Automated poetry scoring using BERT with multi-scale poetry representation

Abstract

Automated poetry scoring is an emerging task in automated text scoring, which is receiving increasing attention in AI for education. Poetry is distinct from other text in its complexity and specialty in language feature moreover, poems are usually rated from multiple criteria besides the overall impression. However, few existing methods to the best of our knowledge have considered a tailored text representation model for encoding poetry. Moreover, the lack of large poetry corpus and extensive labelled data is another major constraint to construct an effective poetry scoring model. To address such problems, we proposed BERT-based models with multi-scale poetry representation. In addition, we employ multiple losses and R-Drop strategy to align the distribution of manual and model scoring and mitigate the tendency of consistent score in poems. Experiment results demonstrate that our model with multi-scale poetry representation stands out when comparing with single-scale representation model.