基於深層類神經網路及表示學習技術之文件可讀性分類

Classification of Text Readability Based on Deep Neural Network and Representation Learning Techniques

隨著網路快速發展，帶來的是資訊流通的便利性。然而，如此爆炸性成長的網路資訊卻令人忽略最重要的根本：閱讀的材料是否能夠被理解。當閱讀材料愈能夠被讀者所理解，就愈能夠產生好的學習效果。而合理估測文本能夠被理解的程度是可讀性公式的研究目標，它主要透過分析文件上的資訊來轉化成可讀性特徵，再利用這些可讀性特徵來訓練出可讀性模型，以便能自動化預測文件的可讀性。而近年來隨著深層類神經網路的崛起，不論應用在語音辨識、圖像處理及自然語言都有明顯效能上的提升。因此，本論文將研究這個技術應用於文本可讀性這個有趣且富實際應用的議題，提出一個基於深層類神經網路技術及詞向量表示法來建構出一個能夠分析跨領域文件的可讀性模型，以符合文件內容多元主題的特性。使可讀性模型能夠更精確分析文件的可讀性之外，也能夠兼具領域一般化的能力。

The development of the internet has facilitated the flow of information. However, this explosive growth of information has led to fundamental importance being overlooked: Reading material can be understood. Research on readability formulas aims to predict, to a reasonable extent, the degree to which a text can be understood. It does so mainly by analyzing and translating the information within a text into readability features, which are used to train a readability model, in order to automatically predict the readability of a given text. In recent years, the development of deep neural networks, applied to speech recognition, image processing and natural language processing has improved significantly on the performance. Therefore, this paper proposes a readability model built with deep neural network and word vector representation, and which is capable of analyzing cross-domain texts, in accordance with the diverse topics of text contents. The authors aim to make the readability model capable of analyzing text readability with more accurate, as well as possess domain generalization capacity.