Integrating ordinal, multitask deep learning with faceted item response theory: debiased, explainable, interval measurement of hate speech