結合BERT與Wav2vec 2.0提升第二外語受試者之自動英語口說評測

Enhancing Automated English Speaking Assessment for L2 Speakers with BERT and Wav2vec2.0 Fusion

英語逐漸作為許多國家的第二語言(English as a Second Language, ESL)，同時也帶動電腦輔助語言學習的發展，近年來又以發展自動口語評測較為熱門。然而，英語口語能力評測的過程需要耗費許多人力，也相當費時。因此，建立出一套自動英語口語評分的方法不但能節省人力、時間，亦能提供更加一致的評估標準。本研究中我們使用公開資料集ICNALE，建構一套融合BERT和Wav2vec 2.0模型來進行自動英語口說能力分級。研究結果顯示，整合文字與語音的模型表現優於以人工轉錄訓練的BERT模型和單獨的Wav2vec 2.0模型。

Due to the increasing popularity of English as a second language, there has been a growing interest in developing Compute-assisted Language Learning (CALL) applications that focus on automated assessing of spoken language proficiency. In the past, evaluating English speaking proficiency has been a time-consuming and labor-intensive process. Therefore, developing an efficient method for automated grading can establish consistent evaluation standards in a more timely and cost-effective manner. In this study, we explore the fusion of BERT and Wave2vec2.0 modeling strategies to assess holistic English speaking proficiency scores, withe an extensive set of experiments conducted on the publicly available ICNALE dataset. The experimental results indicate the superiority of our approach in relation to the existing baselines.