@cataldomusto
cataldo.musto@uniba.it
Towards Queryable User Profiles:
Introducing Conversational Agents in
a Platform for Holistic User Modeling
CATALDO MUSTO*, FEDELUCIO NARDUCCI^, MARCO POLIGNANO*, PASQUALE LOPS*, MARCO DE GEMMIS*, GIOVANNI SEMERARO*
*UNIVERSITÀ DEGLI STUDI DI BARI ‘ALDO MORO’ – ITALY
^POLYTECHNIC UNIVERSITY OF BARI - ITALY
ExUM 2020 – International
Workshop on Explainable User
Models and Personalised Systems
jointly held with UMAP 2020
Online from Genoa (Italy) – July 18, 2020
2
User Modeling (in the ‘egosystem’ era)
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
Holistic User Model*
Affects
Demographics
Interests
Behaviors
Social
Relations
Knowledge
and Skills
Physical
States
Psychological
Aspects
6
(*) Musto, C., Polignano, M.,
Semeraro, G. et al. MYRROR: a
platform for holistic user
modeling. User Modeling and
User-Adapted
Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
7
Myrror: a platform for holistic user modeling*
(*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic
user modeling. User Modeling and User-Adapted Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
8
Myrror: a platform for holistic user modeling*
(*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic
user modeling. User Modeling and User-Adapted Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
10
Myrror: a platform for holistic user modeling*
(*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic
user modeling. User Modeling and User-Adapted Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
11
Myrror: a platform for holistic user modeling*
(*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic
user modeling. User Modeling and User-Adapted Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
13
Myrror: a platform for holistic user modeling*
(*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic
user modeling. User Modeling and User-Adapted Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
14
Myrror: a platform for holistic user modeling*
(*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic
user modeling. User Modeling and User-Adapted Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
15
Myrror: a platform for holistic user modeling*
(*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic
user modeling. User Modeling and User-Adapted Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
Myrror: a platform for holistic user modeling*
• Current Issues
1. Overall transparency may be improved
◦ Users should be aware of the information encoded in their
profiles
◦Interaction with a web-based dashboard can be difficult,
especially for beginners
◦ «How can I check my current interests?»
2. It is difficult to get personal «insights» from the current
dashboard
◦ «Did I sleep enough?»
16
(*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic
user modeling. User Modeling and User-Adapted Interaction (2020).
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 17
Goals
To improve the way people access to the
information encoded in their profiles
18
Intuition: Conversational Agents
What is a conversational
agent?
• A tool that understands
natural-language
requests and provides
natural-language answers
to general-purposes
services
• Popularized by Alexa, Siri,
Google Now, etc.
Alexa, play Africa
by Toto.
Here’s Africa by Toto
from Amazon Music.
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
19
Intuition: Conversational Agents
Intuition
• To exploit Conversational
Agents to allow the user
to query her holistic user
model in natural
language
• Easier to access to the
information encoded
in the profile
• Easier to get insights
Alexa, play Africa
by Toto.
Here’s Africa by Toto
from Amazon Music.
Did I sleep enough?
No, your sleep is not
enough. Currently, you
sleep 6h per day.
A first step towards Queryable User Profiles
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
21
Contribution: MyrrorBot
MyrrorBot is a conversational
agent built on top of Myrror
Based on HUMs
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
MyrrorBot is a conversational
agent built on top of Myrror
Based on HUMs
Answers to natural-language
requests concerning the
information encoded in the
HUM
Provides personalized services
Alexa-like (Music, News,
Recipes, etc.)
22
Contribution: MyrrorBot
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
23
Contribution: MyrrorBot
MyrrorBot is a conversational
agent built on top of Myrror
Based on HUMs
Answers to natural-language
requests concerning the
information encoded in the
HUM
Provides personalized services
Alexa-like (Music, News,
Recipes, etc.)
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
24
MyrrorBot: Intent Recognizer
MyrrorBot is a conversational
agent built on top of Myrror
Based on HUMs
Answers to natural-language
requests concerning the
information encoded in the
HUM
Provides personalized services
Alexa-like (Music, News,
Recipes, etc.)
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
25
MyrrorBot: Intent Recognizer
Our natural language
understanding module is based
on Google DialogFlow.
Input:
Natural Language
Request
Output:
Intent (i.e., a
representation of the
informative need)
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
Alexa, play
Africa by Toto.
Here’s Africa by Toto
from Amazon Music.
Intent=music
26
MyrrorBot: Intent Recognizer
Our natural language
understanding module is based
on Google DialogFlow.
Input:
Natural Language
Request
Output:
Intent (i.e., a
representation of the
informative need)
What kind of
questions can
we handle?
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
29
Intents in MyrrorBot
Demographics
Interests
Affects
Physical States
Social Relations
Behaviors
Knowledge and Skills
Psychological Traits
Intents are directly mapped to the facets of HUMs
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
30
Intents in MyrrorBot
Demographics
Interests
Affects
Physical States
Social Relations
Behaviors
Knowledge and Skills
Psychological Traits
Each intent is further split into two sub-intents
Informative
(«How much did I
sleep today?»)
Insights
(«Did I sleep
enough today?»)
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
35
Training the Intent Recognizer
Intent Recognizer is
based on Machine
Learning
For each intent, a set of
training sentences
covering all the
possible utterances
should be provided
We feed the algorithm
with 20-30 sentences
for each intent
Behaviors - Insights Behaviors - Insights
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
36
Intent Recognizer @Work
Affects - Informative
Psychological Traits - Informative
Behaviors - Insights
Physical States - Insights
Demographics - Informative
What is my mood right now, according
to your data?
Tell me about my personality traits
Did I burn enough calories?
Did I sleep enough today?
Where do I live?
Intent
Recognizer
(NLU)
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
38
MyrrorBot: Workflow
MyrrorBot is a conversational
agent built on top of Myrror
Based on HUMs
Answers to natural-language
requests concerning the
information encoded in the
HUM
Provides personalized services
Alexa-like (Music, News,
Recipes, etc.)
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
40
MyrrorBot: Generator
The Generator module produces a natural
language answer that fulfill user’s
information need.
In case of Informative intent, no computation
is carried out. The profile is queried and a
template is filled in with the values
returned by the HUM.
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
41
MyrrorBot: Generator
The Generator module
produces a natural language
answer that fulfill user’s
information need.
Differently, intents that aim to
get some insights require some
computation (i.e., compare
current values with some
threshold value) before the
answer is generated.
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 42
MyrrorBot: Recap
Acquires a HUM
Recognizes users’ requests
Generate the answer
Experimental Evaluation
Research Question 1
What is the opinion of the users, in terms of usability, ease of use and learning curve?
Research Question 2
Does a conversational interface reduces the time needed to access to the personal information
encoded in a user profile?
Experimental Design
User Study with a Web Application
76 subjects (previous experience with chatbots= 35.5%)
Metrics RQ1: Questions based on SUS (System Usability Survey) Questionnaire
5-points Likert scale
Metrics RQ2: Seconds required to complete a task (web-based vs. conversational interface)
Randomized Order of interaction, Wilcoxon Test
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 43
Results – Research Question 1
Metric Question %Compl. Agree %Agree %Others
Learning Curve «I became familiar with
the system very quickly»
51.3% 38.2% 10.5%
Interaction «It was easy to formulate
my questions»
40.8% 48.7% 11.8%
Accuracy «The system provided me
with the right answers»
55.3% 35.5% 9.2%
Daily Use «I would use the system
for my daily tasks»
28.9% 40.8% 29.2%
Switching «I would prefer the system
to a web platform»
40.8% 40.8% 18.4%
Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles:
Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 44
Results – Research Question 2
Facet Task MyrrorBot
(sec.)
Myrror
(sec.)
Gap%
Demographics Check your birthday 7.72 10.31 -25.12%
(p<0.001)
Interests Check your interests 9.62 22.11 -56.49%
(p<0.001)
Affects Check your latest mood 10.34 20.02 -48.35%
(p<0.001)
Psychological
Aspects
Check your personality
traits
9.98 12.33 -19.10%
(p<0.05)
45
Significant reduction of the task completion time
Results – Research Question 2 (cont.)
Facet Task MyrrorBot
(sec.)
Myrror
(sec.)
Gap%
Behaviors
Check your daily activities 13.25 24.99 -47.20%
(p<0.001)
Check if you burnt
enough calories
14.18 23.17 -19.05%
(p<0.05)
Physical States
Check if you slept enough 13.44 19.20 -30.47%
(p<0.001)
Check your heart rate 10.20 16.49 -38.14%
(p<0.001)
Significant reduction of the task completion time
46
Research Question 2 – Recap
Avg. Reduction40.3%
11,09
18,58
0,00 2,00 4,00 6,00 8,00 10,00 12,00 14,00 16,00 18,00 20,00
MyrrorBot
Myrror
Average Task Completion Time (sec.)
47
Recap
A first step towards Queryable User Profiles
Starting Point: Myrror
◦ Holistic User Modeling platform
◦ Collects data from several heterogeneous sources
◦ Processes the data and aggregates them to eight pre-defined facets
Contribution: MyrrorBot
◦ A Conversational agent built on top of MyrrorBot
◦ Acquires a HUM and recognizes users’ requests (based on an intent recognizer)
◦ Generate natural language answers fulfilling the requests
◦ Informative and Insight-based intents
◦ Good impact on the users. Significant reduction of task completion time.
48
Demo Available!
cataldo.musto@uniba.it
@cataldomusto
Want to try MyrrorBot?
Contact us!
49
Future Work
Evaluation of personalized services
Evaluation of new interaction
methods
◦ Integration with social robots (e.g.,
Pepper)
◦ Integration with voice-based interfaces
(e.g., Amazon Alexa-like)
Training of more (and more
sophisticated) intents
◦ Goal: to get more useful insights from the
platform (e.g., is my lifestyle correct?)
50
Thank you!
cataldo.musto@uniba.it
@cataldomusto
Contacts
51

Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling

  • 1.
    @cataldomusto cataldo.musto@uniba.it Towards Queryable UserProfiles: Introducing Conversational Agents in a Platform for Holistic User Modeling CATALDO MUSTO*, FEDELUCIO NARDUCCI^, MARCO POLIGNANO*, PASQUALE LOPS*, MARCO DE GEMMIS*, GIOVANNI SEMERARO* *UNIVERSITÀ DEGLI STUDI DI BARI ‘ALDO MORO’ – ITALY ^POLYTECHNIC UNIVERSITY OF BARI - ITALY ExUM 2020 – International Workshop on Explainable User Models and Personalised Systems jointly held with UMAP 2020 Online from Genoa (Italy) – July 18, 2020
  • 2.
    2 User Modeling (inthe ‘egosystem’ era) Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 3.
    Holistic User Model* Affects Demographics Interests Behaviors Social Relations Knowledge andSkills Physical States Psychological Aspects 6 (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 4.
    7 Myrror: a platformfor holistic user modeling* (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 5.
    8 Myrror: a platformfor holistic user modeling* (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 6.
    10 Myrror: a platformfor holistic user modeling* (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 7.
    11 Myrror: a platformfor holistic user modeling* (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 8.
    13 Myrror: a platformfor holistic user modeling* (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 9.
    14 Myrror: a platformfor holistic user modeling* (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 10.
    15 Myrror: a platformfor holistic user modeling* (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 11.
    Myrror: a platformfor holistic user modeling* • Current Issues 1. Overall transparency may be improved ◦ Users should be aware of the information encoded in their profiles ◦Interaction with a web-based dashboard can be difficult, especially for beginners ◦ «How can I check my current interests?» 2. It is difficult to get personal «insights» from the current dashboard ◦ «Did I sleep enough?» 16 (*) Musto, C., Polignano, M., Semeraro, G. et al. MYRROR: a platform for holistic user modeling. User Modeling and User-Adapted Interaction (2020). Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 12.
    Cataldo Musto, FedelucioNarducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 17 Goals To improve the way people access to the information encoded in their profiles
  • 13.
    18 Intuition: Conversational Agents Whatis a conversational agent? • A tool that understands natural-language requests and provides natural-language answers to general-purposes services • Popularized by Alexa, Siri, Google Now, etc. Alexa, play Africa by Toto. Here’s Africa by Toto from Amazon Music. Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 14.
    19 Intuition: Conversational Agents Intuition •To exploit Conversational Agents to allow the user to query her holistic user model in natural language • Easier to access to the information encoded in the profile • Easier to get insights Alexa, play Africa by Toto. Here’s Africa by Toto from Amazon Music. Did I sleep enough? No, your sleep is not enough. Currently, you sleep 6h per day. A first step towards Queryable User Profiles Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 15.
    21 Contribution: MyrrorBot MyrrorBot isa conversational agent built on top of Myrror Based on HUMs Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 16.
    MyrrorBot is aconversational agent built on top of Myrror Based on HUMs Answers to natural-language requests concerning the information encoded in the HUM Provides personalized services Alexa-like (Music, News, Recipes, etc.) 22 Contribution: MyrrorBot Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 17.
    23 Contribution: MyrrorBot MyrrorBot isa conversational agent built on top of Myrror Based on HUMs Answers to natural-language requests concerning the information encoded in the HUM Provides personalized services Alexa-like (Music, News, Recipes, etc.) Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 18.
    24 MyrrorBot: Intent Recognizer MyrrorBotis a conversational agent built on top of Myrror Based on HUMs Answers to natural-language requests concerning the information encoded in the HUM Provides personalized services Alexa-like (Music, News, Recipes, etc.) Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 19.
    25 MyrrorBot: Intent Recognizer Ournatural language understanding module is based on Google DialogFlow. Input: Natural Language Request Output: Intent (i.e., a representation of the informative need) Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 Alexa, play Africa by Toto. Here’s Africa by Toto from Amazon Music. Intent=music
  • 20.
    26 MyrrorBot: Intent Recognizer Ournatural language understanding module is based on Google DialogFlow. Input: Natural Language Request Output: Intent (i.e., a representation of the informative need) What kind of questions can we handle? Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 21.
    29 Intents in MyrrorBot Demographics Interests Affects PhysicalStates Social Relations Behaviors Knowledge and Skills Psychological Traits Intents are directly mapped to the facets of HUMs Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 22.
    30 Intents in MyrrorBot Demographics Interests Affects PhysicalStates Social Relations Behaviors Knowledge and Skills Psychological Traits Each intent is further split into two sub-intents Informative («How much did I sleep today?») Insights («Did I sleep enough today?») Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 23.
    35 Training the IntentRecognizer Intent Recognizer is based on Machine Learning For each intent, a set of training sentences covering all the possible utterances should be provided We feed the algorithm with 20-30 sentences for each intent Behaviors - Insights Behaviors - Insights Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 24.
    36 Intent Recognizer @Work Affects- Informative Psychological Traits - Informative Behaviors - Insights Physical States - Insights Demographics - Informative What is my mood right now, according to your data? Tell me about my personality traits Did I burn enough calories? Did I sleep enough today? Where do I live? Intent Recognizer (NLU) Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 25.
    38 MyrrorBot: Workflow MyrrorBot isa conversational agent built on top of Myrror Based on HUMs Answers to natural-language requests concerning the information encoded in the HUM Provides personalized services Alexa-like (Music, News, Recipes, etc.) Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 26.
    40 MyrrorBot: Generator The Generatormodule produces a natural language answer that fulfill user’s information need. In case of Informative intent, no computation is carried out. The profile is queried and a template is filled in with the values returned by the HUM. Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 27.
    41 MyrrorBot: Generator The Generatormodule produces a natural language answer that fulfill user’s information need. Differently, intents that aim to get some insights require some computation (i.e., compare current values with some threshold value) before the answer is generated. Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020
  • 28.
    Cataldo Musto, FedelucioNarducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 42 MyrrorBot: Recap Acquires a HUM Recognizes users’ requests Generate the answer
  • 29.
    Experimental Evaluation Research Question1 What is the opinion of the users, in terms of usability, ease of use and learning curve? Research Question 2 Does a conversational interface reduces the time needed to access to the personal information encoded in a user profile? Experimental Design User Study with a Web Application 76 subjects (previous experience with chatbots= 35.5%) Metrics RQ1: Questions based on SUS (System Usability Survey) Questionnaire 5-points Likert scale Metrics RQ2: Seconds required to complete a task (web-based vs. conversational interface) Randomized Order of interaction, Wilcoxon Test Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 43
  • 30.
    Results – ResearchQuestion 1 Metric Question %Compl. Agree %Agree %Others Learning Curve «I became familiar with the system very quickly» 51.3% 38.2% 10.5% Interaction «It was easy to formulate my questions» 40.8% 48.7% 11.8% Accuracy «The system provided me with the right answers» 55.3% 35.5% 9.2% Daily Use «I would use the system for my daily tasks» 28.9% 40.8% 29.2% Switching «I would prefer the system to a web platform» 40.8% 40.8% 18.4% Cataldo Musto, Fedelucio Narducci, Marco Polignano, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro. Towards Queryable User Profiles: Introducing Conversational Agents in a Platform for Holistic User Modeling. ExUM Workshop@UMAP 2020 – Online – July 18, 2020 44
  • 31.
    Results – ResearchQuestion 2 Facet Task MyrrorBot (sec.) Myrror (sec.) Gap% Demographics Check your birthday 7.72 10.31 -25.12% (p<0.001) Interests Check your interests 9.62 22.11 -56.49% (p<0.001) Affects Check your latest mood 10.34 20.02 -48.35% (p<0.001) Psychological Aspects Check your personality traits 9.98 12.33 -19.10% (p<0.05) 45 Significant reduction of the task completion time
  • 32.
    Results – ResearchQuestion 2 (cont.) Facet Task MyrrorBot (sec.) Myrror (sec.) Gap% Behaviors Check your daily activities 13.25 24.99 -47.20% (p<0.001) Check if you burnt enough calories 14.18 23.17 -19.05% (p<0.05) Physical States Check if you slept enough 13.44 19.20 -30.47% (p<0.001) Check your heart rate 10.20 16.49 -38.14% (p<0.001) Significant reduction of the task completion time 46
  • 33.
    Research Question 2– Recap Avg. Reduction40.3% 11,09 18,58 0,00 2,00 4,00 6,00 8,00 10,00 12,00 14,00 16,00 18,00 20,00 MyrrorBot Myrror Average Task Completion Time (sec.) 47
  • 34.
    Recap A first steptowards Queryable User Profiles Starting Point: Myrror ◦ Holistic User Modeling platform ◦ Collects data from several heterogeneous sources ◦ Processes the data and aggregates them to eight pre-defined facets Contribution: MyrrorBot ◦ A Conversational agent built on top of MyrrorBot ◦ Acquires a HUM and recognizes users’ requests (based on an intent recognizer) ◦ Generate natural language answers fulfilling the requests ◦ Informative and Insight-based intents ◦ Good impact on the users. Significant reduction of task completion time. 48
  • 35.
  • 36.
    Future Work Evaluation ofpersonalized services Evaluation of new interaction methods ◦ Integration with social robots (e.g., Pepper) ◦ Integration with voice-based interfaces (e.g., Amazon Alexa-like) Training of more (and more sophisticated) intents ◦ Goal: to get more useful insights from the platform (e.g., is my lifestyle correct?) 50
  • 37.