A Study On The Automatic Scoring Method For Chinese-English Interpreting Questions In Terms Of The Key Information Of Answers

The automatic scoring for spoken English test has always been a research hotspot in the field of Computer-Assisted Language Learning.At present,most of the existing automatic scoring methods focus on reading questions,but rarely focus on Chinese-English interpreting questions.Human rating mainly focuses on the key information contained in the speech answer when scoring the Chinese-English interpreting.The key information is generally reflected by keywords.In this context,by simulating the human rating pattern,this paper focuses on designing an automatic scoring method for Chinese-English interpreting questions,in terms of the key information in the answers.During my research,I’ve studied and solved a series of problems in the automatic scoring method for Chinese-English interpreting questions,including speech signal preprocessing of speech answers,identifying key information and extracting key information integrity features from the answers,pronunciation fluency calculation and fluency feature extraction of speech answers,method designing of automatic scoring based on the combination of key information integrity features and fluency features.Main research efforts are as follow:(1)Considering the difficulty of constructing an accurate continuous speech recognition system without a large number of marked voice data,I applied the Query-by-Example Spoken Term Detection method based on dynamic time regularization(DTW)algorithm to recognize the keywords in speech answers.The performance of SLN-DTW(Segmental local-normalized-DTW)keyword detection method has been proven by the comparative experiment on TIMIT corpus.And then I built the keyword recognition library of speech answer through WordNet,and applied the SLN-DTW keyword detection method to recognize the keywords in speech answers.The experiment result shows that the number of keywords detected by SLN-DTW keyword detection method can be used as the effective feature representation of the key information coverage.(2)In order to further obtain its confidence after detecting the keywords,I applied the speech recognition method based on convolution neural network(CNN)to recognize the keywords.By constructing a speech recognition model based on CNN and combining the mean parameter algorithm with speech parameters,the recognition experiment was carried out on the Spoken Arabic Digit dataset of UCI machine learning library.The recognition rate of experiment is better than other models on the same data set.The result of the experiment in corrected keywords dataset is better than the other two commonly used recognition models too,which proves the feasibility of adopting CNN speech recognition method.(3)The key information integrity features are calculated by using the result of the keywords detection and the result of the keywords recognition.The automatic scoring model is constructed by using regression analysis to analyze all the features combining the key information integrity features with the fluency feature extracted on the original speech answers.The correlation coefficent between the scores given by the machine and the scores from human rating is 0.729 in the experiment for testing the performance of the automatic scoring model,which proves that the extracted feature is valid for the machine score.It also proves that the automatic scoring method for Chinese-English interpreting questions for the in terms of the key information of the answers is effective.