Crosslingual Content Scoring in Five Languages Using Machine-Translation and Multilingual Transformer Models

Abstract

AbstractThis paper investigates crosslingual content scoring, a scenario where scoring models trained on learner data in one language are applied to data in a different language. We analyze data in five different languages (Chinese, English, French, German and Spanish) collected for three prompts of the established English ASAP content scoring dataset. We cross the language barrier by means of both shallow and deep learning crosslingual classification models using both machine translation and multilingual transformer models. We find that a combination of machine translation and multilingual models outperforms each method individually - our best results are reached when combining the available data in different languages, i.e. first training a model on the large English ASAP dataset before fine-tuning on smaller amounts of training data in the target language.