论文部分内容阅读
为了改变计算机必须依靠文本才能进行评分的情况,该文在国内首次开展了面向大规模英语口语机考中的复述题型自动评分技术研究,并证明了其技术可行性。首先基于连续语音识别、自然语言理解等技术搭建了复述题自动评分技术流程,并针对复述题无需按原文复述、考试现场录音质量低等难点,通过借助朗读题录音的声学模型自适应处理、基于复述原文的通用语言模型裁剪、基于识别输出词图的机器评分特征提取等一系列的改进工作的开展,最终完成的自动评分系统在339份中国科学技术大学期末考试现场采集的复述题数据集上达到了专家精细评分84%的性能,超过了教师批量阅卷时的性能。实验结果表明该系统在英语复述测试中,可以辅助教师进行更科学客观的评分。
In order to change the situation that the computer must rely on the text to score, this paper firstly studies the auto-scoring technique for the repetition question in large-scale oral English test and proves its technical feasibility. Firstly, based on continuous speech recognition and natural language comprehension, the process of auto-scoring of compound words is established. In the light of the difficulty of repetition of original words and the poor recording quality of the test site, A series of improvement work such as machine language feature extraction based on recognition of output words and a series of improvement work were carried out. The final automatic scoring system was applied to 339 sets of compound data collected on the final exam site of University of Science and Technology of China Achieved 84% of the experts fine score performance, exceeding the performance of teachers when the volume marking. The experimental results show that the system can assist teachers to conduct more scientific and objective scoring in the English repetition test.