Enhancing Emergency Situational Awareness Models by Correcting Mislabelled Data

The data labels that feed machine learning are usually the result of someone's painstaking labor. This task demands focused attention and is not inherently enjoyable, which is why many platforms outsource it. However, we cannot ascertain a 100% accuracy in the result of labeling. Sometimes we use datasets that can be downloaded in bulk from the internet; however, it cannot be ensured that the labeled values are accurate as well. Even though if someone were to collect only the labels with errors, we could relabel them correctly, but finding the ones that are wrong in a large number of labels is not trivial. In this paper, we propose a ensure that method to learning is successful even if there is some error in the labels, and present the effectiveness of the method through proposed experiments.