碩士論文(2021)

基於深度學習之心智圖自動產生方法與技術研發:以數位閱讀與寫作能力培養之應用為例

On deep learning-based method and technology for automatic mind map generation: development of digital reading and writing ability as an example

關鍵字 Keywords

摘要

心智圖是一種用於整理概念的圖像,被廣泛用於語文教學領域,並有多項研究證明,將心智圖應用於教學,能提升學生不同面向的語文能力。但是,將心智圖應用於教學,須為大量的教學內容準備心智圖,造成教師需耗費大量時間。

鑒於心智圖教學之需求,本研究運用近年熱門的深度學習技術,設計並開發「基於深度學習之心智圖自動產生方法」,包括「命名實體辨識」、「實體關係擷取」、「重點實體擷取」與「結構視覺化」等四個步驟。本研究設計三個深度學習模型,用來解決「命名實體辨識」、「實體關係擷取」與「重點實體擷取」步驟遇到的分類問題。最後由「結構視覺化」演算法,整合深度學習模型之預測結果資料,將資料視覺化,產生數位心智圖。

為評估「基於深度學習之心智圖自動產生方法」的有效性,本研究設計「深度學習模型評估流程」與「基於深度學習之心智圖產生方法評估流程」。「深度學習模型評估流程」評估三個不同的BERT預訓練模型,在命名實體辨識、實體關係擷取與重點實體擷取任務中的表現,並將表現最佳之模型應用於「基於深度學習之心智圖產生方法評估流程」。「基於深度學習之心智圖產生方法評估流程」分為兩個評估情境「人工標註之心智圖」與「自動產生之心智圖」,在命名實體辨識、實體關係擷取與重點實體擷取步驟,分別透過人工標註與深度學習模型產生資料,將此資料經過結構視覺化步驟,繪製成心智圖。最後,將人工標註與深度學習模型產生之心智圖混合,透過人工評分的方法進行實驗。實驗結果顯示,「基於深度學習之心智圖自動產生方法」在人工標註下能夠產生足夠品質的心智圖,證明此方法之有效性。透過深度學習模型產生之心智圖品質則較不穩定,少部分可以達到與人工標註相同的品質,但大部分自動產生心智圖之品質則較為不足。

為驗證數位心智圖應用於教學的有效性,本研究設計「以心智圖應用為基之數位閱讀與寫作能力培養模式」,並以此模式開發「數位讀寫學習平台」。經實驗驗證,可以提升學生閱讀時的專注度與作文的豐富度。

Abstract

Mind maps are often used in the field of language teaching in schools and are effective in improving students' reading and writing skills. However, the preparation of such teaching materials is quite a huge burden to teachers.

To reduce such burden on teachers, this study uses deep learning techniques to design the "Deep Learning-based Mind Map Generation Method" for automatic mind map generation. The method includes four steps: "Named Entity Recognition", "Entity Relationship Extraction", "Key Entity Extraction" and "Structural Visualization". In this study, three deep learning models are proposed to solve the classification problems in the steps of "Named Entity Recognition", "Entity Relationship Extraction" and "Key Entity Extraction". Finally, the "Structural Visualization" algorithm organizes the prediction result data of deep learning models, visualizes the data, and generates digital mind maps.

To evaluate the effectiveness of the "Deep Learning-based Mind Map Generation Method", the experimental design of this study includes two scenarios: mind maps generated by manually labelled data, and mind maps generated automatically. After manual evaluation, the evaluation result of mind maps generated by manually labelled data shows that the method could generate a mind map with sufficient quality. The evaluation result of mind maps generated automatically shows that a small number of the mind maps have sufficient quality, but most of the mind maps have poor quality.