search

huggingface / neuralcoref

✨Fast Coreference Resolution in spaCy with Neural Networks
https://huggingface.co/coref/
MIT License
2.85k stars 476 forks source link

In what sense state of the art? #87

Closed matanox closed 5 years ago

matanox commented 6 years ago

This leaves a terrific impression, but how did you conclude this is state of the art? I think this should be noted in some prominent way, even if it is already there in the blog post!

ljinstat commented 6 years ago

The neuralcoref model is inspired by the mention-ranking model which is implemented in previous the state-of-the-art papers (mentioned in the blog and issue #71). Due to the expectation of relatively good performance and easier implementation, the neuralcoref model isn't actually implemented the same as it is mentioned in those two papers. Thus, it loses some performance without surprise. Besides, these two papers are not any more the state-of-the-art (replaced by the end-to-end neural coreference model). It depends on how you define "the-state-of-the-art". As we know, it is tough to improve the performance of the coreference model a lot. Despite of the gap between the neuralcoref model and the current best model, this neuralcoref model is great for inference though.

thomwolf commented 5 years ago

You are right, it's not state-of-the-art anymore. The AllenAI system is. I will change the wording.