Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

Multimodal Test Item Parameter Prediction From Text, Images, and Metadata: Fusing Together AI Vision and Language

Hotaka Maeda1,2, Yikai Ek Lu3

  • 1Smarter Balanced, Santa Cruz, CA, USA.

Educational and Psychological Measurement
|July 15, 2026
PubMed
Summary

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Field-Testing Multiple-Choice Questions With AI Examinees: English Grammar Items.

Educational and psychological measurement·2024
See all related articles

This study introduces a multimodal deep learning model to predict educational item parameters using text, images, and metadata. While versatile, the model

Area of Science:

  • Educational Measurement
  • Artificial Intelligence
  • Computer Science

Background:

  • Traditional item parameter prediction models often require specialized architectures for different item types and data modalities.
  • Integrating diverse data sources like text, images, and metadata presents a challenge for unified prediction models.

Purpose of the Study:

  • To develop and evaluate a flexible multimodal deep learning model for predicting dichotomous and polytomous item parameters.
  • To assess the model's ability to fuse representations from vision and language transformers for heterogeneous item formats.
  • To investigate the contribution of different data modalities (text, images, metadata) to prediction accuracy.

Main Methods:

  • A deep learning model fusing Transformer vision and language models was employed.
Keywords:
DINOv3DeBERTaV3artificial intelligenceitem response theorylarge-scale assessmentnatural language processingquestion difficulty prediction

Related Experiment Videos

  • The model accommodates heterogeneous item formats and uses attention pooling to weight component importance.
  • Item parameters were predicted jointly using the two-parameter logistic and generalized partial credit models with a masking strategy.
  • Item-level and component-level metadata were incorporated into the model.

    • Main Results:

    • A single model successfully predicted parameters for English language arts and mathematics items across 11 types.
    • The full model did not consistently leverage all input data, with feature removal often not impacting accuracy.
    • Images were strong predictors independently but did not consistently add unique information when combined with text and metadata.
    • The most parsimonious model achieved R² values up to .74 for various item parameters.

    Conclusions:

    • Current training methods may limit the learnability of complex multimodal deep fusion models.
    • While promising, the model's ability to fully integrate multimodal data requires further investigation.
    • The findings suggest potential for unified models in educational measurement, but optimization is needed.