Research Internship on Multimodal Deep Learning Techniques for the
Identification of Asymmetric Relations

Gaël Dias, Youssef Chahir and Houssam Akhmouch

CNRS GREYC UMR 6072 - Normandy University

Internship Description

Traditionally in classification, a learning example refers to a unique
object 0 A that must be categorized in some class given an input set
of features. For example in sentiment classification, a text
transformed in some learning representation can be classified either
as positive, negative or neutral [4].

But many other situations exist. In this internship, we are
particularly interested in identifying relations between pairs of
objects. In this case, a learning example (0 A r 0 B ) refers to a
pair of objects 0 A and 0 B that must be transformed in some
learning representation so that the model can predict if some relation
r holds between them, or not. For instance, in image classification, a
model may have to predict if two patches are similar or not [10].

Within this context, a vast majority of works have been tackling
symmet- ric relations (e.g. matching images, synonymy detection),
while asymmetric relations (e.g. textual entailment, temporal image
ordering) have received less attention. Indeed, it has been shown that
asymmetry is usually a difficult prob- lem [8] as it involves the
direction of the relation. So, if two objects O A and O B are in an
asymmetric relation , let's say (O A -> O B ), then the learning
example (O A -> O B ) is a positive example, while (O B -> O A ) is a
negative example.

In this internship, we are particularly interested in designing
learning models that can positively handle asymmetry. Different work
directions will be studied such as (1) the development of specific
asymmetric deep learning architectures [2, 9], (2) the computation of
semantic feature spaces capable of handling asymmetry [7, 1, 3, 6]
or (3) a combination of both ideas [5]. In particular, multimodal
(text, image and text/image) experiments will be performed on gold
standard data sets in order to evaluate the performance of the
different proposed models.

Candidate Profile

The successful candidate must pursue Master or Engineering School
level stud- ies in Data Science, Computer Science, Applied
Mathematics, or related scientific fields and show strong background
in Mathematics, Machine Learning and Programming. Additional knowledge
in Deep Learning (both theoretically and practically) will be highly
appreciated as well as experience in Natural Language Processing
and/or Computer Vision.


Internship Organization

The internship will start at the beginning of 2019 (January, February
or March) and will last up to 6 months. It will take place at the CNRS
GREYC UMR 6072 Laboratory in Caen (France). Some visits to the
internship partner Crédit Agricole Brie Picardie in Serris near Paris
will be planned. The candidate will be compensated following the rules
in force for internships. Note that the internship is subject to the
establishment of a preliminary internship agreement.

Internship Perspectives

Depending on the obtained results and the dedication/motivation of the
success- ful candidate, there shall be a possibility to pursue PhD
studies in collaboration with all the internship partners, i.e. CNRS,
Normandy University and Crédit Agricole Brie Picardie.


Application Procedure

Applicants are requested to submit their application with an academic
CV, copies of academic degree records and certificates, and two
potential references (if possible). Applications must be sent directly
to the internship coordinators:

Gaël Dias (gael.dias@unicaen.fr), Youssef Chahir (youssef.chahir@unicaen.fr)
and Houssam Akhmouch (houssam.akhmouch@unicaen.fr).

Note that the CNRS GREYC UMR 6072 is committed to being a fully
inclusive institution which actively recruits staff from all sectors
of society. It is proud to be an equal opportunities employer and
encourages applications from everybody, regardless of race, sex,
ethnicity, religion, nationality, sexual orientation, age, disability,
gender identity, marital status/civil partnership, pregnancy and
maternity, as well as being open to flexible working practices.

References

[1] Haw-Shiuan Chang, ZiYun Wang, Luke Vilnis, and Andrew McCallum.
Distributional inclusion vector embedding for unsupervised hypernymy de-
tection. In Conference of the North American Chapter of the Association
2for Computational Linguistics: Human Language Technologies (NAACL-
HLT), pages 485-495, 2018.
[2] Goran Glava¨ and Simone Paolo Ponzetto. Dual tensor model for detecting
asymmetric lexico-semantic relations. In Conference on Empirical Methods
in Natural Language Processing (EMNLP), pages 1757-1767, 2017.
[3] Goran Glava¨ and Ivan Vulic. Explicit retrofitting of distributional word
vectors. In 56th Annual Meeting of the Association for Computational
Linguistics (ACL), volume 1, pages 34-45, 2018.
[4] Siwei Lai, Liheng Xu, Kang Liu, and Jun Zhao. Recurrent convolutional
neural networks for text classification. In 29th AAAI Conference on Arti-
ficial Intelligence (AAAI), volume 333, pages 2267-2273, 2015.
[5] Simon Moura, Amir Azarbaev, and Massih-Reza Amini. Heterogeneous
dyadic multi-task learning with implicit feedback. In 25th International
Conference on Neural Information Processing (ICONIP), 2018.
[6] Marek Rei, Daniela Gerz, and Ivan Vulic. Scoring lexical entailment with
a supervised directional similarity network. In 56th Annual Meeting of the
Association for Computational Linguistics (ACL), pages 638-643, 2018.
[7] Ivan Vulic and Nikola Mrksic. Specialising word vectors for lexical entail-
ment. In Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies (NAACL-
HLT), pages 1134-1145, 2018.
[8] Ekaterina Vylomova, Laura Rimell, Trevor Cohn, and Timothy Baldwin.
Take and took, gaggle and goose, book and read: Evaluating the utility
of vector differences for lexical relation learning. In 54th Annual Meeting
of the Association for Computational Linguistics (ACL), pages 1671-1682,
2016.
[9] Qi Wang, Tong Ruan, Yangming Zhou, Daqi Gao, and Ping He. An
attention-based bi-gru-capsnet model for hypernymy detection between
compound entities. arXiv preprint arXiv:1805.04827, 2018.
[10] Sergey Zagoruyko and Nikos Komodakis. Learning to compare image
patches via convolutional neural networks. In IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR), pages 4353-4361, 2015.