Reproduce “SAP: Syntactic Attention Pruning for Transformer-based Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tzu-Yun Lee, Ding-Yong Hong, Jan-Jan Wu
SAP uses grammatical dependency links to guide which Transformer attention heads to prune, ranking heads by alignment between attention and dependencies. Authors compare retrain-free pruning on BERT-base across six GLUE tasks with a gradient-mask baseline. SAP’s performance declines more gradually as sparsity rises, especially at higher sparsity, and Candidate Filtering generally improves SAP; findings cover this model and task set.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.