Reproduce “CorefInst: Leveraging LLMs for Multilingual Coreference Resolution” 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

Tuğba Pamay Arslan, Emircan Erol, Gülşen Eryiğit

Researchers at Istanbul Technical University developed CorefInst, the first instruction-tuned Large Language Model for multilingual coreference resolution, which specifically addresses both overt and zero mentions. The model achieved a 1.8 percentage point improvement in CoNLL score over the state-of-the-art CorPipe24 model and an 8.4 percentage point gain in F1 score for zero mention resolution across multiple languages.

Key insights

  • Instruction-tuned decoder-only LLMs can effectively learn and perform complex multilingual coreference resolution, including the challenging task of zero mention identification.
  • LLMs, through their extensive pre-training and instruction-based fine-tuning, exhibit an enhanced capacity for discerning deep contextual and implicit semantic relations essential for zero mention resolution.
  • The use of instruction engineering and controlled inference methods offers a viable pathway to adapt powerful general-purpose LLMs for specific NLP tasks, potentially replacing labor-intensive task-specific architectures.

Summary by alphaXiv.

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