Simpler installation which should hopefully be easier to maintain compatible with SpaCy from now on.
Only use one model and only store/download/cache the coreference resolution model and not the SpaCy models anymore: 10x smaller to download and store
Compatible w. SpaCy 2.1+: we probably won't seek backward compatibility for older versions of SpaCy since the dependency structure has been cleaned up a lot in SpaCy 2.1 (impressive job by Inès and Matthew! Read about it here)
New feature to incorporate domain knowledge: conversion dictionary (see details in the readme).
3.0
5 years ago
Code moved to Cython and Thinc, 50-100x faster. Currently around 2kWPS.
Removed speakers handling (prevented tighter integration with spaCy)
Re-factored and simplified code
Updated documentation
2.0
6 years ago
You can now train Neuralcoref on your own dataset , e.g. other languages! 🎉🎉
Released our training code + updated code for compatibility with training code