Are Knowledge Graphs Ready for the Real World? Dylan’s Dagstuhl Trip Report
Dylan Van Assche’s trip report on the “Are Knowledge Graphs Ready for the Real World? Challenges and Perspective” Dagstuhl Seminar of Ferbruary 2024
Months ago I received an invite for a Dagstuhl Seminar
‘Are Knowledge Graphs Ready for the Real World? Challenges and Perspective’
to represent my research group KNoWS at the seminar.
This was an amazing opportunity for KNoWS to contribute
to the research domain of Knowledge Graphs.
Dagstuhl Seminars are a way to interact with experts from other universities
in a relaxed manner and can work intensively on specific topics in groups.
This way, a lot of topics quickly move forward because you have
the right people and their expertise, all in the same place!
This seminar was organized by David Chaves-Fraga,
Oscar Corcho,
Anastasia Dimou,
and Maria-Ester Vidal.
Topics
During the first day, the organizers introduced the seminar
and several groups were formed to work on various topics
to bring Knowledge Graphs (KG) into the real world:
- KG Construction & Quality: construction of KGs, evaluating data quality inside KGs, user involvement.
- Federated Semantic Data Management: federation between KGs regarding querying, access control.
- Enhance Machine Learning: empowering Large Language Models (LLMs) with KGs.
- Knowledge Engineering: focus on education around knowledge, KGs, etc.
- Project KG Management: best practices for KG construction, maintenance, etc. inside projects.
Within our KNoWS team, we have a keen interest in KG construction,
federated querying and access control and Knowledge engineering.
During the week, I was mostly active in the KG Construction & Quality
group because I am involved daily in KG construction research and
wanted to get more expertise from the other participants.
Users or should I say people?
During the first days of the seminar, several times participants highlighted
that they should involve users or better say people into the process of
creating, consuming, and refining a KG.
We have spent a lot of time already on the technical part of KGs,
but not so much on the social aspect which blocks the adoption of KGs
in industry and among developers. We agree that investing in training,
books, workshops, and involving non-experts is key for adoption of KGs.
LLMs cannot be ignored
With the introduction of LLMs e.g., ChatGPT, a lot of things were
suddenly solved, the hype was real. Nowadays, people seem to realise
that LLMs also have their limitations, such as hallucinations.
They are actively looking for a way to reduce these hallucinations.
Instead of pushing LLMs away, we as the community should rather embrace them
because KGs can provide the necessary semantic descriptions
to improve the LLMs’ accuracy, thus reducing their hallucinations.
KG lifecycle & ecosystem
Involving the people into the process, puts us on a path of defining properly
what the lifecycle and ecosystem is of a KG. Once we have this defined,
we know in which parts of the process we should involve certain people.
For example: domain experts for modeling the KG’s ontology happens
at the beginning of the process, but they are also needed
to audit the KG’s results.
Main take-aways
Some main take-aways I gathered during this seminar for KGs:
- LLMs & KGs: LLMs were all over the place in the seminar,
since the launch of ChatGPT you cannot ignore them anymore.
Luckily, there’s still a good opportunity for KGs given that LLMs suffer
from hallucinations. KGs provide the semantic structure to LLMs which
they desperately need to increase their accuracy.
- KG lifecycle needs users: We need to involve users way more
into the KG lifecycle from construction until consumption.
Users can provide so much vital information regarding
the domain modeled in the KG, the data quality, etc.
- Workflow & best practices: All of the people at the seminar
are constructing, using, modifying KGs, but we still haven’t written
down a good workflow and its corresponding best practices.
What’s next?
Within KNoWS we will follow up the discussions
from this Dagstuhl Seminar during the next coming months
to write the Dagstuhl Seminar’s report.
We are already planning a publication around the KG ecosystem and lifecycle,
together with other participants in the seminar, stay tuned!
If you want to collaborate on any of these topics,
feel free to reach out to us!
You can contact me (Dylan Van Assche - dylan.vanassche@ugent.be)
or any of my PhD advisors (Pieter Colpaert - pieter.colpaert@ugent.be
and Anastasia Dimou - anastasia.dimou@kuleuven.be).