ADDC 2018 - Cristina Santamarina: Humans vs Bots: Typos, Trolls and other challenges for NLP

Humans vs Bots explores the technical limitations of NLP in a world of high expectations.

A mix of faster computers, a more mature artificial intelligence field and the growing collaboration of technologists and humanists is shaping the new generation of conversational interfaces. How much magic is there really involved and where are we in the practice? How are real chatbots from real brands performing?

In this talk Cristina Santamarina will review the 10 most important natural language challenges she faces when designing conversational interfaces. From slang and typos to rants and trolls she will provide examples and propose dialogue design techniques that can help mitigate them.

More about the talk, authors & slides: https://addconf.com/2018/schedule/humans-vs-bots-typos-trolls-and-other-challenges-for-nlp/
Read about the conference: https://addconf.com

Transcript

1. Humans vs
   Chatbots
   @CRISSANTAMARINA
   

   View Slide

2. View Slide

3. View Slide

4. View Slide

5. View Slide

6. View Slide

7. View Slide

8. View Slide

9. @ tnpgram
   

   View Slide

10. View Slide

11. View Slide

12. View Slide

13. View Slide

14. @ tnpgram
    

    View Slide

15. View Slide

16. View Slide

17. View Slide

18. View Slide

19. View Slide

20. View Slide

21. View Slide

22. View Slide

23. View Slide

24. View Slide

25. FEATURES LANGUAGES
    SLANG MISTEAKS
    LONG INPUT EMPATHY
    CONTEXT
    ABBREVIATIONS TROLLS
    REFERENCES
    10 COMMON CHALLENGES IN NLP CHATBOT PROJECTS
    @ tnpgram
    

    View Slide

26. FEATURES
    

    View Slide

27. View Slide

28. View Slide

29. View Slide

30. View Slide

31. Basic features
    What’s the single most important
    action your users want to complete
    using your bot? Surveys, workflow
    support, online shopping, daily news,
    storytelling… pick your most
    important features and polish them
    until they’re smooth
    1
    Navigation and Support
    Once the basics are ready, make your
    chatbot better. Reorient users to the
    happy path, answer FAQs about
    chatbots and navigation and get your
    users human help when the chatbot
    can’t handle the conversation alone.
    2
    Small talk
    If you have minimised errors now you
    can improve your chatbot with small
    talk, cultural references, answers to
    other questions… Learn from the
    input your chatbot can’t process to
    discover new features.
    3
    

    View Slide

32. View Slide

33. FEATURES
    ● Have a purpose and focus
    ● Build from the bottom up
    ● Start stating your purpose and limitations
    ● Help lost users discover features
    ● Use buttons and rich elements as support
    ● Don’t get trapped in small talk
    What to do about?
    

    View Slide

34. LANGUAGES
    

    View Slide

35. View Slide

36. View Slide

37. View Slide

38. LANGUAGES
    ● Focus on your market
    ● Use your tools right
    ● Use the right tools
    What to do about?
    

    View Slide

39. CONTEXT
    

    View Slide

40. View Slide

41. View Slide

42. View Slide

43. View Slide

44. CONTEXT
    ● Quick, simple
    ● Diagram your processes and subprocesses
    ● Use buttons and other rich elements
    ● Use your tools right
    What to do about?
    

    View Slide

45. LONG INPUT
    

    View Slide

46. View Slide

47. LONG INPUT
    ● Consider limiting input length
    ● Most long inputs are trolls or lost users
    What to do about?
    

    View Slide

48. EMPATHY
    

    View Slide

49. View Slide

50. View Slide

51. EMPATHY
    ● Do you really need pre-processing?
    ● Pay attention to the most important keywords
    ● Emojis are the way we express feelings online
    What to do about?
    

    View Slide

52. SLANG
    

    View Slide

53. View Slide

54. SLANG
    ● Reconsider your need to accept slang input
    ● Analyse the way your users talk
    ● Don’t sound TOO cool
    What
    to do about
    ?
    

    View Slide

55. MISTEAKS
    

    View Slide

56. View Slide

57. View Slide

58. MISTEAKS
    ● Think: how literate are your users?
    ● Use quick replies!
    ● There’s some magic involved in this
    ● Be creative: Think of ways to mess up words
    ● The Real World™ will help you get better
    What to do about?
    

    View Slide

59. REFERENCES
    

    View Slide

60. View Slide

61. View Slide

62. REFERENCES
    ● Focus on the happy path
    ● Analyse users to select important references
    ● Be up to date with news and memes
    ● Use references as a wow / PR effect
    What to do about?
    

    View Slide

63. ABBREVIATIONS
    

    View Slide

64. View Slide

65. ABBREVIATIONS
    ● Add common abbreviations
    ● Think of vertical-specific language
    ● Train on demand
    What to do about?
    

    View Slide

66. TROLLS
    

    View Slide

67. TIME + WORDS
    =
    $$$
    

    View Slide

68. TROLLS
    ● Look for long conversations and long entries
    ● Block recurring trolls
    ● Use trolls to improve your sarcasm
    ● Don’t let them distract you, always redirect!
    What to do about?
    

    View Slide

69. View Slide

70. Learn. By.
    Experience
    

    View Slide

71. FEATURES LANGUAGES
    SLANG MISTEAKS
    LONG INPUT EMPATHY
    CONTEXT
    ABBREVIATIONS TROLLS
    REFERENCES
    10 COMMON CHALLENGES IN NLP CHATBOT PROJECTS
    @ tnpgram
    

    View Slide

72. THANK YOU!
    Questions?
    @CRISSANTAMARINA
    CRISTINA@THENEONPROJECT.ORG
    

    View Slide