Reproduce “Retrofitting Multilingual Sentence Embeddings with Abstract Meaning Representation” 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

Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, Wai Lam

Researchers from The Chinese University of Hong Kong and DAMO Academy, Alibaba Group, developed a method to retrofit multilingual sentence embeddings using Abstract Meaning Representation (AMR). Their approach, which integrates AMR-derived explicit semantic information, consistently improved state-of-the-art performance across 10 semantic textual similarity datasets and five diverse cross-lingual transfer tasks, while also proposing a more comprehensive evaluation standard for such embeddings.

Key insights

  • Integrating explicit semantic information from AMR consistently enhances the performance of state-of-the-art multilingual sentence embeddings on both semantic similarity and diverse transfer tasks.
  • A comprehensive evaluation benchmark including various transfer tasks is crucial, as strong STS performance does not guarantee utility for all downstream applications.
  • The effectiveness of AMR retrofitting relies heavily on robust multilingual AMR parsing and extensive language coverage in the AMR encoder's training data, particularly for distant languages.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.