A Comparative Study of Ensemble Approaches to Fact-checking for the FEVER Shared Task
Researcher: Oluwabamigbe Oghenetega Oni, University of the Witwatersrand, Johannesburg
Supervisor: Prof. Terence van Zyl, University of the Witwatersrand, Johannesburg
The surge of information globally has motivated for automated rumour detection. Since misinformation is rumour on incorrect information, we use fact-checking when detecting it. The FEVER-shared task is the fact-checking task used for our comparative study. The task is divided into Document Retrieval, Sentence Selection, and Claim Verification components. We standardise TF-IDF for document retrieval, create our pipelines of one of two Sentence Selection options and one of two Claim Verification options. We then evaluate each unique pipeline on the FEVER score, compare our four pipelines to the baseline and state of the art from the FEVER Shared Task. We find that our 2-way classification task using the Siamese BiLSTM achieves better Evidence Retrieval F1 scores than the state of the art models, and that our pipeline combinations rival the state of the art for the Shared Task.