Reference library
Attention is not Explanation
Paper summary
Tests whether attention weights explain predictions in neural language tasks. Weak agreement with gradient-based importance and alternative attention distributions giving similar predictions challenge the use of attention maps as direct causal explanations.
Relevance to the project
Supports intervention-based analysis of communication and influence in the simulation. Visible connection strengths or routing weights should be checked against changes in behavior.
Conclusions depend on the tested models and definition of explanation.
Source record
- Reference ID
- article-https-aclanthology-org-n19-1357
- Archived version
- Publisher/institution PDF advertised by article metadata
- Summary provenance
- reference-library/papers/article-https-aclanthology-org-n19-1357/metadata.json · summary
- Review status
- initial orientation summary; not a critical full-paper review
- Reading scope
- Selected abstract/opening and available closing discussion/conclusion passages in the preserved text extract; broader project notes provide context. No reproduction of experiments, exhaustive methods review, or equation/figure audit.