[bull-ia] PhD Position : Active Learning for Axiom Discovery

Open PhD Position on “Active Learning for Axiom Discovery” at the I3S
Laboratory in Sophia Antipolis.

Applications : a curriculum vitae together with a motivation letter
should be sent to Andrea Tettamanzi (andrea.tettamanzi@unice.fr) and
Célia da Costa Pereira (celia.pereira@unice.fr).

The deadline for application is Friday, May 17, 2019, through the
following link: http://edstic.unice.fr/edsticTheses2019/.

Context

The rise of deep learning has been made possible by the availability of
large computing powers at affordable prices. Its performance has made it
possible to solve problems that previously seemed out of reach,
especially in the field of perception, such as computer vision and the
analysis of natural language texts. However, tasks that require
reasoning still require knowledge that is symbolically represented.
Substantial breakthroughs can be achieved by combining deep learning
with symbolic reasoning. What hinders these developments is mainly the
cost of building knowledge bases rich not only in factual information,
which is relatively abundant and easy to capture, but also in rules,
constraints, and relationships (in summary, axioms) that make it
possible to infer implicit knowledge by reasoning.
The definition of the standards that collectively go under the name of
“semantic Web” has provided a technological framework to produce open
data as well as to define vocabularies and ontologies to make those data
interoperable. Nowadays, a huge mass of machine-readable knowledge is
available on the semantic Web, which opens up enormous opportunities for
research. An obvious thing to do is to analyze it and learn new
knowledge from it. Potential applications range from bio-informatics to
computational finance.

Objectives

The main goal of this thesis is to combine symbolic reasoning and active
learning to make the automatic discovery of axioms possible, thus
helping to overcome the knowledge acquisition bottleneck, while
radically changing the way we look at the semantic Web: instead of
postulating an a priori conceptualization of reality (i.e., an ontology)
and requiring that our knowledge about facts complies with it, we
propose to start from collected observations about facts and learn an
ontology which is able to account for them.
Discovering new axioms from a knowledge base containing both axioms
(background knowledge) and assertions (facts) may be regarded as a sort
of generate and test procedure, whereas candidate axioms are generated
following some heuristics and then tested to determine whether they are
compatible with the facts recorded in the knowledge base and consistent
with background knowledge.
The main problem is that testing a candidate axiom requires reasoning
with the knowledge base plus the axiom, which can be computationally
very expensive. Therefore, testing every candidate axiom would be
prohibitive. A way to overcome this problem is to learn a model capable
of predicting whether a candidate axiom will fit the knowledge base or
not, as a surrogate for reasoning. Reasoning, however, remains an option
which can be used, every once in a while, as an “oracle” to classify
those candidate axioms for which the trained model has a hard time to
make a reliable prediction. This looks like a perfect scenario for
applying active learning. Indeed, the intuition behind active learning
is that a machine learning algorithm with few labeled data can improve
its result if it is allowed to choose which data to use during the
learning process. The general context can be described as follows: (i)
the learner has few labeled data which it uses to construct an initial
model, (ii) a large set of unlabeled data is also available, (iii) an
“oracle” can be asked to associate labels to some unlabeled data. The
main problem is then to determine when and to which data the learner
will ask a label, the aim being to revise the model in order to improve
it. In this thesis, the reasoner will take the place of the “oracle” and
the new axioms to be tested will take the place of the set of unlabeled
data.

Organization

The thesis will be carried out in the SPARKS research group of the I3S
Laboratory, located in the technology park of Sophia Antipolis, on the
French Riviera/Côte d’Azur, famous worldwide as a landmark of science,
invention, innovation, and research.

References

[1] Thu Huong Nguyen, Andrea G. B. Tettamanzi: Learning Class
Disjointness Axioms Using Grammatical Evolution. EuroGP 2019: 278-294
[2] Duc Minh Tran, Claudia d’Amato, Binh Thanh Nguyen, Andrea G. B.
Tettamanzi: Comparing Rule Evaluation Metrics for the Evolutionary
Discovery of Multi-relational Association Rules in the Semantic Web.
EuroGP 2018: 289-305
[3] Dario Malchiodi, Andrea G. B. Tettamanzi: Predicting the
possibilistic score of OWL axioms through modified support vector
clustering. SAC 2018: 1984-1991
[4] Dario Malchiodi, Célia da Costa Pereira, Andrea G. B.
Tettamanzi: Predicting the Possibilistic Score of OWL Axioms Through
Support Vector Regression. SUM 2018: 380-386
[5] Andrea G. B. Tettamanzi, Catherine Faron-Zucker, Fabien Gandon:
Possibilistic testing of OWL axioms against RDF data. Int. J. Approx.
Reasoning 91: 114-130 (2017)
[6] Duc Minh Tran, Claudia d’Amato, Binh Thanh Nguyen, Andrea G. B.
Tettamanzi: An evolutionary algorithm for discovering multi-relational
association rules in the semantic web. GECCO 2017: 513-520
[7] Claudia d’Amato, Andrea G. B. Tettamanzi, Duc Minh Tran:
Evolutionary Discovery of Multi-relational Association Rules from
Ontological Knowledge Bases. EKAW 2016: 113-128
[8] Claudia d’Amato, Steffen Staab, Andrea G. B. Tettamanzi, Duc
Minh Tran, Fabien L. Gandon: Ontology enrichment by discovering
multi-relational association rules from ontological knowledge bases. SAC
2016: 333-338
[9] Andrea G. B. Tettamanzi, Catherine Faron-Zucker, Fabien L.
Gandon: Dynamically Time-Capped Possibilistic Testing of SubClassOf
Axioms Against RDF Data to Enrich Schemas. K-CAP 2015: 7:1-7:8
[10] Andrea G. B. Tettamanzi, Catherine Faron-Zucker, Fabien L.
Gandon: Testing OWL Axioms against RDF Facts: A Possibilistic Approach.
EKAW 2014: 519-530
[11] Burr Settles. Active learning literature survey. Technical
report, University of Wisconsin Madison Department of Computer Sciences,
2009


Prof. Andrea G. B. Tettamanzi
Université Nice Sophia Antipolis, Dept. Informatique
Equipe SPARKS – WIMMICS (I3S – INRIA)
Tel. +33 (0)4 89 15 42 96, Mobile +33 (0)7 81 86 26 40
URL: “http://www.i3s.unice.fr/~tettaman/”

———————————————————————
Desinscription: envoyez un message a: bull-ia-unsubscribe@gdria.fr
Pour obtenir de l’aide, ecrivez a: bull-ia-help@gdria.fr