased availability of larger, more diverse historical text databases.