任務導向對話系統輔助語言學習以英語教學系統為例

本論文主軸為改進任務導向(Task-Oriented)對話系統的評分機制，其目標為提升對話系統的使用情境，包括自然語言處理、拆解意圖、對話狀態追蹤等模組，使對話系統能更精準的決策最佳的回應句子，在評分機制上能更趨近於教師的人工評分。對話系統是以口語的方式與機器溝通，分為任務導向及非任務導向兩種，而本論文針對「任務導向」進行探討。本研究以任務導向型的對話系統作為電腦輔助教學的媒介，輔助教師讓學習者增強他們的英語口說能力，模擬具有目標導向特性的對話情境，學習者反覆的練習以提升口說能力及語言流暢度來建立自信心。基於樣板式模型建置任務導向對話系統來輔助語言學習，提供學習者以遊戲的方式進行，教材由專業的英語教師建構涵蓋日常生活用語的對話劇本，此方向之特點為將語言模型導入對話管理者模組，使系統能依據學習者輸入的語句給予不同的回應及對話評分，同時提出三種不同的模型來預測分數，並驗證是否能校準於教師所產生的人工評分。研究對象為大學部應用外國語文學系的學生，採用Willis提出的三段式教學步驟，於教學實驗前讓學習者熟悉系統以及任務描述的事項，同時教師在此階段進行任務解說，實驗過程中由學習者自行發揮口說能力，此階段教師則扮演監督者的角色而不干預測驗，待任務結束後由系統根據對答狀況給予評分，讓教師日後能針對不同學習者的狀況給予適當的教材，達到語言加強階段的效益。本研究提出三種不同的評分機制，以均方根誤差及平均絕對誤差來衡量，平均絕對誤差最少的為系統評分，換言之，系統評分最趨近於人工評分，但均方根誤差最少的為機器學習預測評分。本研究貢獻為建置任務導向對話系統輔助語言學習，擺脫受時間及地點的限制，提出一個系統框架，依照此框架能發展其他語言學習的系統。

關鍵字

任務導向教學法；任務導向對話系統；電腦輔助教學；數位學習

並列摘要

This study improves the scoring mechanism of the Task-Oriented dialogue system, the goal is to improve the context of the dialogue system, including natural language processing, natural language understanding, and dialog state tracker module. Make the dialogue system more accurate find best response sentences, experiment with a mechanism similar to the teacher scoring mechanism. dialogue system can communicate with the machine and can be divided into Task-Oriented and Non-Task-Oriented, this study explores task orientation. Previous research has shown that Computer Assisted Instruction is effective for language learning, can assist teachers help learners enhance English speaking ability, simulate conversation situation, after the learners continuous to practice, improve conversational fluency and confidence. This study developed a system of Task-Oriented dialogue assisted language learning, Rule-based model design, provides learners with a game-based learning experience. Dialogue tree is designed by a professional English teacher, contain food, transportation, etc., scripts. Learners can easily talk to the system, because this study, imported a language model into dialogue management module. When the learner say different sentences will get different responses and scoring. This study proposes three different ways to predict scores, and evaluation is similar to teacher scoring. The research target is a college student in Chung Yuan Christian University, Department of Applied Linguistics and Language Studies, using the three-stage teaching steps proposed by Willis in 1996. Make the learner familiar with the system and task description before the teaching experiment, the teacher is the observer during the experiment, after the task is over, the learner will get a rating from the system and the teacher. This study proposes three different ways to scoring mechanism, evaluation method includes Root Mean Squared Error and Mean Absolute Error. minimum Mean Absolute Error is system scoring, but minimum Root Mean Square Error is the machine learning prediction scores. The contribution of this research is to developed a Task-Oriented dialogue assisted language learning system, it is not limited by time and place, according to this framework, can develop systems for other language learning.