What question did this study set out to answer?

This research aims to evaluate how well Pseudo-Factor Analysis can replicate the outcomes of Exploratory Factor Analysis using Pretrained Language Models in multilingual contexts.

March 13, 2026Open Access

Expanding Psychometrics with Pretrained Language Models: Evaluating Pseudo-Factor Analysis in Applied and Multilingual Contexts

Key Points

This research aims to evaluate how well Pseudo-Factor Analysis can replicate the outcomes of Exploratory Factor Analysis using Pretrained Language Models in multilingual contexts.
Implemented Pseudo-Factor Analysis using item-level embeddings from Pretrained Language Models.
Compared PFA results with traditional Exploratory Factor Analysis across psychometric measures.
Evaluated performance in both English and Italian contexts.
PFA achieved up to 100% factor recovery in English measures with high congruence (>0.90).
Italian results displayed more variability in factor recovery outcomes.
The study highlights the potential and limitations of PFA's language sensitivity in psychometric applications.

Abstract

The integration of Pretrained Language Models (PLMs) into psychometrics introduces new opportunities for scale development and validation. Pseudo-Factor Analysis (PFA), which employs item-level embeddings as proxies for human responses, offers a data-efficient alternative to Exploratory Factor Analysis (EFA). This study examines PFA’s capacity to replicate EFA outcomes across clinical and health psychology measures in English and Italian. PFA showed high performance in English (factor recovery up to 100%, congruence >.90), while Italian results were more variable. Findings highlight both the promise and language sensitivity of PFA, underscoring its potential for cross-linguistic psychometric applications.

Read Full Paperexternally

Demander à l'IA

Bookmark

View Full Paper