PolitiKweli: A Swahili-English Code-Switched Twitter Political Misinformation Classification Dataset

In the age of freedom of speech, users of the various social media platforms post millions of unverified messages, resulting in misinformation. Despite these platforms’ set policies against misinformation, there is an alarming rise in misleading news dissemination. On political matters, misinformation online can result in defamation and in extreme cases, violence offline. Misinformation classification involves classifying text as fake or fact. Most of the existing studies address misinformation classification for posts in a single language only. Among most bilingual or multilingual social media users, code-switching such as Swahili-English, is common practice. This poses a threat to code-switching being used to spread misinformation. There is need for more research in low-resource languages such as Swahili, especially their use to spread misinformation. This study curated the PolitiKweli (dataset: https://github.com/jayneamol/kweli ) dataset, a Swahili-English code-switched misinformation classification dataset, containing 6,345 Swahili-English code-switched texts, 22,954 English texts and 211 Swahili texts. The texts are labeled as fake, fact or neutral as compared to fact-checked Twitter dataset also created as part of this study. The paper discusses the dataset curation process including data collection, data processing and data annotation. It also highlights the challenges during the annotation. The study develops a benchmark classification model based on pretrained language model BERT that achieves an F-score of 0.62. The results of the experiment show promising results on the usefulness of the Swahili-English code-switched misinformation classification models. When applied, the classification model can be able to flag instances of misinformation on social media platforms.