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

```python 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="encorewebsm", device=-1, modelname="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", "infoxlm", "xlmroberta", "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

```python from crosslingual_coreference import Predictor

predictor = Predictor( language="encorewebsm", device=0, modelname="minilm", chunksize=2500, chunkoverlap=2, ) ```

Use spaCy pipeline

```python 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("encorewebsm") nlp.addpipe( "xxcoref", config={"chunksize": 2500, "chunk_overlap": 2, "device": 0} )

doc = nlp(text) print(doc..corefclusters)

Output

[[[4, 5], [7, 7], [27, 27], [36, 36]],

[[12, 12], [15, 16]],

[[9, 10], [27, 28]],

[[22, 23], [31, 31]]]

print(doc..resolvedtext)

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..clusterheads)

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. ```python 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("nlcorenewssm") nlp.addpipe("xxcoref", config={"modelname": "minilm"}) doc = nlp(text) spans = [] for idx, cluster in enumerate(doc..corefclusters): 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