Designing Practical NLP Solutions

Video: https://www.youtube.com/watch?v=JpkzK58lkmA

Transcript

1. Designing
   practical
   Ines Montani
   Explosion
   NLP solutions
   

   View Slide

2. Early 2015
   spaCy is first released
   • open-source library for industrial-
   strength Natural Language
   Processing
   • focused on production use
   

   View Slide

3. Early 2015
   spaCy is first released
   • open-source library for industrial-
   strength Natural Language
   Processing
   • focused on production use
   Current stats
   17m+ total downloads
   16k+ stars on GitHub
   400+ contributors
   80+ extension packages
   

   View Slide

4. Late 2016
   Explosion
   • new company for AI developer tools
   • bootstrapped through consulting
   for the first 6 months
   • funded through software sales
   since 2017
   • remote team, centered in Berlin
   

   View Slide

5. Late 2016
   Explosion
   • new company for AI developer tools
   • bootstrapped through consulting
   for the first 6 months
   • funded through software sales
   since 2017
   • remote team, centered in Berlin
   Current stats
   8 team members
   100% independent & profitable
   

   View Slide

6. Late 2017
   Prodigy
   • first commercial product
   • modern annotation tool
   • fully scriptable in Python
   

   View Slide

7. Late 2017
   Prodigy
   • first commercial product
   • modern annotation tool
   • fully scriptable in Python
   Current stats
   4000+ users, including
   500+ companies
   1600+ forum members
   

   View Slide

8. Coming soon
   • spaCy v2.3: Models for Chinese,
   Japanese and many more
   • spaCy v3.0: Transformer-based
   pipelines, custom models using
   any library, new training workflow
   • Prodigy v1.10: Dependencies &
   relation annotation, audio & video
   annotation & lots of new features
   • Prodigy Teams: Manage large
   annotation projects in your cloud
   

   View Slide

9. NLP project are
   like start-ups:
   they fail a lot
   

   View Slide

10. How to maximize your
    project’s risk of failure
    

    View Slide

11. How to maximize your
    project’s risk of failure
    Imagineer.
    Forecast.
    Outsource.
    Wire.
    Ship.
    1
    2
    3
    4
    5
    Decide what your
    application ought to do.
    Be ambitious! Nobody
    changed the world saying
    “uh, will that work?”
    

    View Slide

12. Imagineer.
    Forecast.
    Outsource.
    Wire.
    Ship.
    How to maximize your
    project’s risk of failure
    1
    2
    3
    4
    5
    Figure out
    what accuracy
    you’ll need. If you’re
    not sure here, just
    say 90%.
    

    View Slide

13. How to maximize your
    project’s risk of failure
    1
    2
    3
    4
    5
    Imagineer.
    Forecast.
    Outsource.
    Wire.
    Ship.
    Pay
    someone else to
    gather your data.
    Think carefully about your
    accuracy requirements,
    and then ask for 10k rows.
    

    View Slide

14. Imagineer.
    Forecast.
    Outsource.
    Wire.
    Ship.
    How to maximize your
    project’s risk of failure
    1
    2
    3
    4
    5
    Implement your
    network. This is the
    fun part! Tensor all your
    flows, descend every
    gradient!
    

    View Slide

15. How to maximize your
    project’s risk of failure
    1
    2
    3
    4
    5
    Imagineer.
    Forecast.
    Outsource.
    Wire.
    Ship.
    Put it all
    together. If it doesn’t
    work, maybe blame
    the intern?
    

    View Slide

16. Failure
    sucks
    

    View Slide

17. accuracy
    estimate
    training &
    evaluation
    labelled
    data
    annotation
    scheme
    product
    vision
    A difficult
    chicken-and-
    egg problem
    

    View Slide

18. You need to
    iterate on your code
    and your data
    

    View Slide

19. Requirements
    We’re building a crime database based on news
    reports. We want to label the following:
    victim name
    perpetrator name
    crime location
    offence date
    arrest date
    #1
    

    View Slide

20. View Slide

21. View Slide

22. View Slide

23. Requirements
    We’re adding data from financial news about
    company sales to our internal database, so we can
    connect it to our analytics.
    We need to extract:
    buyer (official company name) and stock ticker
    acquired company with stock ticker
    sale price and currency
    #2
    

    View Slide

24. pytorch predict company acquisitions with prices and stock tickers
    No results.
    

    View Slide

25. “Microsoft acquires software development
    platform GitHub for $7.5 billion”
    

    View Slide

26. “Microsoft acquires software development
    platform GitHub for $7.5 billion”
    

    View Slide

27. TEXT CLASSIFIER
    “Microsoft acquires software development
    platform GitHub for $7.5 billion”
    

    View Slide

28. TEXT CLASSIFIER
    ENTITY RECOGNIZER
    “Microsoft acquires software development
    platform GitHub for $7.5 billion”
    

    View Slide

29. TEXT CLASSIFIER
    ENTITY RECOGNIZER
    ENTITY LINKER
    “Microsoft acquires software development
    platform GitHub for $7.5 billion”
    

    View Slide

30. TEXT CLASSIFIER
    ENTITY RECOGNIZER
    ENTITY LINKER
    ATTRIBUTE LOOKUP
    “Microsoft acquires software development
    platform GitHub for $7.5 billion”
    

    View Slide

31. TEXT CLASSIFIER
    ENTITY RECOGNIZER
    ENTITY LINKER
    ATTRIBUTE LOOKUP
    CURRENCY NORMALIZER
    “Microsoft acquires software development
    platform GitHub for $7.5 billion”
    

    View Slide

32. Reality is not
    an end-to-end
    prediction problem
    

    View Slide

33. The great thing about practical NLP:
    you can choose to make the problem
    simpler and the solution cheaper.
    #1
    

    View Slide

34. The most interesting problems are
    very specific and also need specific
    solutions. That’s what makes them
    valuable.
    #2
    

    View Slide

35. Transfer learning means we don’t
    always need “big data” anymore.
    But we need some.
    #3
    

    View Slide

36. Thank
    Explosion
    explosion.ai
    Twitter
    @_inesmontani
    you!
    

    View Slide