Pseudo Fake Data Generation using Round-Trip Translation for Fact Verification on Political Debates

In this work, data augmentation methods for the fact verification task on political debate are proposed. One of the methods employs round-trip translation to generate natural and fluent claims as invented by humans. In our experimental evaluation, we investigated if the accuracy of classification could be improved by retraining the baseline model, which was trained by using only existing training data, by adding the training data automatically generated by the proposed methods. The result showed that the fact verification models were successfully improved by using the pseudo training data created by our methods. It also showed that using round-trip translation in the fake manipulation process helps to create effective training samples for fake detection.