Support Azure Cognitive Services in some pipelines

[danielbichuetti]

Haystack is deeply integrated with Transformers. But what are your thoughts on being able to make usage of Azure Cognitive Services as an integration in Translator, Summarizer, Entity Extractor and Classification pipeline? Azure is already used on form recognizer.

With the addition of OpenAI in the other pipeline, maybe it would be acceptable. If the user doesn't wish to migrate his existent custom “models”. They could just use haystack on the parts they need, with deep integration in Documents.

[ZanSara]

Hey @danielbichuetti! Let me ask my colleagues before proceeding with this one. Sounds like a big addition to the library! I personally see no problem with it but I'd like to make sure nobody has objections before moving on.

[tstadel]

@danielbichuetti I think we had very specific needs regarding form/table parsing which led us towards Azure Form Recognizer. Could you please elaborate more on why to use the Translator, Summarizer, Entity Extractor and Classification services from Azure Cognitive Services? Is there any functionality that you do not get from the currently implemented nodes?

[danielbichuetti]

Hi @tstadel! Indeed, our usage is pointed towards Translator, entity extractor and classification.

And currently we are getting medium translation from available transformers models for specific domain (legal). Of course, we could train one new model, but training a new model is costly regarding team time and costs. There are some custom translators already done by us for Cognitive Services.

The same apply for entity extraction and classification, since there are exclusive entities that commonly you need to tag into legal documents, and unique types of kinds of documents. For this reason, we also have custom trained endpoints on Azure. Throwing away without having at least near models would be a waste of time and work for us.

Probably there are enterprises and other companies that area already using some of these custom endpoints. If they changed their NLP pipeline to haystack framework, they would have this overlap for some time. Where they have some old endpoints and are preparing a whole new pipeline.

The main point about this integration isn't the transformers pitfalls, which are none (as long as you have a decent model in your language or enough time and resources to build a new). The main point is about make easier the transition from other stacks.

This feature isn't a problem of haystack, I would agree. But a plus. Specially when you look for adoption into already in-game companies using Azure or AWS ML stack.

Maybe this could be a custom node, which I have thought about it. But I decided to discuss here first, as I think we won't be the last to face this.