Do Transformer Models Show Similar Attention Patterns to Task-Specific Human Gaze?

TL;DR

This study compares the attention patterns of large-scale pre-trained language models with human gaze during task reading, revealing that their correlation depends on rare context features and is unaffected by task-specific fine-tuning.

Contribution

It provides a systematic comparison of model attention and human gaze, highlighting the influence of rare contexts and the limited impact of fine-tuning on attention alignment.

Findings

Model attention correlates with human gaze more in rare, syntactic contexts.

Fine-tuning does not significantly improve the alignment with human attention.

Lower-entropy attention vectors are more faithful to human gaze patterns.

Abstract

Learned self-attention functions in state-of-the-art NLP models often correlate with human attention. We investigate whether self-attention in large-scale pre-trained language models is as predictive of human eye fixation patterns during task-reading as classical cognitive models of human attention. We compare attention functions across two task-specific reading datasets for sentiment analysis and relation extraction. We find the predictiveness of large-scale pre-trained self-attention for human attention depends on `what is in the tail', e.g., the syntactic nature of rare contexts. Further, we observe that task-specific fine-tuning does not increase the correlation with human task-specific reading. Through an input reduction experiment we give complementary insights on the sparsity and fidelity trade-off, showing that lower-entropy attention vectors are more faithful.