Veuillez trouver ci-dessous une proposition d'un stage M2 financé par
le projet TIRIS Text-Ecop.

Master thesis offer: Polarization in Natural Language

Student profile: Master 2 student in Machine Learning, Data Science,
Natural Language Processing.

Keywords: Natural Language Processing, Machine Learning

Supervision and Contacts:
Stergos Afantenos, stergos.afantenos@irit.fr
Karine Van Der Straeten, Karine.Van-Der-Straeten@TSE-fr.eu
Seebastien Gadat, Sebastien.Gadat@TSE-fr.eu

Working place Institut de Recherche en Informatique &
Toulouse School of: Economics

Length: 5 to 6 months, starting in February/March 2025

Remuneration: Standard internship grant : 550 euros/month.

Context: The proposed Masters thesis will be developed within the
context of the TIRIS project entitled "Text-mining et nouvelles
stratégies de mesure en économie politique" (Text-Ecop), a
multi-disciplinary research project betweenToulouse School of Economics
and Université Paul Sabatier. The project deals with the use of applied
mathematics, computer science and A.I. on textual data analysis for
applications in various fields, including political economy as a
specific focus in Text-Ecop. Objectives and Contents of the Internship:
Polarisation is a societal phenomenon which can take many forms.
Usually it is manifested by the differenttopic choices that opposite
parties choose to discuss or focus on. On the other hand, these parties
are not isolated from society and thus they often need to discuss the
same topic. In such cases, polarization is manifested via the
linguistic choices they make in order to express the same concept.
Consider for examplethe following two phrases (retrieved from chat GPT
on Sep 9th, 2024) :

A: The city invested in much needed green spaces to promote wellbeing
and community activities.
B: The city wasted millions on useless patches of grass that no-oneeven
wants.

In this example we can clearly see the framing bias that each party
has, as manifested by the lexical choices that were made in order to
express the same concept. Furthermore, we can see that this framing
bias lies on aspectrum of polarization (from neutral to more polarized).
To make it more explicit, in this particular case we have two instances
of polarized lexical choices. The first one concerns the action that is
denoted by the main verbs of the sentence ("invested" vs "wasted")
while the second concerns the description of the object ("much needed
green spaces" vs "useless patches of grass"). Note, at this point, that
a third party might have made another lexical choice such as "spent" or
something else more or less neutral. The same holds between "green
spaces" and "useless patches of grass." In a certain sense, an analogy
is formed between these lexical choices, both referring to the same
concept, but expressed in a way that reflects polarization. The goal of
this masters thesis is the development of Machine Learning models based
on Neural Networks or other approaches that will enable us to identify
the existence of framing bias, expressed via particular lexical
choices, showing a degree of polarization. The dataset that we will use
in order to build the Machine Learning models will be the NEUS corpus
described in Lee et al. (2022).In terms of measuring the degree of
polarity, various approaches, based on thedistance (Euclidean or other)
in a latent embedding space of the aforemen-tioned lexical choices, can
be examined in order to provide a numerical valuereflecting the degree
of polarization. The successful candidate is expected to have a
background in Machine Learn-ing or Computational Linguistics. This
masters thesis has the possibility to beextended to a PhD thesis funded
by the same project, depending on the resultsobtained.

References
Nayeon Lee, Yejin Bang, Tiezheng Yu, Andrea Madotto, and Pascale Fung.
Neus:Neutral multi-news summarization for mitigating framing bias. 2022.