Research And Application Of Multi-Feature Intelligent Correction Model

With the development of computer technology,computer-aided language learning system(CALL)has becoming more and more mature.Many CALL systems have been used in teaching and examination to assist teachers in teaching and improve students' learning efficiency.At present,the oral English corrections of CALL systems are all for text-related questions such as reading questions.The automatic correction of open-type oral English questions has been a problem in the field CALL,it is a text-independent question type.At present,the automatic correction system for open oral English test questions has not been put into practical application.The existing open-type oral automatic scoring system has problems such as insufficient relevance to expert ratings and scoring dimensions.In order to solve these problems,this paper proposed a multi-feature intelligent scoring model,based on the current mature English composition correction system and speech recognition engine,and automatically scores for open oral questions.We scored the multi-features of the candidates' oral recordings.There were five characteristics:pronunciation,fluency,vocabulary,grammar,and semantics.The deep neural network model was constructed to model the features to obtain the final score.Based on previous studies,we used the deep neural network training model instead of the linear regression method,improved the correlation between model scores and expert scores,and improved the method of semantic scoring using LDA model instead of word frequency witch has no expert labeling keywords,realized the real automated correction.In addition,this paper also introduces text cleaning after speech recognition and deep learning based speech suppression technology in the scoring model which improves the performance of speech recognition and scoring model.This paper uses the audio data collected by the 480 English examinations of the Beijing University of Posts and Telecommunications as a data set to train the model,and uses the expert scoring mapping model.Our model finally achieved the 89.3%of expert scoring performance,reaching a practical level.