On the Effectiveness of Curriculum Learning in Educational Text Scoring

Automatic Text Scoring (ATS) is a widely-investigated task in education. Existing approaches often stressed the structure design of an ATS model and neglected training process model. Considering difficult nature this task, we argued that performance could be potentially boosted by carefully selecting data varying complexities process. Therefore, aimed to investigate effectiveness curriculum learning (CL) scoring educational text. Specifically, designed two types difficulty measurers: (i) pre-defined, calculated measuring sample's readability, length, number grammatical errors or unique words it contains; (ii) automatic, based on whether epoch can accurately score samples. These measurers were tested both easy-to-hard hard-to-easy paradigms. Through extensive evaluations widely-used datasets (one for short answer other long essay scoring), demonstrated (a) CL indeed boost state-of-the-art models, maximum improvement up 4.5%, but most improvements achieved when assessing easy answers; (b) pre-defined measurer contained text sample tended outperform across different