Crosslingual Coreference

Coreference is amazing but the data required for training a model is very scarce. In our case, the available training for non-English languages also proved to be poorly annotated. Crosslingual Coreference, therefore, uses the assumption a trained model with English data and cross-lingual embeddings should work for languages with similar sentence structures.

Install

pip install crosslingual-coreference

Quickstart

from crosslingual_coreference import Predictor

text = (
    "Do not forget about Momofuku Ando! He created instant noodles in Osaka. At"
    " that location, Nissin was founded. Many students survived by eating these"
    " noodles, but they don't even know him."
)

# choose minilm for speed/memory and info_xlm for accuracy
predictor = Predictor(
    language="en_core_web_sm", device=-1, model_name="minilm"
)

print(predictor.predict(text)["resolved_text"])
print(predictor.pipe([text])[0]["resolved_text"])
# Note you can also get 'cluster_heads' and 'clusters'
# Output
#
# Do not forget about Momofuku Ando!
# Momofuku Ando created instant noodles in Osaka.
# At Osaka, Nissin was founded.
# Many students survived by eating instant noodles,
# but Many students don't even know Momofuku Ando.

Models

As of now, there are two models available "spanbert", "info_xlm", "xlm_roberta", "minilm", which scored 83, 77, 74 and 74 on OntoNotes Release 5.0 English data, respectively.

• The "minilm" model is the best quality speed trade-off for both mult-lingual and english texts.
• The "info_xlm" model produces the best quality for multi-lingual texts.
• The AllenNLP "spanbert" model produces the best quality for english texts.

Chunking/batching to resolve memory OOM errors

from crosslingual_coreference import Predictor

predictor = Predictor(
    language="en_core_web_sm",
    device=0,
    model_name="minilm",
    chunk_size=2500,
    chunk_overlap=2,
)

Use spaCy pipeline

import spacy

text = (
    "Do not forget about Momofuku Ando! He created instant noodles in Osaka. At"
    " that location, Nissin was founded. Many students survived by eating these"
    " noodles, but they don't even know him."
)

nlp = spacy.load("en_core_web_sm")
nlp.add_pipe(
    "xx_coref", config={"chunk_size": 2500, "chunk_overlap": 2, "device": 0}
)

doc = nlp(text)
print(doc._.coref_clusters)
# Output
#
# [[[4, 5], [7, 7], [27, 27], [36, 36]],
# [[12, 12], [15, 16]],
# [[9, 10], [27, 28]],
# [[22, 23], [31, 31]]]
print(doc._.resolved_text)
# Output
#
# Do not forget about Momofuku Ando!
# Momofuku Ando created instant noodles in Osaka.
# At Osaka, Nissin was founded.
# Many students survived by eating instant noodles,
# but Many students don't even know Momofuku Ando.
print(doc._.cluster_heads)
# Output
#
# {Momofuku Ando: [5, 6],
# instant noodles: [11, 12],
# Osaka: [14, 14],
# Nissin: [21, 21],
# Many students: [26, 27]}

Visualize spacy pipeline

This only works with spacy >= 3.3.

import spacy
from spacy.tokens import Span
from spacy import displacy

text = (
    "Do not forget about Momofuku Ando! He created instant noodles in Osaka. At"
    " that location, Nissin was founded. Many students survived by eating these"
    " noodles, but they don't even know him."
)

nlp = spacy.load("nl_core_news_sm")
nlp.add_pipe("xx_coref", config={"model_name": "minilm"})
doc = nlp(text)
spans = []
for idx, cluster in enumerate(doc._.coref_clusters):
    for span in cluster:
        spans.append(
            Span(doc, span[0], span[1]+1, str(idx).upper())
        )

doc.spans["custom"] = spans

displacy.render(doc, style="span", options={"spans_key": "custom"})

More Examples