New similarity measures for automatic short answer scoring in spontaneous non-native speech

This paper focuses on similarity measures about automatic short answer scoring in spontaneous non-native speech. We propose two new corpus-based measures of text similarity, which are Manhattan similarity measure and keywords coverage rate based on edit distance. Manhattan similarity measure is converted by Manhattan distance. Furthermore, automatic speech recognition (ASR) system and students´ pronunciation may lead words´ variation including pronunciation and morphology. Improved keywords coverage rate combined edit distance can distinguish students´ grade and provide more objective scores. Experiment results show that these similarity measures are effective to our scoring system and performance of scoring system can achieve 96% of it by human raters.