Using Coreference Resolver with [spacy-experimental]

Coreference?

Coreference is one the most common practices we do every time (at least for many western language users). Literally, it is to find out what multiple words (or phrases) represent in the text. That technique is so crucial as we often (and always in written texts) change the same word to other equivalent synonyms to avoid repetitions (Have you ever heard of Thesaurus dictionary? Of course you have!)

As you know well, the concept is as follows.

Yesterday, **Google** announced its own **_AI chatbot_**, _**Bard**_, _**a competitor**_ to ChatGPT developed by OpenAI.
However, **the tech giant** embarrassed itself by sharing an inaccurate information generated with the new platform.
As a result, **the company**'s stock plunged pretrading before recouping **its** losses during the day.

Here, all bold w/o italics mean Google while bold w/ italics do Bard.

For this simple example, it seems fairly easy, maybe because we know the mechanism intrinsically or because the text does not involve a too complicate structure complicating the process. Whatever is true, perfect coreference with machine has been an illusion in the NLP domain.

Why spacy-experimental?

Disclaimer: I'm not an employee of Explosion.AI or a contributor to spaCy.

Since I came across spaCy and after its release of Korean language model pipleine, I never looked back. Until now, I've been pleased with stable version (not experimental one). In terms of Coreference, various solutions or (Python) modules have been proposed.

From my own experience, most weren't for me.

Putting asise the quality of the result, most critical snag was the version compatibility of spaCy. Most tools depend on both AllenNLP and spaCy. However, most rely on previous versions of spaCy, some of which are caused by their reliance on previous versions of AllenNLP. As I needed spaCy version 3.3 at the least, most of them were out of the option.

Then, with version 0.6, spacy-experimental started to offer Coreference Resolver with its transformer-based pipeline. Upon testing on some texts, I was quickly into it.

How to use?

Installation

Obviously, installing spacy-experimental is required. Even spaCy hasn't been installed, the installer will install spaCy as one of dependencies. Also, model pipeline capable of Coreference Resolver is needed as well (or you can train it but I do not recommend).

pip install spacy-experimental
pip install https://github.com/explosion/spacy-experimental/releases/download/v0.6.1/en_coreference_web_trf-3.4.0a2-py3-none-any.whl

Action!

Note that we do NOT IMPORT spacy-experimental, rather indirectly by importing spacy as usual. Only difference is the model to load. Also, to check the result, see contents of doc.spans.

import spacy
coref = spacy.load('en_coreference_web_trf')
text = "Yesterday, Google announced its own AI chatbot, Bard, a competitor to ChatGPT, developed by OpenAI. However, the tech giant embarrassed itself by sharing an inaccurate information generated with the new platform. As a result, the company's stock plunged pretrading before recouping its losses during the day."
doc = coref(text)
print(doc.spans)

And here is the result.

{'coref_clusters_1': [Google, its, the tech giant, itself, the company's, its], 'coref_clusters_2': [its own AI chatbot, Bard, a competitor to ChatGPT, developed by OpenAI, the new platform]}

As you can see, the resolved coreferences are in the same coref_clusters_{number}. In other words, the model concludes that items within a cluster represents the same. For coref_clusters_1, Google, its, the tech giant, itself, the company's, its and for coref_clusters_2, its own AI chatbot, Bard, a competitor to ChatGPT, developed by OpenAI, the new platform .

Looks good, isn't it?

Overall, it seems pretty decent. The only drags are its limited support of language and heavy model (transformer-trained).