||Construction of an Automatic Scoring Model for English Interpretations by using a Recurrent Neural Network
|| Recurrent neural network; Posterior probability; English interpretation; Automatic scoring; Deep neural network
||Automatic scoring of English interpretations is a technology that evaluates content accuracy and pronunciation. The technology is of great practical significance for English interpreting tests. This paper uses DNN-HMM speech recognition to evaluate the interpretations of English translators. Post-probabilistic features of phonemes to the speech vector are decoded in the sub-recognition framework. The correlation between posterior probability and manual scoring is analyzed. The calculation of a posterior probability feature depends highly on recognition performance. To improve recognition performance, a Recurrent Neural Network (RNN) is used to construct speech recognition models. Through comparison experiments, results show that the correlation between the RNN model’s scoring method and manual scoring is 0.781. The recognition accuracy of the RNN model is 0.923 on the training set and 0.915 on the test set. The posterior probability error of the RNN model is 0.028 for the training set and 0.033 for the test set. Therefore, the RNN structure has more reasonable and scientific scoring performance in automatic English interpretation.