Constructing a Psychometric Testbed for Fair Natural Language Processing

Ahmed Abbasi, David G. Dobolyi, John P. Lalor, Richard G. Netemeyer, Kendall Smith, Yi YangEMNLP 2021 · DOI

Introduction

Psychometrics—the scientific measurement of latent psychological constructs such as trust, anxiety, or literacy—traditionally relies on extensive, validated surveys. While these instruments provide high-quality data, they are often time-consuming and invasive. Natural Language Processing (NLP) offers a potential alternative: inferring these traits from the text individuals naturally produce in clinical notes, forum posts, or patient portals. However, the field has lacked a robust dataset that pairs raw text with "gold-standard" psychometric scores and demographic metadata necessary for fairness auditing.

This research introduces a novel psychometric testbed designed to bridge this gap. By collecting over 33,000 text responses from 8,502 participants alongside validated survey results, the authors provide a framework for training NLP models that can predict complex psychological states while being audited for racial and gender bias.

Methodology: Building the Testbed

The researchers developed their testbed through a multi-stage process focused on four critical health-related dimensions:

  1. Trust in Physicians (TD): Measuring the degree of confidence a patient has in their doctor's decisions.
  2. Anxiety Visiting the Doctor (AV): Assessing the emotional distress associated with medical examinations.
  3. Subjective Health Literacy (HL): Evaluating a person's self-perceived ability to navigate hospital materials.
  4. Objective Health Numeracy (HN): Testing the ability to interpret medical statistics and data.

To establish ground truth, the study utilized well-validated scales from social science literature. For example, the Trust in Physician scale uses a Likert-style format to gauge patient sentiment.

Example of the validated survey items used to measure Patient Trust in a Physician.

The core innovation lies in the Text Elicitation phase. Participants were asked to write short responses to prompts designed to "prime" the specific trait being measured. The researchers found that "alignment"—placing the text box immediately after the corresponding survey items—resulted in the most relevant and predictive text.

Predictive Performance and the "Fairness Trap"

The study evaluated five NLP architectures, including BERT and simpler models like Linear Regression and WordCNN. BERT generally emerged as the most powerful predictor, achieving Area Under the Curve (AUC) scores between 0.72 and 0.84 for binary classification tasks. "Trust in Physicians" proved to be the most accessible trait for text-based prediction, while "Anxiety" was more elusive.

However, the high accuracy of complex models like BERT came with a significant caveat: the "Fairness Trap." When auditing for racial bias using metrics like Disparate Impact (DI), the researchers discovered that BERT was significantly less fair in the "Numeracy" category.

A sample question from the Health Numeracy assessment requiring the interpretation of a nutrition label.

Specifically, BERT was 30% less likely to assign a high numeracy score to non-white participants compared to white participants, even when their actual survey-based numeracy levels were identical. This suggests that the model may be picking up on linguistic patterns associated with race rather than the underlying construct of numeracy, highlighting a critical tension between model accuracy and social equity.

Conclusion and Future Directions

The creation of this psychometric testbed demonstrates that text-based proxies for psychological traits are viable, with correlations on par with other established NLP tasks. However, the findings underscore that a model’s performance on paper does not guarantee its fairness in practice.

Future work should aim to expand this methodology to other domains, such as financial literacy or cybersecurity perception. Furthermore, as these models move closer to clinical application, researchers must account for intersectional identities—how the combination of race, age, and income affects model predictions—to ensure that the benefits of automated psychometrics are distributed equitably across all populations.

Summary updated

BibTeX
@inproceedings{abbasi2021constructing,
  author = {Ahmed Abbasi and David G. Dobolyi and John P. Lalor and Richard G. Netemeyer and Kendall Smith and Yi Yang},
  title = {{Constructing a Psychometric Testbed for Fair Natural Language Processing}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.emnlp-main.304},
}
Figure