Combating propaganda texts using transfer learning

Abstrak

Recently, it has been observed that people are shifting away from traditional news media sources towards trusting social networks to gather news information. Social networks have become the primary news source, although the validity and reliability of the information provided are uncertain. Memes are crucial content types that are very popular among young people and play a vital role in social media. It spreads quickly and continues to spread rapidly among people in a peer-to-peer manner rather than a prescriptive. Unfortunately, promoters and propagandists have adopted memes to indirectly manipulate public opinion and influence their attitudes using psychological and rhetorical techniques. This type of content could lead to unpleasant consequences in communities. This paper introduces an ensemble model system that resolves one of the most recent natural language processing research topics; propaganda techniques detection in texts extracted from memes. The paper also explores state-of-the-art pretrained language models. The proposed model also uses different optimization techniques, such as data augmentation and model ensemble. It has been evaluated using a reference dataset from SemEval-2021 task 6. Our system outperforms the baseline and state-of-the-art results by achieving an F1-micro score of 0.604% on the test set.