A domain-independent framework for building conversational recommender systems.
Slides presented @ Google Workshop on Conversational Search and Recommendation, London, 28-29 August 2019
1. ConveRSE:
a domain-independent
Framework for building
Conversational Recommender Systems
Fedelucio Narducci
University of Bari Aldo Moro - Italy
Google Conversational Search & Recommendation Workshop - August 28-29, 2019 - London
2. BACKGROUND
➤ Conversational
Recommender Systems
(CoRSs) belong to the class of
dialog agents and interact
with the user during the
recommendation process
➤ CoRSs guide the users through an
interactive dialog
➤ the preference acquisition
is an incremental process
that does not have to be necessarily
finalized in a single step
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
3. PROBLEMS
➤ Developing dialog
agents is becoming more
and more popular for several
domains and applications
➤ The implementation is a
complex task since it requires
knowledge about NLP,
HCI, machine
learning
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
4. CONVERSE
➤ a domain-
independent
framework for building
conversational
recommender
systems
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
5. CAPABILITIES
➤ Acquire preferences
➤ Explore user profile
➤ Exploit feedback
➤ Recommendation explanation
➤ Offer different interaction
modes
➤ natural language
➤ buttons
➤ hybrid
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
7. ARCHITECTURE: DIALOG MANAGER
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ core component of the
framework
➤ supervises the whole
recommendation process
➤ keeps track of the dialog
state
➤receives/sends
messages from/to the user
➤ is completely independent
from the client (JSON
message)
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
8. INTENT RECOGNIZER
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ identifies the intent of the
user expressed by a natural
language sentence
➤ is based on DialogFlow
➤ recognizes four user intents
➤ preference
➤ recommendation
➤ show profile
➤ help
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
9. ARCHITECTURE: INTENT RECOGNIZER
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ each intent can be composed
of set of sub-intents
➤ show profile
➤ delete preference
➤ update preference
➤ reset profile
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
10. ARCHITECTURE: SENTIMENT ANALYZER
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ Sentiment Tagger of
Stanford CoreNLP
➤ returns the sentiment
tags identified in a sentence
➤ links the sentiment tag
to an entity in the sentence
➤ I like The Matrix, but I hate
Keanu Reeves
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
11. ARCHITECTURE: SENTIMENT ANALYZER
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ Sentiment Tagger of
Stanford CoreNLP
➤ returns the sentiment
tags identified in a sentence
➤ links the sentiment tag
to an entity in the sentence
➤ I like The Matrix, but I hate
Keanu Reeves
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
12. ARCHITECTURE: ENTITY RECOGNIZER
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ Entity Recognizer finds
entities in the user sentence
➤ links them to the
Knowledge Base Wikidata
➤ does not require
annotated data for
training
➤recognizes alias
➤ Steven Allan Spielberg, Spielberg,
Steven Spielberg
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
13. ARCHITECTURE: ENTITY RECOGNIZER - PROBLEMS
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ different surface forms can
refer to the same entity
➤ Steven Spielberg, Spielberg ->
Steven_Spielberg:director
➤ the same surface form
can refer to more than one
entities
➤ Spielberg ->
Steven_Spielberg:director
Sasha_Spielberg:actor
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
14. ARCHITECTURE: ENTITY RECOGNIZER@WORK
➤Step 1
➤ I like Spielberg and Jurassic Park ~ I like Jurassic Park and its director
Spielberg
Steven_Spielberg:director
Sasha_Spielberg:actor
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
15. ARCHITECTURE: ENTITY RECOGNIZER@WORK
➤Step 1
➤ I like Spielberg and Jurassic Park
Spielberg
Steven_Spielberg:director
Sasha_Spielberg:actor
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
16. ARCHITECTURE: ENTITY RECOGNIZER@WORK
➤Step 2
➤ User sentence: I like Spielberg and Jurassic Park
➤ Surface form1: Spielberg
➤ Candidate entities: Steven_Spielberg:director, Sasha_Spielberg:actor
➤ Context: Jurassic Park
sim2 (Jurassic Park,Steven_Spielberg:director) = 0.90
sim (Jurassic Park, Sasha_Spielberg:actor)= 0.15
Spielberg = Steven_Spielberg:director
1Basile, P., Caputo, A., Semeraro, G., Narducci, F.: Uniba: Exploiting a distributional semantic model for disambiguating
and linking entities in tweets. Making Sense of Microposts (# Microposts2015) (2015)
2Maximilian Nickel, Lorenzo Rosasco, Tomaso A Poggio, and others. 2016. Holographic Embeddings of Knowledge Graphs.
In The Thirtieth AAAI Conference on Artificial Intelligence (AAAI-16). 1955–1961.
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
17. ARCHITECTURE: RECOMMENDATION SERVICES
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ Recommendation algorithm
is PageRank with priors
➤ nodes are entities from
DBpedia (eg, actors, movies,
directors)
➤ Explanation1
➤ exploits links between liked
items and
recommendations in the
DBpedia graph
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
1Musto, C., Narducci, F., Lops, P., de Gemmis, M., Semeraro, G.
Linked open data-based explanations for transparent recommender systems.
(2019) International Journal of Human Computer Studies, 121, pp. 93-107.
18. ARCHITECTURE: RECOMMENDATION SERVICES
Intent Recognizer
Dialog Manager
Entity Recognizer
Sentiment
Analyzer
Recommendation
Services
➤ Critiquing
➤ the user can provides
complex feedback on
the recommended
items
➤ I like the movie Titanic,
but I don’t like the actor
Bill Paxton
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
19. EXPLANATION@WORK
American Epic Films
Tom
Hanks
Dystopian Films
The
Wachowskis
I recommend you Cloud Atlas because you often like films with Tom
Hanks as Saving Private Ryan and Da Vinci Code. In addition, you
sometimes like films directed by The Wachowskis as The Matrix.
dbpedia-owl:starring
dbpedia-owl:starring
dcterms:subject!
dcterms:subject!
dbpedia-owl:director
dbpedia-owl:director
dcterm
s:subject!
dbpedia-owl:starring
dcterms:subject!
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
20. THE FRAMEWORK@WORK
➤ three instances on Telegram (movie, music, and
book)
➤ @movierecsys2_bot
➤ @musicrecsys_bot
➤ @bookrecsys_bot
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
21. EXPERIMENTAL EVALUATION
➤ first session: in-vitro experiment on
two datasets
➤ second session: user study
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
22. EXPERIMENTAL EVALUATION: FIRST SESSION
➤ In-vitro experiment
➤ to assess the accuracy of each
component of our framework
and its impact on the
recommendation process
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
23. EXPERIMENTAL EVALUATION: GOAL
Test separately
➤ Intent Recognizer
➤ Entity Recognizer
➤ Sentiment Recognizer
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
24. EXPERIMENTAL EVALUATION: DATASET
bAbI by Facebook Research
collects utterances like
Beauty and the Beast, Aladdin, Schindler’s List, and The
Silence of the Lambs are movies I loved.
Would you recommend something I might like?
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
25. EXPERIMENTAL EVALUATION
Intent Recognizer test
Entities and Sentiments are set programmatically
Beauty and the Beast, Aladdin, Schindler’s List, and The
Silence of the Lambs are movies I loved. (Preference)
Would you recommend something I might like?
(Recommendation request)
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
26. EXPERIMENTAL EVALUATION
Entity Recognizer test
Intents and Sentiments are set programmatically
Beauty and the Beast, Aladdin, Schindler’s List, and The
Silence of the Lambs are movies I loved.
Would you recommend something I might like?
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
27. EXPERIMENTAL EVALUATION
Sentiment Recognizer test
Intents and Entities are set programmatically
Beauty and the Beast, Aladdin, Schindler’s List, and The
Silence of the Lambs are movies I loved.
Would you recommend something I might like?
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
28. EXPERIMENTAL EVALUATION
➤ Intent Recognizer Test
➤ Entity Recognizer Test
➤ Sentiment Recognizer Test
compared to
➤ Upper bound
recommendations generated by setting intents,
entities, and sentiments by code
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
30. EXPERIMENTAL EVALUATION: SECOND IN-VITRO EXPERIMENT
➤ Dataset released by Grouplens1
➤ collects recommendation requests of
real users to a (simulated) conversational
recommender system
➤ 694 sentences
➤ Results
➤ 7.4% intents (very difficult task, requests like “action
movies”,”exploitations films”, ”film with sharks”, ”i’m looking for a hard
sci-fi movie”)
➤ 64.39% entities
➤
1Kang, J., Condiff, K., Chang, S., Konstan, J. A., Terveen, L., & Harper, F. M. (2017, August). Understanding how people use natural langua
ask for recommendations. In Proceedings of the Eleventh ACM Conference on Recommender Systems (pp. 229-237). ACM.
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
31. EXPERIMENTAL EVALUATION: USER STUDY
➤ 50 users tested three domains and three
interaction modes:
➤ movie, book, and music
➤ natural language, buttons, and mixed
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
32. EXPERIMENTAL EVALUATION: USER STUDY
➤ Goal
➤ to assess the impact of using natural language in a CoRS
➤ Research Questions
➤ Can natural language improve a CoRS in terms of
cost of interaction?
➤ Can natural language improve a CoRS in terms of
quality of recommendations?
➤ What is the impact of each component of a CoRS to
the accuracy of the recommendations?
➤ What are the most critical aspects to consider when
modelling a natural language-based dialog for a
CoRS?
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
33. EXPERIMENTAL EVALUATION: USER STUDY
➤ Metrics
➤ Objective metrics
➤ MAP, Accuracy
➤ Number of Questions, Time per Question,
Interaction Time, Query Density
➤ Questionnaire based on the ResQue model [1]
and the questionnaire proposed in [2]
[1] P. Pu, L. Chen, R. Hu, A user-centric evaluation framework for recommender systems, in:
Proceedings of the fifth ACM conference on Recommender systems, ACM, 2011, pp. 157–164
[2] A. Silvervarg, A. Jonsson, Subjective and objective evaluation of conversational agents in learning
environments for young teenagers, in: Proceedings of the 7th IJCAI Workshop on Knowledge and
Reasoning in Practical Dialogue Systems, 2011
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
34. EXPERIMENTAL EVALUATION: USER STUDY
➤ Results 1/2
➤ Pure NL-based interfaces need to help the user
when precise input is needed
➤ Pure NL-based interfaces did not perform
significantly better than button-based
interfaces
➤ Mixed interactions (NL + buttons) generally
perform better than NL and buttons taken
individually
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
35. EXPERIMENTAL EVALUATION: USER STUDY
➤ Results 2/2
➤ Recognition of ratings and entities is a
crucial step for achieving a good accuracy,
specifically in the first phases of the interaction
➤ When users are asked to choose from a set of
predefined answers, this activity needs to be
facilitated in some way
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems
36. CONVERSE DATASET
➤ During the user study, we collected three
datasets of real dialogs useful for training-
testing Intent Recognizer, Entity Recognizer, or
Sentiment Analyzer
➤ Messages
➤ 5,318 movie domain
➤ 1,862 book domain
➤ 2,096 music domain
➤ split in recommendation requests, preferences,
criticisms, and explanation requests
➤ https://tinyurl.com/converse-dat
ConveRSE: A domain-independent Framework for building Conversational Recommender Systems