Masters Thesis Defense: Zhanfan Zhou

The spread of offensive language has become a severe social problem and may stress unmeasurable mental health illness. The rapid usage of social media worsens the situation. We develop a lite but robust offensive language identification system and evaluate the system on two SemEval offensive language identification shared tasks: SemEval 2019 Task 6 and SemEval 2020 Task 12. In order to take the advantage of a large semi-supervised dataset, and reduce the processing complexity of such huge data, we investigate approaches to adapt a model to the silver standards via curriculum learning and dynamic loss functions. By adapting a model to such data with the curriculum learning or dynamic loss functions, the systems are capable of scattering the focus properly on data of different difficulty levels. Experiments show both help the model learn effectively and acquire more messages from the hard cases without impairing the performance on easy cases. The best run on each task achieves competitive F1 score of 81.6% and 91.7% on the official test data of SemEval 2019 Task 6 and SemEval 2020 Task 12 respectively with at least 50% parameters and less data overhead, compared to the state-of-the-art systems.