Speech-Based Automatic Assessment of Question Making Skill in L2 Language

摘要

In this paper, we present a spoken educational system to automatically assess Arabic-native children's skill in forming English questions for different presented prompts. These prompts consist of images with a sentence that includes the answer to the required question. The answer key is colored to indicate what to ask. The main methodology of the proposed system is to record the spoken response of the child and pass it through state-of-the-art ASR to convert it into text. The output transcription is passed through three pipelined subsystems; Wh-question word checker, English grammar checker, which returns the number of grammar errors in the given question, and machine learning based grammar/language checker. The student response is accepted only if it is accepted by the three subsystems. The system was trained on 650 recorded responses made by 60 students (5th to 8th grades) as response to 75 different prompts. The number of grammar errors produced by the English grammar checker, best cosine similarity, best edit distance and best Jaccard distance between student response and the corresponding reference possible responses, are used to train KNN and SVM models with different parameters. The best precision, recall, f-measure and accuracy were achieved by SVM with linear kernel and degree of 2, 91%, 88%, 89% and 89%, respectively.