Harnessing Large Language Models to De-escalate Online Polarisation (DOP)

Socio-political context

Polarization has been a significant challenge for modern liberal democracies since their inception. Throughout the twentieth century, the world witnessed major events such as the two world wars, the civil rights movement, and the rise of extremist political movements, all of which were marked by deep ideological divides and societal fragmentation. The emergence of the Web and social media in recent decades has intensified and accelerated this phenomenon, despite the progress made towards greater social cohesion through direct connectivity with distant people, access to diverse information, pluralism, and democratic participation. Paradoxically, this has also resulted in unprecedented levels of online polarization.

The dangers of online polarization extend far beyond the erosion of public trust and the proliferation of disinformation. Perhaps the most significant risk is the potential for online polarization to lead to radicalization and violent extremism, which can pose a serious threat to democratic societies. Online polarization is often exploited by extremist groups who seek to undermine democratic institutions by manipulating public opinion, exploiting social divisions, and sowing discord. This exploitation of online polarization represents one of democracy’s Achilles’ heels as it undermines social cohesion and can result in the weakening of democratic institutions and values.

 

Computer science approach

With this project, we will develop a Machine Learning (ML) system based on Large Language Models (LLMs) that will measure the online polarisation of social media users based on the content they consume and produce in natural language.

Potential research directions include (but are not limited to): 

  • Identifying (political) debates subject to division online
  • Exploring relevant datasets (Reddit, Twitter, Wikipedia, etc.) that will be used to train and evaluate our models
  • Collecting data from users showing different levels of polarization and analyzing their behavior based on the content produced and consumed
  • Training / evaluating the predictive capabilities of the latest neural natural language processing methods and models
  • Assessing the few-shot learning (and zero-shot) performances of Large Language Models
  • Comparing methods to gain visibility in this exploratory domain 

Software development objectives:

  • Designing and implementing a pipeline that will help to analyze individual users’ polarisation
  • Developing flexible tools to predict and visualize users’ polarisation levels
  • Developing a Python package that will be used by the research community for further investigation

PREREQUISITES

  • Familiar with Python, PyTorch, HuggingFace librairies
  • Creativity, initiative and proactive spirit
  • Knowledge of Linux and related tools

PREFERRED, BUT NOT REQUIRED

  • Experience in Machine Learning and Deep Learning
  • Experience in Natural Language Processing

Send me your CV: [email protected]