Unsupervised Text Segmentation Predicts Eye Fixations During Reading.

Area of Science:

• Psycholinguistics
• Computational Linguistics
• Cognitive Science

Background:

• Traditional studies in psycholinguistics and computational linguistics focus on words as the basic units of sentence processing.
• Recent research suggests that the mental lexicon may also contain sub-word and supra-word units, requiring cognitive mechanisms for their detection.
• Human readers possess a cognitive mechanism to learn and detect these units during reading.

Purpose of the Study:

• To investigate whether cognitive units during reading are exclusively words or if they encompass sub-word and supra-word segments.
• To determine if unsupervised text segmentation models can identify cognitive units relevant to reading.
• To test the hypothesis that eye fixation locations during reading correspond to these cognitive units.

Main Methods:

• Eye movement data from English and Dutch reading tasks were analyzed.
• Unsupervised text segmentation models were employed to identify potential cognitive units.
• The predictive accuracy of model-segmented units on eye fixation locations was compared against word-based units.
• The Less-is-Better (LiB) model was utilized to assess memory load minimization in unit identification.

Main Results:

• Model-segmented units demonstrated superior prediction of eye fixations compared to traditional word units.
• The Less-is-Better (LiB) model showed advantages in both prediction score and computational efficiency.
• The findings support the principle of least effort in managing long-term and working memory during reading unit inference.

Conclusions:

• The mental lexicon stores units beyond single words, including smaller and larger segments.
• Eye fixation patterns during reading are influenced by these discovered cognitive units.
• Unsupervised segmentation models are effective tools for identifying these cognitive units, advancing our understanding of reading processes.