Automatic scoring of scene question-answer in English spoken test

摘要

This paper describes our studies on the automatic scoring of scene question and answer in English spoken test. The system includes three important parts: speech recognition, scoring features computation and scoring model. According to the assessment of English spoken test, scoring features should describe accurately speakers' answer, that's to say, they should cover different aspects of the student's answer including speech fluency, pronunciation quality, content relevance and grammar accuracy in order to get a proper machine score. Our system put a list of 16 initial features. Finally, features are mapped to ultimate machine score with SVM classification model. The performance measure that has been typically used is the correlation between the machine scores and the corresponding human scores [1]. Based on the same measure, the correlation coefficient is 0.72 between the machine scores and human scores while the coefficient is 0.69 within raters.