Automatic Proficiency Evaluation of Spoken English by Japanese Learners for Dialogue-Based Language Learning System Based on Deep Learning
Year: 2021
Type: dissertation
Abstract: The rapid development of globalization leads to an increase in the demand for the second language (L2) learners to study foreign languages.English is the most universal language for speech communication along the world.Traditional classroom English teaching indoors contributes to timely interaction between the teachers and the students.However, such a language teaching method tends to be costly and time-consuming.With the widely used electrical devices and the Internet, it is possible to provide a variety of materials for the needs of English self-study.This method is still hard for them to check the language proficiency as there is no teacher to give feedback directly.In most cases, this boring way of learning alone can make learners lose interest in learning a new language.Due to the increasing computing power and modern information technology in the past two decades, computer-assisted language learning (CALL) system has attracted much attention in the field of language teaching, as an efficient tool for the L2 learners.Speech-based computer-assisted language learning (CALL) systems should recognize the utterances of the learner with high accuracy and evaluate the language proficiency of the specific speaker with appropriate methods.In this thesis, we discuss the automatic proficiency assessment of the second language (L2) for non-native speakers.For the language proficiency evaluation, pronunciation evaluation and grammatical error detection (GED) are both important components in CALL systems.Regarding the automatic assessment on language pronunciation, there are many existing works iii "optimum" in some senses, as the rules were heuristically established, the generated sentences do not necessarily reflect the comprehensive situation of English spoken by L2 learners.For the GED task in our CALL system, we describe methods for improving the recognition accuracy of speech with grammatical errors.We introduce the process of sentence generation with grammatical errors by adopting the state-of-the-art neural machine translation (NMT) for training a new language model.Then we compare five language models with different training data: 1) baseline LM (language model), 2) rule-based LM, 3) NMT-based LM, 4) rule & NMT-based LM, 5) gold LM (closed-condition).As a result, by comparing the five different language models, the proposed method gives more accuracy on speech recognition rate, which indicates that if the language model contains the human transcribed text of the recorded speech (sentences with realistic errors), the ASR could get a high recognition rate for non-native speech.Therefore, the NMT models generated related errors, which would promote the performance of ASR in the spoken GED module of our CALL system.To a certain extent, it is worth to use the NMT for generating sentences with errors, which will increase the efficiency of the entire CALL system and save labor costs.
Author: FU, JIANG
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7146152255