Skip to content

Latest commit

 

History

History
245 lines (191 loc) · 13.3 KB

measuring-values_human-ai.md

File metadata and controls

245 lines (191 loc) · 13.3 KB
Raw

Measuring Human and AI Values based on Generative Psychometrics with Large Language Models

by Haoran Ye, Yuhang Xie, Yuanyi Ren, Hanjun Fang, Xin Zhang, Guojie Song

https://arxiv.org/abs/2409.12106

Contents

Abstract

Generative Psychometrics for Values (GPV)

  • Human values and measurement: interdisciplinary inquiry since long time
  • Renewed interest due to advances in AI, specifically large language models (LLMs)
  • Introducing GPV: LLM-based data-driven value measurement paradigm

GPV Components:

  1. Fine-tuning an LLM: for accurate perception-level value measurement
  2. Verifying LLM capability to parse texts into perceptions, forming GPV pipeline core

Demonstration of Stability, Validity, and Superiority:

  • Applying GPV to human-authored blogs
  • Results indicate stability, validity, and superiority over prior psychological tools

Extensions for LLM Value Measurement:

  1. Psychometric methodology: measures LLM values based on scalable, free-form outputs
  2. Comparative analysis of measurement paradigms: reveals response biases in prior methods
  3. Attempt to bridge LLM values and safety: predictive power of different value systems, impact on LLM safety

Future Goals:

  • Leverage AI for next-generation psychometrics
  • Use psychometrics for value-aligned AI development.

Resources:

1 Introduction

Introduction:

  • Value theory is a cornerstone of philosophical inquiry, guiding ethical decision-making and shaping societal norms [66]
  • Traditional psychometric methods for measuring values have limitations such as response biases, resource demands, and inability to handle historical or subjective data [59]
  • Data-driven tools, like social media post analysis, have been developed but fail to grasp nuances of semantic meaning and context-dependent value expressions [26, 59]
  • The rise of Large Language Models (LLMs) opens up new possibilities for data-driven value measurement [86, 63]

Generative Psychometrics for Values (GPV):

  • Overcomes limitations of self-reports and dictionary-based tools by leveraging LLMs' advanced semantic understanding
  • Extracts contextualized and value-laden perceptions from texts and decodes underlying values for arbitrary value systems
  • Enables automatic generation of such items and their adaptation to any given data [76]

ValueLlama:

  • A fine-tuned LLM, ValueLlama, demonstrates outperformance in perception-level value measurement compared to state-of-the-art general and task-specific LLMs.

GPV Pipeline:

  • Parses texts into perceptions, which function similarly to psychometric items in self-report questionnaires
  • Demonstrates stability, validity, and superiority over prior psychological tools in measuring individual values.

Implications of LLMs:

  • Recent literature treats LLMs as subjects of value measurement using static, inflexible, and unscalable self-report questionnaires [50]
  • GPV constitutes a novel evaluation methodology that measures LLM values based on their scalable, free-form, and context-specific outputs
  • Mitigates response bias demonstrated in prior tools and enables context-specific value measurements.

Contributions:

  1. Introduce GPV, a novel LLM-based value measurement paradigm grounded in text-revealed selective perceptions
  2. Fine-tune Llama 3 for accurate perception-level value measurement and demonstrate its outperformance using ValueLlama
  3. Apply GPV to human-authored blogs, demonstrating its stability, validity, and superiority over prior psychological tools
  4. Enable LLM value measurements based on their scalable, free-form, and context-specific outputs by applying GPV across 17 LLMs and 4 value theories.

2 Related Work

Related Work

Value Measurements for Human Behavior:

  • Measuring individual values is important to understand the driving forces behind human behavior
  • Different measurement methods have been developed, including:
    • Self-report questionnaires: Participants rate their agreement with expert-defined perceptions
    • Behavioral observation: Experts analyze how personal values manifest in real-life actions
    • Experimental techniques: Structured scenarios are used to isolate and analyze variables affecting human behavior
  • These methods have limitations, such as:
    • Response biases
    • Resource demands
    • Inaccuracies in capturing authentic behaviors
    • Inability to handle historical or subjective data

Data-Driven Measurement Tools:

  • Dictionary-based tools: Determine values by analyzing the frequency of value-related lexicons, but overlook nuanced semantics and contexts
  • Deep learning models: Trained to identify values, but largely focused on specific value systems and not validated for individual-level measurements

Value Measurements for Language Models (LLMs):

  • LLMs are being integrated into public-facing applications, requiring comprehensive and reliable value measurements
  • Psychometric tests designed for humans have been applied to LLMs, including:
    • Dark triad traits
    • Big Five Inventory (BFI)
    • Myers–Briggs Type Indicator (MBTI)
    • Morality inventories
  • These test results are used to investigate the attributes of LLMs concerning:
    • Political positions
    • Cultural differences
    • Belief systems
  • Researchers have observed discrepancies between constrained and free-form LLM responses, with the latter being more practically relevant
  • Variability in LLM responses to subtle contextual changes necessitates scalable and context-specific evaluation methods

3 Generative Psychometrics for Values

Generative Psychometrics for Values

Value Measurement with Selective Perceptions

  • Values: individual's concepts of transitional goals, reflecting interests within motivational domains and guiding principles in life [76]
  • Value measurement quantitatively evaluates significance attributed to various values through behavioral and linguistic data [3, 53, 66]

Value Measurement Process (Definition 3.1)

  • V: value system, where each vi represents a particular value dimension
  • D: individuals' behavioral and linguistic data
  • w: value vector indicating the relative importance of each vi

GPV Instantiates Value Measurement through Selective Perceptions

  • Personal values are determinants of what individuals select to perceive, observe, and prioritize [60, 4]
  • Differing perceptions encode value-laden information and value orientations

Traditional Psychometric Inventories vs. GPV

  • Traditional inventories compile static and unscalable perceptions (items) as organized stimulus
  • GPV uses language models to dynamically generate perceptions according to behavioral and linguistic data
  • GPV effectively mitigates response bias, resource demands, and handling historical or subjective data [86]

Perception-level Value Measurement

  • Defines perceptions as value-laden, unambiguous, well-contextualized, and comprehensive measurement units
  • Trains Llama-3-8B to perform perception-level value measurement in an open-ended value space
  • Relevance classification determines if a perception is relevant to a value
  • Valence classification determines if a perception supports, opposes, or remains neutral towards a value

Evaluation of Perception-level Value Measurements

  • Compares the accuracy of ValueLlama with Kaleido and GPT-4 Turbo on relevance and valence classification [86] (Table 1)

Parsing and Aggregation Model

  • Parses individual's textual data into perceptions using GPT-3.5 Turbo guided by human values, definitions of perceptions, and few-shot examples
  • Evaluates parsing results with specifically trained human annotators [B.1]
  • Aggregates perception-level measurements for each value to obtain individual-level measurements

4 GPV for Humans

GPV for Humans

Measurement of Human Values:

  • Using 791 blogs from the Blog Authorship Corpus
  • Evaluating GPV through standard Psychological metrics:
    • Stability
    • Construct validity
    • Concurrent validity
    • Predictive validity
    • Demonstrates superiority over established psychological tools

Validation:

  • Stability: 86.6% of perception-level measurement results are consistent with individual-level aggregated results, indicating desirable stability

Construct Validity:

  • Evaluated using multidimensional scaling (MDS) to project 10 basic values and 4 higher-order values onto a two-dimensional plot
  • Relative positions of values align with the theoretically expected structure, indicating desirable construct validity

Concurrent Validity:

  • Compared GPV to the Personal Values Dictionary (PVD), a well-established measurement tool
  • Correlations between GPV and PVD measurements:
    • Identical/compatible values show positive correlations
    • Opposing values exhibit negative correlations
  • Theoretically expected correlations support the concurrent validity of GPV

Predictive Validity:

  • Measured by examining if GPV results align with blog authors' gender-related socio-demographic traits
  • Men prioritize power, stimulation, hedonism, achievement, and self-direction, while women emphasize benevolence and universalism
  • GPV measurement results align with these established statistical theories

Case Study:

  • Exemplifies the advantage of GPV over PVD in capturing implicit values in text
  • GPV effectively captures the author's intentions, while PVD fails to reflect the intended values or align with the measurement subject in context

5 GPV for Large Language Models

Evaluation of Large Language Models using GPV

GPV for Large Language Models:

  • Evaluated using:
    • Self-report questionnaires
    • ValueBench
    • GPV
  • Used LLM-generated value-eliciting questions for GPV to ensure comprehensive measurement
  • Across 19,910 perception-value pairs, 86.8% of results were consistent with LLM aggregated results

Comparative Analysis of Construct Validity:

  • Compared GPV against prior measurement tools: Self-Direction, Stimulation, Hedonism, Achievement, Power, Security, Conformity, Tradition, Benevolence, Universalism
  • Examined correlation between Schwartz's values using different measurement tools (Figure 4)
    • GPV showed superior construct validity as its measurements aligned more closely with the theoretical structure
    • Prior tools exhibited positive correlations between distant values, indicating response bias
  • Evaluated construct validity by relating values from different value theories (Table 4)
  • Concluded that GPV showed superior construct validity over prior tools prone to response bias.

Comparative Analysis of Value Representation Utility Tools

Accuracy of Measurement Tools (percentage)

  • Self-report: 56.7 ±26.0
  • ValueBench: 67.8 ±20.6
  • GPV: 85.6 ±14.1

Utility in Predicting LLM Safety Scores:

  • Human value measurements have predictive power for human behavior [83]
  • Few studies connect LLM values with their safety
  • Evaluate the predictive power of different measurement tools for LLM safety scores using GPV's safety scores as ground truth

Results:

  • Using linear probing classifier to predict relative safety of LLMs based on value measurement results from SALAD-Bench [43]
  • Train 30 times with randomly sampled data splits for statistically meaningful results

Findings:

  • Different value systems lead to different results
  • GPV is more predictive of LLM safety scores than prior tools
  • VSM (Value Systems Measurement) [31] is more predictive of LLM safety and has positive/negative impact on LLM safety based on values like Long-term Orientation or Masculinity

Discussions - Superiority of GPV:

  • Knowledge embedded within ValueLlama enhances the measurement process and ensures construct validity
  • Context-specific value measurements are necessary for LLMs [67]
  • GPV enables context-specific measurements, mitigating response bias, being more practically relevant, and scalable.

Limitations and Future Work:

  • Current studies are limited to English language evaluations
  • Future research should explore multi-lingual measurements
  • Investigate the spectrum of values an LLM can exhibit and how different profiling prompts affect this spectrum.

6 Conclusion

GPV: A Tool for Value Measurement

  • Introduction: Introduces GPV, an LLM-based tool designed for value measurement, theoretically based on text-revealed selective perceptions.
  • Superiority of GPV: Experiments demonstrate the superiority of GPV in measuring both human and AI values.
  • Potential Applications: Offers promising opportunities for both sociological and technical research.

Sociological Research:

  • Scalable, Automated, Cost-Effective Measurements: Enables scalable, automated, and cost-effective value measurements.
  • Reduces Response Bias: Reduces response bias compared to self-reports.
  • Provides More Nuance: Provides more semantic nuance than prior data-driven tools.
  • Flexibility: Can be used independently of specific value systems or measurement contexts.

Technical Research:

  • New Perspective on Value Alignment: Presents a new perspective on value alignment by offering interpretable and actionable value representations for LLMs.