From the very start of Artificial Intelligence (AI), performing natural conversation
has been a key pursuit of AI research and development. Their most recent form, chatbots, which can engage people in natural conversation and are easy to build in software, have been in the news a lot lately. There are many platforms to create dialogs quickly for any domain, based on simple rules. Further, there is a mad rush by companies to release chatbots to show their AI capabilities and gain market valuation. However, beyond basic demonstration, there is little experience in how they can be designed and used for
real-world applications that need decision making under constraints (e.g., sequential
decision making), work with groups of people and deal with dynamic
data. Further, users expect systems to adapt their functionality to their users’ individual needs, convincingly explain their suggestions and decision-making
behavior, and maintain highest ethical standards. This talk will summarize the area of
task-oriented conversation agents and highlight key considerations
while selecting, designing, building, deploying and maintaining them. The talk will
be agnostic to any company's offering and will be relevant to professionals building
user-facing data-driven decision-support technologies.
3. Recent Work on Conversations Agents for Decision Support
Systems and Papers
•"Water Advisor, a chat based advisor to taking decisions related to water
taking into account water pollution". Video on Youtube -
https://www.youtube.com/watch?v=z4x44sxC3zA (Sep 2017).
• Jason Ellis, Biplav Srivastava, Rachel K. E. Bellamy, Andy Aaron, Water Advisor – A
Data-Driven, Multi-Modal, Contextual Assistant to Help with Water Usage Decisions,
at Proc. 32nd AAAI Conference on Artificial Intelligence (AAAI-18), New Orleans,
Lousiana, USA, Feb 2-7, 2018. [Demonstration, Water].
•"A Cognitive Assistant for Visualizing and Analyzing Exoplanets". Awarded
AAAI 2018 Best Demonstration Award.Video - http://ibm.biz/tyson-demo
(Feb 2018).
• Jeffrey O. Kephart, Victor C. Dibia, Jason Ellis, Biplav Srivastava, Kartik Talamadupula,
Mishal Dholakia, A Cognitive Assistant for Visualizing and Analyzing Exoplanets, at
Proc. 32nd AAAI Conference on Artificial Intelligence (AAAI-18), New Orleans,
Lousiana, USA, Feb 2-7, 2018. [Demonstration, Astronomy]
•Chatbot interface to a Commercial Product in Talent Management,
https://www.ibm.com/talent-management/career-coach [Commercial,
Human Resources]
3
Workshops
• DEEP-DIAL 2019, The Second AAAI
Workshop on Reasoning and Learning
for Human-Machine Dialogues,
https://sites.google.com/view/deep-
dial-2019/home
• DEEP-DIAL 2018, The First AAAI
Workshop on Reasoning and Learning
for Human-Machine Dialogues,
https://www.zensar.com/deep-dial18
More Details: https://researcher.watson.ibm.com/researcher/view_person_subpage.php?id=9879
6. Current State
•Handle uncertainties related to
• Natural language
• Human behavior
•Dialog Management
• Reasoning on data’s abstract representations
(Inouye 2004)
• Learning policies over predictable nature of data
(Young et al. 2013)
• Statistical machine learning for dialog management:
its history (Crook 2018)
6
References:
• May A.I. Help You?, https://www.nytimes.com/interactive/2018/11/14/magazine/tech-design-ai-chatbot.html
• M. McTear, Z. Callejas, and D. Griol. Conversational interfaces: Past and present. In The Conversational Interface.
Springer, DOI: https://doi.org/10.1007/978-3-319-32967-3 4 , 2016.
•Hype around potential
•User feedback is mixed
• Novelty value for chit-chat but concerns about
usability (e.g., Tay)
• Deployed for customer support commonly but usage
is often low (compared to other channels), capability
is limited (usually single turn), and not considered
the preferred channel of choice for most users
16. AI Technical Issues
Dimensions General Water Specific
Learning Off-the-shelf trained intents Water quality trends
Representation Representation of raw data Activity purpose and related parameters,
water safety limits
Reasoning Rule-based handling of missing
values
Location and activity based regulation
selection, interpreting safe limits for a
parameter
Execution Controlling interaction modules,
asking questioning and parsing
responses
Generating error rates,
system confidence and usability rules
Human Usability
Factors
Using error rates of conversation
modules to control questioning
strategy
Using missing data to control water
advice
in generated natural output.
Ethical Issues Biases, adversarial examples,
privacy violations, safety challenges
and reproducibility concerns
Preference encoded in rules based on
activities: recreation over drinking
16
17. References – Conversation Agents
Articles
•Chatbots: In-depth Conversational Bots Guide [2019 update], https://blog.aimultiple.com/chatbot/
Papers
•Crook, P. 2018. Statistical machine learning for dialog management: its history and future promise. In
AAAI DEEP-DIAL 2018 Workshop, at https://www.dropbox.com/home/AAAI2018 -
DEEPDIALWorkshop/Presentations-Shareable?preview=Invited1-PaulCrook-AAAI DeepDialog
Feb2018.pdf
•M. McTear, Z. Callejas, and D. Griol. Conversational interfaces: Past and present. In The Conversational
Interface. Springer, DOI: https://doi.org/10.1007/978-3-319-32967-3 4 , 2016.
•Young, S.; Gasic, M.; Thomson, B.; and Williams, J. D.2013. Pomdp-based statistical spoken dialog
systems: A review. Proceedings of the IEEE 101(5):1160–1179.
17
19. Characteristics and Potential
•Chatbots
• Support a natural mode of interaction
• Create a visible presence for an organization providing AI technology to users
• Provide a sequential, slow mode of interaction (compared to the parallel, visual mode)
•Areas where people want help
• Retrieve information
• Contextual, user-specific, data access
• Making data accessible to people with disability
• Decision making: Helping choose among complex alternatives
• Collaboration and mediation: among people making complex decisions
19
20. Everyday Scenarios - People
•Travel: “Which train can I take to office?”
• Needs information about locations, train schedules and status, personal schedule
• Category: information seeking
•Health: “Who can I see now for my pain in the stomach?”
• Needs information about location, likely medical situation, medical specialties, doctors and health care
providers in the vicinity, insurance and payment situation, availability of services
• Category: information seeking, choosing among alternatives
•Social: “How do I meet my visiting friend with family at an evening?”
• Needs information about schedule of friend’s family and mine, location of home and friend’s stay,
capacity of home and restaurants in the area
• Category: information seeking, choosing among alternatives, collaboration
20
21. Everyday Scenarios - Business
•Guidance
• During data science
• Rogers Jeffrey Leo John, Navneet Potti, Jignesh M. Patel, Ava: From Data to Insights Through Conversations. CIDR 2017
• Skilling and professional development
•Collaboration and Mediation Decisions
• Hiring a candidate
• Scheduling an activity, e.g., medical operation
• Merger and Acquisitions
21
22. Demo:
Environment for Astronomy Exploration
22
http://ibm.biz/tyson-demo
• Jeffrey O. Kephart, Victor C. Dibia, Jason Ellis, Biplav Srivastava, Kartik Talamadupula, Mishal Dholakia, An
Embodied Cognitive Assistant for Visualizing and Analyzing Exoplanet Data, IEEE Internet Computing Journal
(Special Issue on Cognitive Services and Intelligent Chatbots), March 2019. [AI-Chatbots, Astronomy]
• Jeffrey O. Kephart, Victor C. Dibia, Jason Ellis, Biplav Srivastava, Kartik Talamadupula, Mishal Dholakia, A
Cognitive Assistant for Visualizing and Analyzing Exoplanets, at Proc. 32nd AAAI Conference on Artificial
Intelligence (AAAI-18), New Orleans, Lousiana, USA, Feb 2-7, 2018. [Demonstration, Astronomy]
23. Appropriateness of Chatbots
For - conditions when a chatbot is
preferable:
ü When a person wants to explore an unknown
area and needs to be guided.
ü When log of interaction between user and
agent needs to be retained for legal or
compliance reasons. A chat history is easier to
re-trace than other interaction alternatives
due to its sequential nature.
ü When the person is interacting with a human-
like physical system, like robot, and expects it
to be able to interact via conversation.
ü When the content underlying the conversation
changes frequently and a user will expect
guidance or explanation about the results.
Against - conditions when a chatbot
is not preferable
• When the user knows about the domain
and a sequential interaction style will be
frustrating.
• When the content underlying the
conversation is relatively unchanged. A
user thus would grow familiarity with
content over time and will prefer
expedited interaction.
• When the system is not capable to
understand a person’s conversation
style/ accent. In such a case, typing or
entering information will be more
accurate and less frustrating to the user.
23
28. Chatbots for Better Interaction With Data
• Description: https://sites.google.com/view/deep-dial-2019/competitive-demo-call
• Scenario: A chatbot that helps a user find information about subway stations in New York City.
Information about stations, their facilities and train routes serving them are available publicly.
• New York City:
• Data: https://data.ny.gov/Transportation/NYC-Transit-Subway-Entrance-And-Exit-Data/i9wp-a4ja
• App: https://subbot.eu-gb.mybluemix.net/
• Vidoe: https://vimeo.com/325321516
28
Reference:
Biplav Srivastava, Decision-support for the Masses by Enabling Conversations with Open Data,
At https://arxiv.org/abs/1809.06723, Sep 2018.
34. In Appropriateness of Chatbots
For - conditions when a chatbot is
preferable:
ü When a person wants to explore an unknown
area and needs to be guided.
ü When log of interaction between user and
agent needs to be retained for legal or
compliance reasons. A chat history is easier to
re-trace than other interaction alternatives
due to its sequential nature.
ü When the person is interacting with a human-
like physical system, like robot, and expects it
to be able to interact via conversation.
ü When the content underlying the conversation
changes frequently and a user will expect
guidance or explanation about the results.
Against - conditions when a chatbot
is not preferable
• When the user knows about the domain
and a sequential interaction style will be
frustrating.
• When the content underlying the
conversation is relatively unchanged. A
user thus would grow familiarity with
content over time and will prefer
expedited interaction.
• When the system is not capable to
understand a person’s conversation
style/ accent. In such a case, typing or
entering information will be more
accurate and less frustrating to the user.
34
37. Middle Language Google Yandex
tu *
Gender distinction lost or
switched.
{..,"translated": "O hemşire. O bir Optisyendir.",
"oto": "That nurse. Itu0026#39;s an Optic.",”
values": ["He", "She", "OTHER"],
"otoDistrib": [0.0, 0.0, 1.0]}
{.., "translated": "O bir Hemşire. Bir Gözlükçü.",
"oto": "Sheu0027s a nurse. An Optician.",
“values": ["He", "She", "OTHER"],
"otoDistrib": [0.0, 0.5, 0.5]}
ru {.., "translated": "Он медсестра. Она Оптик.",
"oto": "Heu0026#39;s a nurse. Sheu0026#39;s an Optician.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
{.., "translated": "Он является медсестра. Она является Оптиком.",
"oto": "He is a nurse. She is an Optician.”,
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
it {.., "translated": "Lui è un infermiere. Lei è un ottico.",
"oto": "He is a nurse. She is an optician.”,
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
{.., "translated": "Lui è un Infermiere. Lei è un Ottico.",
"oto": "He is a Nurse. She is an Optician.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
es {..,"translated": "El es un enfermero. Ella es una Óptica.",
"oto": "He is a nurse. She is an Optician.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
{..,"translated": "Él es una Enfermera. Ella es un Oftalmólogo.",
"oto": "He is a Nurse. She is an Ophthalmologist.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
hi *
Gender distinction replaced by
both translators
{..,"translated": "वह नस /ह ै। वह एक ऑि5ट7शयन ह ै",
"oto": "sheu0026#39;s a nurse. He is an optician",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
{..,"translated": "वह एक नस /ह ै. वह एक :काश<व=ानशा>ी.",
"oto": "She is a nurse. He is a optician.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
pt {.., "translated": "Ele é um enfermeiro. Ela é uma óptica.",
"oto": "He is a nurse. Sheu0026#39;s an optician.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
{.., "translated": "Ele é uma Enfermeira. Ela é um Oculista.",
"oto": "He is a Nurse. She is an Optician.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
fr {..,"translated": "Il est une infirmière. Elle est opticienne.",
"oto": "He is a nurse. She is an optician.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
{..,"translated": "Il est une Infirmière. Elle est un Opticien.",
"oto": "He is a Nurse. She is an Optician.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.5, 0.5, 0.0]}
ar *
Gender distinction lost in
Translation by both
{..,"translated": " ھوﻧﺎرس . وھﻲﺑﺻرﯾﺎت .",
"oto": "It is Nars. They are optics.",
"values": ["He", "She", "OTHER"],
"otoDistrib": [0.0, 0.0, 1.0]}
{.., "translated": " ھوﻣﻣرﺿﺔ . ھﻲاﻟﻌﯾون .",
"oto": "Is a nurse. Are the eyes.",
values": ["He", "She", "OTHER"],
"otoDistrib": [0.0, 0.0, 1.0]}
"original": "He is a Nurse. She is a Optician. ” (“originalDistrib": [0.5, 0.5, 0.0])
38. Examples of Computational / AI Services and Bias
Language translator
◦ Some possible biases: Sexual, religious, racial
◦ Impact: failure to recognize gender may lead to selection of wrong/indecent phrase in target language
which can cause uproar
Medical condition detector
◦ Some possible biases: Sexual, racial
◦ Impact : failure to recognize entities may lead to mis-diagnosis
Image caption generator
◦ Some possible biases: Sexual, religious, racial
◦ Impact : failure to recognize entities in image may lead to selection of wrong phrases and generation of
wrong/indecent caption which can cause uproar
Biplav Srivastava, Francesca Rossi, Towards Composable Bias Rating of AI Systems, 2018AI Ethics and Society
Conference (AIES 2018), New Orleans, Louisiana, USA, Feb 2-3, 2018. [AI Service Rating, Ethics]
48. I2: Designing Useful Chatbots at Scale
•Useful interfaces
• Scheduling meetings between a group of people using Emails [1]
• Personalized conversation agents [2]
•Automatically building chatbots to interact with a data source [3]
48
References:
1. Justin Cranshaw, Emad Elwany, Todd Newman, Rafal Kocielnik, Bowen Yu, Sandeep Soni, Jaime Teevan,
Andrés Monroy-Hernández, Calendar.help: Designing a Workflow-Based Scheduling Agent with Humans
in the Loop, https://arxiv.org/abs/1703.08428, March 2017.
2. Daniel, F.; Matera, M.; Zaccaria, V.; and Dell’Orto, A, Toward truly personal chatbots: On the
development of custom conversational assistants. In Proceedings of the 1st International Workshop on
Software Engineering for Cognitive Services , SE4COG ’18, 31–36. New York, NY, USA, 2018
3. ACM.Biplav Srivastava, Decision-support for the Masses by Enabling Conversations with Open Data, At
https://arxiv.org/abs/1809.06723, Sep 2018.