基於深度學習之英語教材適性化研究

Abstract

閱讀能力是語言能力的重要組成部分，然而學生的閱讀水平、理解能力以及興趣愛好往往不盡相同。隨著網路的發展，網路中的閱讀材料也越來越多，為每個學生都找到適合的課外閱讀教材對教師來說是一項繁重的工作。這項任務可以通過評估教材的可讀性來解決，早期的可讀性研究多採用簡單的語法及詞彙特徵去設計模型，以此來預測教材的可讀性。近年來，深度學習技術在影像與語音辨識等應用獲得很大的進步與關注，如何將其優勢應用在數位學習方面，是一個值得研究的方向。本研究提出一種基於深度神經網路中長短期記憶(Long Short-Term Memory, LSTM)模型的文本可讀性預測的方法，針對過去可讀性研究的特性及困難點，選取文本的表面特徵、語法結構樹特徵和單詞特徵，並加入深度學習的詞向量特徵，以此解決傳統研究中沒有考慮到的詞彙間關係的問題。本研究以多個資料集進行實驗，透過與過去方法的對比，驗證了本研究所提出的文本可讀性預測方法的可行性。除此之外，本研究應用所提出的文本可讀性預測方法，開發一套具有教材分級功能和時事測驗題自動產生功能的數位學習系統，以此提供適合使用者學習的內容，減輕英語教師的工作負擔。

Reading ability is an important part of language ability. However, students' reading level, comprehension and hobbies are often different in reality. With the development of the Internet, there are more and more reading materials on the Internet. It is an arduous job for teachers to find suitable reading materials for each student from huge online reading materials. Actually, it can be solved by evaluating the readability of the textbook. Early readability studies used simple grammar and lexical features to design the model to predict the readability of the textbook. In recent years, deep learning technology has made great progress and attention in applications such as image and speech recognition. Thus, how to apply the advantages of deep learning technology to e-learning is a worthwhile research direction. This study proposes a text readability prediction method based on the Long Short-Term Memory (LSTM) model in deep neural networks. According to characteristics and difficulties of past readability studies, we selected surface features of the text, syntactic parse tree features, word features and word vector features to solve the problem of lexical relation is not considered in traditional research. In this study, experiments used multiple data sets, and by comparing with the previous methods, the feasibility of the proposed text readability prediction method was verified. What’s more, in order to provide more suitable content for users and reduce the workload of English teachers. An e-learning system with the function of textbook grading and automatic generation of current test questions is developed, which applies text readability prediction method proposed in this study.