Statistical models of morphology predict eye-tracking measures during visual word recognition

Area of Science:

• Cognitive Psychology
• Computational Linguistics
• Psycholinguistics

Background:

• Human word recognition involves complex morphological processing.
• Statistical models offer computational approaches to understanding this process.
• Different representational units (holistic vs. decomposition) in models yield varying predictive power.

Purpose of the Study:

• To evaluate statistical models of morphology based on different representational units for predicting human word recognition.
• To investigate the performance of these models at early versus late stages of word recognition.
• To determine the influence of task demands on morphological processing.

Main Methods:

• Eye-tracking was used during two tasks: a lexical decision task and a multi-word recognition task.
• Statistical models emphasizing holistic units versus decomposition were compared.
• Performance was assessed at early and late stages of word recognition.

Main Results:

• Morfessor models, which segment words at morpheme boundaries, performed well at both early and late stages in the lexical decision task.
• Models based on full word forms performed better in late than early stages.
• The multi-word recognition task revealed that early and late processing involve morphological constituents, with late stages including morpheme prediction.

Conclusions:

• Models that allow for morphological decomposition better capture human word recognition processes than those based solely on full forms.
• Task demands are crucial for understanding morphological processing, as they influence the reliance on decomposition versus holistic processing.
• Findings support dual or multiple-route cognitive models of morphological processing.