The democratization of NLP and what it means for the industry

The developments of the recent two or three years have moved the field quite significantly. There was a lot of focus in the community on the Transformer architecture, with a lot of results and implications. Low-resource tasks became more tractable due to a wide variety of available pretrained models including cross-lingual ones. Reproducibility is in a much better place now than it was five years before, and the list goes on.

This will be a friendly and not too formal discussion – let’s exchange thoughts and talk. From the perspective of building NLP products, let’s touch on how different aspects were affected in practice or whether they are likely to be affected in the near future:

• Building and productizing NLP models
• Time/effort to find a suitable approach (research)
• Time/effort to build something workable (PoC)
• Time/effort to make it ready for users
• Engineering/science roles and how things they do have changed
• What barriers are still there to overcome for NLP to be broadly available, especially from the application side

The discussion will be held in English.

Moderator:

Dmitry Unkovsky (Lead ML Engineer at Grammarly, GitHub, Linkedin, Facebook)

Experts:

• Chad Mills (Leads Grammarly's Applied Research)
• Thomas Wolf (Co-founder and CEO at Hugging Face Inc.)
• Vitaliy Radchenko (Data Scientist at YouScan)