DMRC Visitor Seminar: Assoc. Prof. Fabio Giglietto, University of Urbino, Italy (Brisbane, 22 June 2023)
Unveiling Political Discourse: Harnessing OpenAI Models for Topic Modelling in Social Media Analysis of Italian Elections
The advent of OpenAI’s ChatGPT signalled a watershed moment in our research methodologies, particularly in efficiently understanding topics discussed on social media. While similar models, such as BERT, existed, they necessitated a fine-tuning process to adapt to specific languages and domains—a process that is both computationally demanding and time-consuming. Moreover, despite a plethora of domain-oriented pre-trained models, selecting an optimal model often proves challenging and can lead to unsatisfactory results, necessitating a specialised training session for the task at hand.
In contrast, OpenAI’s models are versatile, excelling in summarisation and classification tasks across a multitude of domains and languages. Importantly, these models can be customised to meet researchers’ needs by fine-tuning them—a process that is straightforward, leverages server-side computational resources, and is both rapid and cost-effective.
In this presentation, I explore the three distinct methods we employed using OpenAI models to identify the most salient topics circulated via Facebook links in the run-up to the two most recent Italian general elections. I will elaborate on our techniques for constructing a classifier to detect political links shared on Facebook, performing a cluster analysis on the document embeddings provided by OpenAI’s API embedding endpoint for political links, and autonomously labeling the identified clusters. We also utilised Meta’s URL Shares Dataset to characterise each cluster based on their exposure and interaction patterns.
Although our methods were applied to the context of the Italian elections, they are easily adaptable to other countries and scenarios. While this talk is predominantly methodological, it will also touch on the broader social implications of this approach, including its impact on future election strategies, online influence operations, and public policy decision-making.
