Download PDFOpen PDF in browser

Mirela at CheckThat! 2024: Check-Worthiness of Tweets with Multilingual Embeddings and Adversarial Training

EasyChair Preprint 13860

6 pagesDate: July 8, 2024

Abstract

Accurately assessing the credibility and significance of texts is crucial in today's digital age where misinformation and disinformation abound, especially in social media. In this  paper, we propose an approach for check-worthiness of tweets that integrates adversarial learning techniques to optimize classification accuracy and language identification simultaneously. We conduct fine-tuning of DistilBERT-multilingual and XLM-RoBERTa-base for English, Dutch, Spanish, and Arabic to allow the models to adapt to the intricacies of different languages. Furthermore, we introduce an adversarial training approach to enhance the performance of multilingual sentence transformers, ensuring their effectiveness across linguistic contexts. The proposed approach ranks 4th in Dutch, 11th in Arabic, and 16th in English with an F\textsubscript{1}-score (positive class) of 0.65, 0.48, and 0.66, respectively.

Keyphrases: Multilingual Classification, Sentence Transformers, check-worthiness, disinformation, misinformation, social media

BibTeX entry
BibTeX does not have the right entry for preprints. This is a hack for producing the correct reference:
@booklet{EasyChair:13860,
  author    = {Mirela Dryankova and Dimitar Dimitrov and Ivan Koychev and Preslav Nakov},
  title     = {Mirela at CheckThat! 2024: Check-Worthiness of Tweets with Multilingual Embeddings and Adversarial Training},
  howpublished = {EasyChair Preprint 13860},
  year      = {EasyChair, 2024}}
Download PDFOpen PDF in browser