An Eye Opener Regarding Task-Based Text Gradient Saliency

Anonymous

Abstract: Eye movements in reading reveal humans' cognitive processes during language understanding. As such, the time a reader's eyes dwell on a token has been utilized as a measure for the visual attention paid to that word, or the importance of that word to the reader. This study investigates the alignment of the importance attributed to input tokens by language models (LMs) on the one hand and humans, in the form of fixation durations, on the other hand. While previous research on the internal processes of LMs have employed the models' attention weights, recent studies have argued in favor of gradient-based methods. Moreover, previous approaches to interpret LMs' internals with human gaze have neglected the tasks readers performed during reading, even though psycholinguistic research underlines that reading patterns are task-dependent. We thus introduce a novel approach that employs a gradient-based saliency method designed to emulate task-specific human reading strategies to align model and human importance, and we find that task specificity plays a crucial role in this alignment.

Paper Type: short

Research Area: Interpretability and Analysis of Models for NLP

Contribution Types: Model analysis & interpretability

Languages Studied: English

Preprint Status: There is no non-anonymous preprint and we do not intend to release one.

A1: yes

A1 Elaboration For Yes Or No: Section Limitations

A2: yes

A2 Elaboration For Yes Or No: Section Ethics Statement

A3: yes

A3 Elaboration For Yes Or No: Abstract and Section 1

B: yes

B1: yes

B1 Elaboration For Yes Or No: Section 4 and Appendix A

B2: yes

B2 Elaboration For Yes Or No: Section Ethics Statement

B3: yes

B3 Elaboration For Yes Or No: Section 3 and 4

B4: no

B4 Elaboration For Yes Or No: The creators of the artifacts have discussed about the steps to protect personal information.

B5: yes

B5 Elaboration For Yes Or No: Section 4 and Limitation; Appendix A

B6: yes

B6 Elaboration For Yes Or No: Appendix A

C: yes

C1: no

C1 Elaboration For Yes Or No: Our computational experiments are not resource-intensive and can be completed in under 2 hours using a single GPU.

C2: yes

C2 Elaboration For Yes Or No: Appendix A

C3: yes

C3 Elaboration For Yes Or No: Section 5 and 6; Appenddix A and B

C4: yes

C4 Elaboration For Yes Or No: Section 4, Appendix A

D: no

E: yes

E1: no

E1 Elaboration For Yes Or No: ChatGPT is utilized for code explanation purposes, while Copilot aids in inline code completion.