基於英語第二外語學習者之自動口說評測模型

Abstract

none

In the post-COVID-19 era, globalization and online education have increased the demand for language-learning tools that enable independent skill assessment and improvement. This trend has spurred considerable interest and research in the field of automated speech assessment (ASA). The general goal of ASA is to deliver a consistent, objective evaluation of the spoken language proficiency of an L2 learner or test-taker. Unlike most previous work, which treats ASA as a nominal multiclassification task and thus neglects the sequential nature of proficiency grades, this study aims to explore the notion of ordinal-related optimization in ASA. In particular, we aim to enhance ASA performance by examining two critical issues: (1) the impact of using ordinal-related optimization instead of hard labels in the optimization of ordinal classification for ASA and (2) the effects of integrating self-supervised learning with handcrafted indicator features through a novel modeling paradigm. Our results show that the proposed model significantly enhances performance compared to existing strong baselines. The enhancement is apparent in both the test dataset of seen prompts and that of unseen prompts, indicating our method's robust generalization and adaptability.