spacy-coref

Lightweight, fast co-reference resolution using a distilled version of AllenNLP's coreference model (exported to ONNX).

✨ Features

• 🧠 Cross-lingual coreference resolution
• 🪶 Lightweight model based on AllenNLP’s coref modeling
• ⚡ Fast inference via ONNX
• 🔌 Easy integration with spaCy

📦 Installation

$ pip install spacy-coref

🚀 Quickstart

Usage as a standalone component

from spacy_coref import CoreferenceResolver, decode_clusters

resolver = CoreferenceResolver.from_pretrained("talmago/allennlp-coref-onnx-mMiniLMv2-L12-H384-distilled-from-XLMR-Large")

sentences = [
    ["Barack", "Obama", "was", "the", "44th", "President", "of", "the", "United", "States", "."],
    ["He", "was", "born", "in", "Hawaii", "."]
]

pred = resolver(sentences)

print(decode_clusters(sentences, pred["clusters"][0]))

# Output is:
# [['Barack Obama', 'He']]

Usage with spaCy

import spacy

from spacy_coref import create_coref_minilm_component

nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("coref_minilm")

doc = nlp("Barack Obama was born in Hawaii. He was elected president in 2008.")
print(doc._.coref_clusters[0])
print(doc._.cluster_heads)
print(doc._.resolved_text)

# Output is:
# [Barack Obama, He]
# {'Barack Obama': Barack Obama}
# Barack Obama was born in Hawaii. Barack Obama was elected president in 2008.

🛠️ Development

Set up virtualenv

$ make env

Set PYTHONPATH

$ export PYTHONPATH=$PYTHONPATH:/Users/talmago/git/spacy-coref/src

Code formatting

$ make format

📚 References

This project builds on the work of the following repositories:

• crosslingual-coreference
  David Berenstein's implementation of multilingual coreference resolution, adapted from the original AllenNLP coref model.
  GitHub: davidberenstein1957/crosslingual-coreference

• AllenNLP coreference model
  Official AllenNLP implementation of coreference resolution.
  GitHub: allenai/allennlp-models