Generative Coreference Resolution

The final project of the course AMNLP (Advanced methods in NLP). Shimon Malnick and I tried to solve the CR resolution task on the OntoNotes benchmark by fine-tuning BART and T5. Our best model achieved an average F1 score of 36.8% on CoNLL, while limiting the input length (due to resource constraints). Our analysis provides a few explanations for our low performance.

More details and code can be found in our Github repository and report.