A study on surprisal and semantic relatedness for eye-tracking data prediction

Area of Science:

• Computational Linguistics
• Cognitive Science
• Natural Language Processing

Background:

• Previous studies explored language modeling and distributional semantics for predicting eye-tracking metrics.
• The distinct contributions of surprisal and semantic relatedness remain unclear, with some research suggesting surprisal subsumes relatedness.

Purpose of the Study:

• To investigate the independent and combined predictive power of surprisal and semantic relatedness on eye-tracking metrics.
• To compare different types of relatedness scores from static and contextual models.
• To evaluate the impact of BERT contextual embeddings on prediction accuracy.

Main Methods:

• A regression experiment was conducted on two English corpora.
• Eye-tracking metrics were predicted using language model surprisal and various semantic relatedness scores.
• Predictions were compared with and without surprisal and relatedness components.
• Static and contextual embeddings (including BERT) were utilized for relatedness scores.

Main Results:

• Both surprisal and semantic relatedness significantly contribute to predicting eye-tracking metrics.
• Semantic relatedness unexpectedly aids in predicting function word processing.
• BERT contextual embeddings yielded higher prediction accuracy, explaining more variance.

Conclusions:

• Language model surprisal and semantic relatedness offer distinct and complementary information for eye-tracking prediction.
• Contextual embeddings, particularly from BERT, enhance the predictive power of semantic relatedness.
• Future research should leverage both components for more robust eye-movement modeling.