Recognition language classifiers demonstrate far transfer of learning.

Area of Science:

• Cognitive Psychology
• Machine Learning
• Forensic Science

Background:

• Machine learning models trained on verbal justifications of recognition decisions can predict accuracy.
• The generalization capability of these models, termed recognition language classifiers, is crucial for broader application.
• Testing generalization requires assessing performance on tasks dissimilar to the training paradigms.

Purpose of the Study:

• To determine if recognition language classifiers, trained on single-item verbal recognition tasks, generalize to predict eyewitness accuracy in lineup selections.
• To investigate the recollection sensitivity of these classifiers by testing their ability to differentiate accurate from inaccurate eyewitness accounts.
• To assess the convergent and divergent validity of the classifiers concerning the recollection sensitivity hypothesis.

Main Methods:

• Three machine learning classifiers were trained to distinguish verbal justifications in different recognition paradigms (hits vs. false alarms, confidence levels, remember/know judgments).
• These trained classifiers were then applied to score language from eyewitness lineup selections (correct vs. incorrect).
• Mixed modeling analysis was used to evaluate classifier performance and validity concerning eyewitness accuracy and confidence.

Main Results:

• All three classifiers reliably predicted eyewitness accuracy (correct vs. incorrect selections) despite significant differences in materials, populations, and procedures compared to training.
• Classifiers showed strong predictive power for eyewitness selections (hits vs. false alarms) but not for rejections (correct rejections vs. misses), supporting convergent validity.
• One classifier predicted eyewitness confidence, even without metacognitive training data, indicating sensitivity to underlying memory processes.

Conclusions:

• Recognition language classifiers demonstrate robust generalization capabilities, extending beyond initial training paradigms to diverse real-world scenarios like eyewitness testimony.
• The findings support the hypothesis that these classifiers capture recollection-based information embedded in language used to justify memory decisions.
• The study highlights the potential of machine learning-based linguistic analysis as a tool for evaluating the reliability of eyewitness accounts.