spaCy and Explosion: past, present & future

Transcript

1. spaCy and Explosion:
   past, present & future
   Matthew Honnibal

   Ines Montani
   Explosion
   

   View Slide

2. 1982 - 2014
   Before spaCy
   

   View Slide

3. 1982 - 2014
   Before spaCy
   Some time before 2005
   

   View Slide

4. 1982 - 2014
   Before spaCy
   PhD 2005-2010 (Sydney)
   Some time before 2005
   

   View Slide

5. 1982 - 2014
   Before spaCy
   PhD 2005-2010 (Sydney)
   ACL 2009 (Singapore)
   Some time before 2005
   

   View Slide

6. July 2014
   First commit on spaCy
   

   View Slide

7. Early 2015
   spaCy is first released
   

   View Slide

8. 2015
   First collaborations
   • demos & visualizers like displaCy
   • first concepts of a modern
   approach to NLP annotation tools
   

   View Slide

9. 2015
   First collaborations
   “Baskets” concept
   • demos & visualizers like displaCy
   • first concepts of a modern
   approach to NLP annotation tools
   

   View Slide

10. 2015
    First collaborations
    “Baskets” concept
    Binary annotation tool concept
    • demos & visualizers like displaCy
    • first concepts of a modern
    approach to NLP annotation tools
    

    View Slide

11. Early 2016
    First highlights: sense2vec
    

    View Slide

12. Early 2016
    First highlights: German model
    • first non-English model
    • non-projective dependencies
    • developed by Wolfgang Seeker
    

    View Slide

13. Late 2016
    Explosion
    • new company for AI developer tools
    • bootstrapped through consulting 

    for the first 6 months
    • funded through software sales 

    since 2017
    • 100% independent and profitable
    

    View Slide

14. Late 2016
    Explosion
    • new company for AI developer tools
    • bootstrapped through consulting 

    for the first 6 months
    • funded through software sales 

    since 2017
    • 100% independent and profitable
    Our bets about NLP
    • NLP won’t just be a cloud API
    • number of developers will increase
    • annotation is better in-house
    

    View Slide

15. 2017
    spaCy v2.0
    • shift to deep learning
    • smaller and updatable models
    • custom pipeline components
    • custom extension attributes
    • built-in text classification
    • built-in displaCy visualizers
    • many other improvements
    Thinc, spaCy’s machine learning library
    

    View Slide

16. July 2017
    neuralcoref by Hugging Face
    Community package for coreference

    resolution with spaCy
    

    View Slide

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

    View Slide

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

    View Slide

19. Late 2017
    Prodigy
    • first commercial product
    • modern annotation tool
    • fully scriptable in Python
    users
    2,000+
    250+ companies
    incl.
    

    View Slide

20. July 2018
    10,000 stars on GitHub
    

    View Slide

21. ,
    Early 2019
    spaCy v2.1
    • transfer learning and pretraining
    • 2-3 times faster tokenization
    • enhanced match pattern API
    • built-in rule-based NER
    • many other improvements
    

    View Slide

22. ,
    Early 2019
    spaCy v2.1
    • transfer learning and pretraining
    • 2-3 times faster tokenization
    • enhanced match pattern API
    • built-in rule-based NER
    • many other improvements
    Transfer learning
    • better models with less data –
    huge win!
    • how to adapt for spaCy without
    bigger (and slower) models?
    • spacy pretrain is a pretty cool
    compromise
    

    View Slide

23. April 2019
    Free interactive online course
    course.spacy.io
    

    View Slide

24. So what’s next 

    for spaCy?
    

    View Slide

25. July 2019
    Explosion team
    Matthew
    Ines Montani
    Honnibal
    

    View Slide

26. July 2019
    Explosion team
    Matthew
    Ines Montani
    Justin DuJardin
    Honnibal
    

    View Slide

27. July 2019
    Explosion team
    Matthew
    Ines Montani
    Sebastián Ramírez Guadalupe Romero
    Giannis Daras
    Justin DuJardin
    Van Landeghem
    Sofie
    Honnibal
    

    View Slide

28. Today
    spaCy IRL
    • 2 sold-out corporate training days
    • 200+ conference attendees
    

    View Slide

29. What’s next?
    spaCy v3.0
    • morphological features
    • entity linking
    • non-entity span tagging
    • static analysis of processing
    pipeline and its components
    

    View Slide

30. What’s next?
    spaCy v3.0
    Vision for spaCy
    • focus on data structures and pipeline
    • build support for new tasks even if we
    don’t have a model
    • make sure it’s easy to BYO model
    • keep shipping good defaults
    • morphological features
    • entity linking
    • non-entity span tagging
    • static analysis of processing
    pipeline and its components
    

    View Slide

31. What’s next?
    spaCy v3.0
    Vision for spaCy
    • focus on data structures and pipeline
    • build support for new tasks even if we
    don’t have a model
    • make sure it's easy to BYO model
    • keep shipping good defaults
    What’s out-of-scope?
    • anything generative: summarization, machine
    translation, etc.
    • multi-modal: audio, video, etc.
    • research assistance: plenty of good
    frameworks for developing novel techniques
    • morphological features
    • entity linking
    • non-entity span tagging
    • static analysis of processing
    pipeline and its components
    

    View Slide

32. ,
    What’s next?
    spaCy ecosystem in your cloud
    • whole systems, not just libraries
    • programmable, extensible cluster
    • running under your control
    • automated setup, good defaults
    • full data privacy – we don’t want
    your data!
    

    View Slide

33. ,
    What’s next?
    spaCy ecosystem in your cloud
    • whole systems, not just libraries
    • programmable, extensible cluster
    • running under your control
    • automated setup, good defaults
    • full data privacy – we don’t want
    your data!
    Processing with Dask
    

    View Slide

34. ,
    What’s next?
    spaCy ecosystem in your cloud
    • whole systems, not just libraries
    • programmable, extensible cluster
    • running under your control
    • automated setup, good defaults
    • full data privacy – we don’t want
    your data!
    Prodigy Scale
    Processing with Dask
    

    View Slide

35. commits
    10,000+
    contributors
    300+
    60+ extension packages
    13,500+ GitHub stars
    2,800+ closed issues
    80+ releases
    

    View Slide

36. Thank you!
    

    View Slide