DIGITAL.CSIC. Repositorio Institucional del CSIC
Publication
Artículo científico (JournalArticle)
Geopolitically-Informed multiModal BERT for propaganda detection in political tweets
IberLEF 2024, September 2024, Valladolid, Spain., People are becoming increasingly dependent on social media for their political information, with political accounts
often being their primary source. The overwhelming flood of information makes it challenging for the average
person to compare sources, critically evaluate the content, and form well-informed opinions. This context leave
individuals susceptible to manipulation by political actors who exploit these vulnerabilities. Following this line,
we presented a model for the DIPROMATS Task 1 of IberLEF 2024, which challenges participants to detect
propaganda from diplomatic tweets in English and Spanish. Our approach proposed a MultiModal BERT-like
model that combined a pre-trained TwHIN-BERT encoder with a selection of contextual features aimed at helping
distinguishing the underlying geopolitical intentions behind propaganda use and at mitigating typical issues
in text-stream scenarios, as spurious correlations or concept drift. Although our results do not reflect a perfect
handling of this issues, we achieved competitive results that positioned us in the middle of the rankings, with
better performance for the Spanish track. Adding contextual information to TwHIN-BERT improves the F1 score
for the propaganda class, motivating further research in this domain., This project has received funding from the European Union’s Horizon 2020 research and innovation
programme under the Marie Skłodowska-Curie grant agreement Nº 101034328., Peer reviewed
Project:
EC / H2020 101034328
View at: http://hdl.handle.net/10261/378034