A Natural Language Pipeline

Presentation from the spaCy IRL 2019 conference.

Transcript

1. A Natural Language Pipeline
   

   View Slide

2. More Input
   

   View Slide

3. Knowledge”
   “A compendium
   of human...
   

   View Slide

4. Library
   

   View Slide

5. Physical archives became digital records,
   encoded with metadata
   

   View Slide

6. The internet promised rich dynamic experiences
   

   View Slide

7. The internet promised rich dynamic experiences
   but served us banner ads
   

   View Slide

8. Advertising has and continues to fuel a substantial
   portion of the innovation on the internet
   

   View Slide

9. What would The Economist look like if it were
   founded in 2012?
   

   View Slide

10. User
    

    View Slide

11. First
    

    View Slide

12. Experience
    

    View Slide

13. “There’s a reason that tech companies are topping the
    lists of most valuable companies and brands. Every
    company is a tech company.”
    Maggie Chan Jones
    

    View Slide

14. Every story, at its core, is a business story
    

    View Slide

15. Language
    

    View Slide

16. View Slide

17. View Slide

18. Stage -> Stenographer -> Editors -> spaCy -> Data Store <-> Backend <- Slack <- Users
    Proto-Pipeline
    

    View Slide

19. Over eight hours we created data from the content
    of the event, building the model in real-time
    

    View Slide

20. The model evolved over time
    

    View Slide

21. This was the experiment that would evolve into SiO
    2
    

    View Slide

22. Silicon, a key element in everything from glass to
    microchips, is at the core of global business
    

    View Slide

23. Oxygen, the journalistic voice Quartz breathes into
    the global business news cycle
    

    View Slide

24. Entities are linguistic anchors, defined by context and
    around which context can be inferred
    

    View Slide

25. Standard Entities
    PERSON FACILITY
    ORG PRODUCT
    GPE EVENT...
    Additional Entities
    TECHNOLOGY PROCESS
    NATURE MEDIA
    CONSTRUCT
    

    View Slide

26. 70K articles
    1.4M blocks of text
    85K labeled sentences
    

    View Slide

27. Entities
    

    View Slide

28. This spaCy model made rich analysis for any given
    text easy to do on the fly
    

    View Slide

29. Stored analysis of a large corpus is a vital resource
    

    View Slide

30. The language graph...
    

    View Slide

31. Graph
    

    View Slide

32. The language graph is a mutable map of the language
    model
    

    View Slide

33. Any new content is analyzed and then mapped onto
    the language graph
    

    View Slide

34. Changes made to the graph can then be incorporated
    into the next model iteration
    

    View Slide

35. The language graph becomes a primary resource for
    extracting training data
    

    View Slide

36. Snapshots of time can be extracted from the
    language graph
    

    View Slide

37. Context can be derived by looking at the relationships
    in the language graph
    

    View Slide

38. Elon Musk
    

    View Slide

39. Jeff Bezos
    

    View Slide

40. Mark Zuckerberg
    

    View Slide

41. Context
    

    View Slide

42. SiO
    2
    is a living Natural Language Pipeline of networked
    algorithms trained on the corpus of Quartz to understand
    the linguistic patterns of global business news
    

    View Slide

43. The Pipeline(s)
    Quartz Corpus -> Training Sentences -> spaCy
    Content -> spaCy -> Language Graph
    Language Graph -> Training Data -> Statistical Models / Classifiers
    Language Graph -> Training Sentences -> spaCy
    Unseen Content -> spaCy -> Pre-Processed Text / Vectors -> Statistical Models / Classifiers
    

    View Slide

44. Thank you
    

    View Slide