Attention-aware semantic relevance predicting Chinese sentence reading.

Area of Science:

• Computational linguistics
• Cognitive science
• Natural language processing

Background:

• Existing computational models predict human sentence comprehension and processing.
• Contextual semantic similarity is a promising approach.
• Transformer attention algorithms and human memory mechanisms inspire new models.

Purpose of the Study:

• Propose an attention-aware approach for computing contextual semantic relevance.
• Incorporate contextual information fully by considering contributions of different parts and expectation effects.
• Evaluate the new approach against existing reading models and eye-tracking data.

Main Methods:

• Developed an attention-aware approach inspired by Transformer attention and human memory.
• Computed contextual semantic relevance metrics.
• Utilized an eye-tracking corpus of Chinese reading tasks.
• Compared attention-aware metrics with existing approaches for predicting fixation durations.

Main Results:

• Attention-aware semantic relevance metrics more accurately predict fixation durations in Chinese reading tasks.
• Findings support the presence of semantic preview benefits in naturalistic Chinese reading.
• The approach facilitates simulation and evaluation of existing reading models.

Conclusions:

• Attention-aware metrics offer a valuable, interpretable computational tool for modeling eye-movements in reading.
• These metrics advance understanding of language comprehension and processing.
• The approach highlights the potential of memory-based metrics in cognitive and linguistic research.