Title : A study on sampling datasets for linguistic diversity
(Natural Language Processing)

Subject in brief :
This internship can be adapted to the candidate preferences among two
possibilities:
1.  Impact of diversity-driven raw data sampling on downstream
    applications (possibly based on deep learning)
2.  When evaluating NLP systems, how much diversity does the test
    corpus add to train and how does it impact the systems scores?
For the first proposal, the intern will focus on a task involving the
discovery of multiword expressions (MWEs). MWEs are sets of words whose
meaning is not necessarily directly related to the meaning of the
individual words ("They spilled the beans"). One possible method of
discovery is to select sentences in which a word (e.g. "spill ")
appears, then cluster these sentences to group together those that have
a similar meaning. In theory, we could have a cluster where all the
sentences with the MWE "spill the beans", and another cluster with the
sentences where the word "spill " appears in the sense "to make fall".
The selection of these sentences is currently random. We would like to
investigate diversity-driven data-sampling methods to study the impact
of this selection. For the second proposal, we are looking at the
diversity of one corpus compared to another. If we have two corpora, a
train corpus and a test corpus, does the test increase diversity of the
train corpus? We also want to study the impact of a biased test on the
results of a task linked to these train and test corpora. Different
diversity can be measured in different linguistic objects (words,
syntactic subtrees, word meanings, etc...). Both proposals include
development, although the second one is more focused on linguistic
analysis. Possible outcomes of this internship : a pipeline including
diversity quantification and a scientific paper.

A full description of the subject is available here :
https://perso.limsi.fr/savary/Projects/2024_Internship_subject.pdf.

Location : France, Orsay, Université Paris Saclay, LISN

Duration : 6 months (starting around March)

How to apply  : Applicants should send a CV, transcripts of bachelor
and master grades to Louis Estève, Manon Scholivet, Agata Savary and
Thomas Lavergne (firstname.lastname@universite-paris-saclay.fr without
accents) before January 6.
Applicants may also link past work, or git repositories of their code.