Deep Learning for Dialogic Spoken Language Assessment and Learning

世界上有超过10亿的英语学习者对评估和提高口语水平的自动化方法的需求越来越大。目前，所有的口语评估系统都假定是独白(单个说话人)或非常受限的对话语音。口语评估和学习的下一个基本挑战之一是在非常自然的互动场景中自动评估考生。对话式评估是通过两三个人的对话对候选人进行评估，允许他们通过最自然的互动形式来展示他们的口语能力。这通常用于评估，包括雅思等高风险考试，以及在学习/教学中，并允许测量和提高语言和交流技能。自动化对话评估带来了许多重大挑战，包括没有明确的评估方法，与标准语音识别任务相比，带注释的训练数据有限，具有广泛的口音和说话风格，以及需要在数据中自动识别说话人。这项工作旨在利用深度学习的最新进展，并将其应用于对话评估和学习。

There is a growing demand for automatic approaches for assessing and improving spoken language proficiency with over one billion learners of English world-wide. Currently all spoken language assessment systems assume monologic (single speaker) speech or very constrained dialogic speech. One of the next fundamental challenges in spoken language assessment and learning is automatically assessing candidates in very natural interactive scenarios. Dialogic assessment, where the candidate is assessed through a 2 or 3 person conversation, allows them to demonstrate their speaking ability through the most natural form of interaction. This is commonly used for assessment, including for high stakes exams such as IELTS, and in learning/teaching, and allows both linguistic and communicative skills to be measured and improved. Automating dialogic assessment presents a number of significant challenges including no clear methodology for approaching the assessment, limited annotated training data compared with standard speech recognition tasks, with a wide range of accents and speaking styles, and the need to identify speakers automatically in the data. This work aims to take the latest advances in deep learning and apply them to dialogic assessment and learning.