The document discusses how knowledge graphs are needed for dialogue systems to avoid issues like the "mushroom effect" where answers provided by a system become increasingly vague and less applicable. It provides an example of a dialogue exchange where a system is unable to determine the architectural style of a building mentioned without access to a knowledge graph. The document argues that knowledge graphs allow systems to better constrain answer spaces and perform more complex question answering and reasoning by aggregating information across graphs.

1. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
The Mushroom Effect
Or Why You Need Knowledge Graphs for Dialogue Systems
Dr. Michael Galkin
Fraunhofer IAIS & Smart Data Analytics
Dresden

2. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Outline
Background
Mushrooms?
With KG Flavors
About us: CEE AI Dresden

3. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Background
We build conversational AI platforms

4. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Background
We build conversational AI platforms
Powered by knowledge graphs

5. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Background
We build conversational AI platforms
Powered by knowledge graphs
Obtained by integrating heterogeneous data

6. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Mushrooms?
Berlin Hbf
What is this building? Q

7. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Mushrooms?
Berlin Hbf
What is this building? Q
This is Berlin HauptbahnhofA

8. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Mushrooms?
What is this building? Q
This is Berlin HauptbahnhofA
What is its architectural style? Q
Berlin Hbf

9. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
What is its architectural style?

10. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
What is its architectural style?
Wikipedia

11. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
What is its architectural style?
SQuAD

12. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
What is its architectural style?
MS MARCO

13. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
What is its architectural style?
HotPot QA

14. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Mushrooms!
What is this building? Q
This is Berlin HauptbahnhofA
What is its architectural style? Q
MushroomA

15. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Mushrooms! What is its architectural style? Q

16. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Mushrooms! What is its architectural style? Q
✓ its = Berlin Hauptbahnhof
✓ In Berlin
✓ Concept ~ Style
❌ Applicable to the city concept, not
architectural style of the station
❌ Close, but not correct

17. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
With KG flavors
Can mushrooms be
an architectural
style?

18. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
With KG flavors
Can mushrooms be
an architectural
style?
Probably not
If not, what can be
an appropriate
value?

19. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
So why you need graphs?
Implicit or explicit constraints on produced answersHow many children
does Berlin Hbf have?

20. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
So why you need graphs?
Implicit or explicit constraints on produced answers
- reduce candidates space
- help to fight the mushroom effect
- ontologies help
How many children
does Berlin Hbf have?
Train stations
don’t have kids

21. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
So why you need graphs?
Implicit or explicit constraints on produced answers
- reduce candidates space
- help to fight the mushroom effect
- ontologies help
Complex QA via (sub)graphs aggregations
What is the busiest
train station in
Germany?
How many children
does Berlin Hbf have?
Train stations
don’t have kids

22. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
So why you need graphs?
Implicit or explicit constraints on produced answers
- reduce candidates space
- help to fight the mushroom effect
- ontologies help
Complex QA via (sub)graphs aggregations
What is the busiest
train station in
Germany?
select ?station ?visits where {
?station wdt:P31 wd:Q18543139 . # central stations
?station wdt:P17 wd:Q183 . # in Germany
?station wdt:P1373 ?visits . # daily visits
} ORDER BY DESC(?visits) LIMIT 1 # sort
How many children
does Berlin Hbf have?
Train stations
don’t have kids
Hamburg Hbf

23. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
So why you need graphs?
Graphs significantly improve reasoning
compared to sole natural language inference
Takeaway 1

24. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
So why you need graphs?
Graphs significantly improve reasoning
compared to sole natural language inference
Reasoning outcomes are explainable and
traceable
Takeaway 1
Takeaway 2

25. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
KDDS
Knowledge-driven
in-car dialogue
system (EN/DE)
Full DBpedia
2019 (wikidata
branch)
> 50M entities
> 4B triples

26. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
KDDS @
Hannover-Messe
Building KGs from enterprise
sources is still a challenge

27. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
KDDS @
Hannover-Messe
Building KGs from enterprise
sources is still a challenge
Depth of knowledge vs
variety of domains

28. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
KDDS @
Hannover-Messe
Building KGs from enterprise
sources is still a challenge
Depth of knowledge vs
variety of domains
Explainability is crucial

29. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Fraunhofer IAIS
Standort Dresden
ML2R
National Competence Center for
Machine Learning Rhein-Ruhr
Fraunhofer-Alliance Big Data AI
The biggest Fraunhofer alliance
with > 30 institutes led by IAIS
AI4EU
EU Lighthouse
Project for AI
International Data Spaces
Association
Data sovereignty for Big
Data and AI, 100+
companies
CEE AI
Center for Explainable and Efficient AI
Technologies with TU Dresden
Fraunhofer Center for
Machine Learning
IAIS-led part of Fraunhofer
Cluster of Excellence
Cognitive Internet
Technologies

30. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Fraunhofer IAIS Dresden
Center for Efficient and Explainable AI
Prof. FitzekProf. Lehmann
● Efficient AI
○ Hardware and software are usually
decoupled instead of embedded
○ 100-times more energy efficient by 2030
● Explainable AI
○ Make AI more traceable
○ Make AI decisions more explainable

31. Michael Galkin 7. Leipziger Semantic Web Tag 22.05.2019
Center for Explainable and Efficient AI