The document discusses the evolution of chatbots, highlighting their current applications like in Facebook Messenger and Alexa, alongside their historical development from early interactions with machines. It covers key technological milestones, significant figures in AI history, and the challenges of building effective conversational interfaces while also forecasting future trends in the chatbot landscape. The importance of data-rich environments and interdisciplinary approaches is emphasized for the advancement of chatbots.

1. Chatbots
Present, Past and Future
Paul Houle
Ontology2

2. Chatbots Present

3. Facebook Messenger Amazon Echo (Alexa)
Interact with chatbots using the same UI you use to
interact with people
Conversational voice interface oriented around tasks;
voice is the primary, if not the only interface

4. Facebook Messenger Platform
Message content:
1. Text
2. Video, Audio and Image Media Objects
3. Hyperlinks, Buttons, “Call to Action”
Application Server
Back-End Business
Systems
Real World
Human Supervision

5. Facebook Messenger: Air Travel Application

6. Air Travel Application
• Large Commercial Value (3 tickets, $4000)
• Complex Interaction that takes place over many days
• Takes place in moderately large, but finite world
• Mixed-initiative, changes can occur on either end:
• Passenger: change flight, seats, upgrade, ...
• Airline: cancellation, delay, standby, ...

7. Your IT Systems
Alexa Skills: System architecture

8. Alexa Skills: Examples

9. The trouble with mobile applications
Complex Business
Complex
Applications
Client
Communication Networks
App Stores
A complex business means a complex application: hard to distribute as a mobile app if all screens are sent on the client
a real business application can have 200-2000 screens!

10. Asynchronous/Mixed Initiative User Interface

11. Chatbots past

12. From telegraph to teletype...
20 words/minute ≈ 2 character/sec TELEX
15 character/sec ≈ 110 bits/sec (bps)

13. 1770: The Turk
Chess playing “robot” controlled by a very
small chess player

14. 1942 Alan Turing: War Hero

15. Alan Turing: Computer Scientist
1952 First Computer Chess ProgramFundamentals of Computation

16. 1950: Alan Turing’s “Imitation Game”
Interrogator tries to answer: “who is the male and who
is the female?” Both players are pretending to be the
other

17. Can a computer understand gender?
Can a computer pretend to be gendered? Can it pretend
to be the other gender?
What social skills are necessary?

18. Can a computer pass as a human at all?

20. 1956 1957

21. 1957 Client-Centered Therapy
Psychotherapist
Carl Rogers
Client-centered therapy
mirrored
not mirrored

22. 1960 SABRE Airline Booking System
1958 SAGE Air Defense System

23. Two response strategies are sufficient to elicit empathy: user fills in imagined motives for simulated computer therapist
(is that different from experiencing fiction?)
1966 The Eliza Effect
Active Listening
Changing the Subject
ELIZA: Maybe your plans have something to do with this.
ELIZA: Don't you think computers can help people?
ELIZA: I've told you that apologies are not required.
ELIZA: Can You Elaborate On That?

24. Science Fiction Robots

25. 1970 SHRDLU
> Pick up a big red block.
OK.
> Grasp the pyramid
I DON'T UNDERSTAND WHICH PYRAMID YOU MEAN.
> Find a block which is taller than the one you are holding and put it into the box.
BY "IT", I ASSUME YOU MEAN THE BLOCK WHICH IS TALLER THAN THE ONE I AM HOLDING.
OK.
> What does the box contain?
THE BLUE PYRAMID AND THE BLUE BLOCK.
> What is the pyramid supported by?
THE BOX.
SHRDLU makes conversation about an imaginary
“blocks world” with the computer and can order
the computer to make changes to the block
world within some constraints.

26. id type color picked_up above inside height width ...
1 block green 0 0 0 2.0 2.0 ...
2 block red 1 0 0 3.0 1.5 ...
3 pyramid red 0 1 0 3 0.5 ...
4 block blue 0 0 0 4.0 2.5 ...
“The Blocks World”

27. SHRDLU capabilities
“Have a conversation about a database”
English-like syntax
Remembers context and...
... asks questions to uncover context
ability to reason and plan

28. Teleprinter Video Display Terminal

29. 1975 Microsoft Basic
1) Using BASIC is like having a conversation, and...
2) ... A conversation-like interface is easy to implement in BASIC

30. 1975 Microsoft Basic
... but the conversation context is encoded in the state of the program
Not Flexible
Can’t handle Mixed Initiative

31. 1975 MYCIN
Medical Diagnostic Expert System

32. 1975 MYCIN
Medical Diagnostic Expert System

33. 1975 MYCIN
Medical Diagnostic Expert System
Rules
encode knowledge & procedures
Facts
describe the problem

34. Production Rules / RETE Engine
Ideal for mixed initiative:
• System accepts facts from both the user and the world (react to multiple inputs)
• Firing rules can (i) cause actions and (ii) cause more rules to fire

35. 1976 Colossal Cave
database describing game world
has a lot in common with SHRDLU....

36. 1983 Infocom

37. 1984 Apple Macintosh

38. “AI Winter”
SHDRLU and MYCIN were not scalable!
Algorithms and tools did not scale
Labor Cost to Create Knowledge Base too High

39. 1991-present Adobe Premiere

40. 1994
Netscape
Navigator

41. 2001 Metal Gear Solid 2

42. 2016 Consumer VR

43. REPL (of many kinds) is alive and well
but not as smart as SHRDLU!

44. 1988 Mathematica Notebooks

45. 1985 GSM and SMS
A PUBL1C SERVICE ANN0UNCEMENT!

46. IRC, ICQ, AIM, Skype...

47. 2007 iPhone

48. 2009-2014 WhatsApp

49. • 1-layer neural network can learn a separating line between two categories
• The input space could have thousands or millions of dimensions (ex. an image)
1957 Perceptrons
This 1969 book by Minsky and Papert demolished Perceptrons by demonstrating many
things Perceptrons could not do

50. 1971 full-text search
tfidf: term frequency/inverse document frequency
Karen-Spärck Jones and Gerard Salton @ Cornell
• is computing a dot product in high-dimensional space just like the Perceptron
• is solving a “subjective” problem, isn’t expected to get 100% right answers
• still the dominant algorithm for full-text search in 2016

51. 1992 US Post Office: Handwritten Digits
Backpropagation makes it practical to train shallow neural
networks.
NIST developed training data for this project
that eventually became the famous MNIST
digits

52. predictive analytics
Problem: given four measurements for an Iris flower, determine species
data from 1936 Ronald Fisher and Edgar Anderson
methods based on Vector Spaces:
methods based on rules:
linear discriminant, support vector machine, neural networks,
k-nearest neighbors, etc.
C4.5, random forests, inductive logic programming

53. data driven competitions
1992-present: yearly competitions and conference to improve accuracy of search
engines and similar systems. Supported by US National Institute of Standards.
2010-present data: set of images annotated with noun concepts from Wordnet,
yearly competitions for classification tasks have led to large breakthroughs in
convolutional neural networks & image recognition
2010-present data: venture backed company solves data science problems for
customers by holding public competitions

54. 2006 deep learning
Two kinds of training data:
1. A large number of unlabeled examples
2. A small number of classified examples
Two phases:
1. Deep Belief Network (DBN) learns statistical regularities in
unlabeled data
2. Backpropagation fine tunes the network for a specific task
based on labeled and/or unlabeled data
information bottleneck
forces network to
generalize rather than
memorize

55. present:Explosion of Neural Network
Architectures

56. 2011 IBM Watson wins at Jeopardy
Watson considers several possible answers to a question and
computes a probability score for each one.
Watson weighs the risk of getting a wrong answer against the risk of
an opponent answering first and takes action at the optimal time. It
calculates bets to ensure a win, if possible.

58. Watson Precision/Recall Curve
Throwaway prototype based on commercial off-
the-shelf (COTS) full-text search engine
Point cloud is estimated performance of
human jeopardy players, the goal is to
get into this region
Progressive improvement of:
• knowledge base
• question answering strategies
• result merging

59. Watson achieves hyperprecision because it can choose whether or
not to answer a question.
Watson reasons about uncertainty in order to maximize a utility
function; act in it’s own “self interest”

60. Chatbots future

61. Overwhelmed?
Intelligent systems use complexity to
cope with the complexity of the world
they inhabit.
Outsource subtasks to agents...

62. present:

63. also on

64. a
company
wit.ai: rapidly generalizes examples of specific patterns of text and links these to “intents”

65. business rules: revenge of the expert system!

66. Business Processing Modelling Language
• Production rules engine scale 1000x larger
• ... are ideal for managing processes which happen over an extended time
• ... that are driven by events that happen in “the real world”

67. Complex Event Processing
Greatly improved RETE algorithms do this efficiently!
Rules can put together a story about a set of related
events, by creating new events when the existing
events meet some condition

68. constraint solving & optimization
Route Optimization
SAT/SMT Solver
Travel PlanningBox Packing
There are tools such as Drools OptaPlanner and IBM CPLEX Optimizer that
marry constraint solving and optimization with rules-based systems.
However, many people who work in this space code everything in C++
because they want to try the largest rate of possibilities per second

69. 2XL
“toy of the year”
-- Disney's Family Fun Magazine
Speak and
Spell
1978

70. 1980s Interactive Voice Response:

71. 2001 <VoiceXML>
VoiceXML (from TellMe) supports text to
speech in voice prompts and lets the
script author write a grammar for things
that the telephone caller is supposed to
say.
Performance is dependent on the system
modelling what the user might say: it can
resolve addresses in the US because it
has a list of all the street names!
Can call out to “Web Services” in order to
implement business tasks

72. 2000 Voice Improvement Program
Today: A chance of rain after 4pm. Increasing clouds, with a high near 41.
Southeast wind 6 to 8 mph. Chance of TODAY: A CHANCE OF RAIN AFTER 4PM.
INCREASING CLOUDS, WITH A HIGH NEAR 41. SOUTHEAST WIND 6 TO 8 MPH. CHANCE OF
PRECIPITATION IS 50%. NEW PRECIPITATION AMOUNTS OF LESS THAN ...
• Six voices: male/female and English/Spanish
• Voices vary speed and pitch to create feeling of urgency
• Requires full attention to listen to
ITHACA WXN59 162.5 MHZ

73. Video games engage players with dialog
that supports the story.
Dialog depends heavily on
writing & voice acting and
is not very interactive.
Text-to-speech can’t keep
users engaged

75. SSML

76. All-in
Amazon Alexa is a new platform where you
can’t fall back to the keyboard, mouse or touch-screen.
Voice function has to be good
Others
Bolted onto full-powered phones, computers, and game
consoles, vendors don’t have to face the hard cases
for voice control because fallback to traditional
controllers is imminent.

77. seeing the world that humans live in
specialized cameras and sensors let robots see the world directly in 3-d

78. varieties of depth camera
multiocular structured light
laser scanner

79. Kinect/A Sensor for a Sessile Robot

82. Inform 7 tricks
and ideas
Controlled English facts and rules Pre-existing Ontologies and Theories
Rules Override Other Rules
Parsing number words as numbers Rule Precedence Managed with
“Rulebooks”

83. Conclusion:
Chatbots
• Popular today because of mobile
application limitations
• Possible “third platform” for
applications
• Use a wide range of tactics to
accomplish goals
• Chatbots in 2017 will depend on
data-rich services
• Deeply interdisciplinary, involving
art as much as science