State-of-the-art coreference resolution librairy using neural nets and spaCy. Try it online !
This coreference resolution module is based on the super fast spaCy parser and uses the scoring neural model described in Deep Reinforcement Learning for Mention-Ranking Coreference Models by Kevin Clark and Christopher D. Manning, EMNLP 2016.
Be sure to check out our medium post where we talk more about neuralcoref and coreference resolution.
Clone the repo (the trained model weights are too large for PyPI)
cd neuralcoref
pip install -r requirements.txt
You will also need an english model for spaCy if you don't already have spaCy installed.
python -m spacy download 'en'
The mention extraction module strongly depends on the quality of the parsing so we recommand selecting a model with a higher accuray than usual. Since the coreference algorithm don't make use of spaCy's word vectors, a medium size spaCy model like spacy's 'en_depent_web_md' model is a good balance between memory footprint and parsing accuracy.
To download and install it use
python -m spacy download 'en_depent_web_md'
python -m spacy link en_depent_web_md en
If you are an early user of spacy 2 alpha, you can use neuralcoref with spacy 2 without any specific modification.
python -m neuralcoref.server starts a wsgiref simple server.
You can retreive coreferences on a text or dialogue utterances by sending a GET request on port 8000.
Example
curl -G "http://localhost:8000/" --data-urlencode "text=My sister has a dog. She loves him so much"
Please refer to the the source for details on the various functions and arguments.
The main class you can import is neuralcoref.Coref.
It offers two main functions for resolving coreference:
one_shot_coref()for a single coreference resolution operation.continuous_coref()for continous coreference resolution (during an on-going dialogue for instance).
The source code contains details of the various arguments you can use. The code of server.py also provides a simple example.
Example:
from neuralcoref import Coref
coref = Coref()
clusters = coref.one_shot_coref(u"My sister has a dog and she loves him.")
print(clusters)