何子岳

He, Tzu-Yue

碩士論文(2022)

基於深度學習之文章摘要提取技術研發：以階層式文章摘要能力培養之應用為例

Development of Deep-learning-based Technology for Article Abstract Extraction：Application of Layered Article Abstraction Ability Developing as an Example

關鍵字 Keywords

人工智慧、機器學習、深度學習、階層式文章摘要、數位閱讀

Artificial intelligence、Machine learning、Deep learning、Layered Article Abstraction、Digital Reading

摘要

隨著數位時代來臨，數位閱讀素養已成為21世紀人類重要的能力之一。許多研究證實，從大量資料中提取重點並加以統整的能力是數位閱讀素養的關鍵，文章摘要提取能力的培養也因此受到相當的重視。

我國教育部認為摘要歷程學習方法是比較適合國小生的方法，比起傳統的摘要教學，將摘要歷程式之階層式文章摘要提取方法應用到教學上，能更有效的提升國小學生的文章摘要提取能力。針對階層式文章摘要提取能力培養之數位學習，教師必須事先準備階層式文章摘要參考內容，除了要花費大量時間且無法因應非預設文章之學習需求外，也很難提供即時評量與輔導。

本研究運用深度學習相關技術，將BERT預訓練模型進行Fine-Tune，再結合本研究設計之演算法，設計並開發「基於深度學習之階層式文章摘要提取技術」，包含「階層式文章摘要分類」、「相似句子分群」、「句子重要度辨識」與「階層式文章摘要提取」等四個步驟。實驗顯示，此技術雖然較不穩定，但有能力產生合乎品質需求之階層式文章摘要。

本研究也設計一個能提升「數位化文章摘要提取能力」之「數位閱讀能力培養模式」，並依此模式開發具自動化文章摘要提取技術之「數位閱讀寫作平台」，以驗證本技術於「文章摘要提取能力培養」與「數位閱讀能力培養」之應用性與有效性。實驗證實，自動化文章摘要提取技術可以實現學習平台即時評量與即時輔導之功能，且「數位閱讀能力培養模式」能提升學生文章摘要提取能力和閱讀理解能力。

Abstract

Due to the advent of the digital age, digital reading literacy is gradually gaining attention. More and more international assessments related to digital reading literacy have been launched. Digital reading literacy has become one of the indispensable skills in the 21st century. As a result, the demand for digital learning has increased dramatically.

In view of the demand for hierarchical article summaries. In this study, the BERT pre-training model is fine-tuned, and the algorithm designed in this study is combined to develop a "deep learning-based method for extracting layered article abstraction". This method can help teachers to generate digital teaching materials and reduce the burden of digital teaching materials production. The method consists of four sub-steps: "layered article abstraction classification", "similar sentence grouping", "sentence importance identification", and "layered article abstraction extraction".

In order to evaluate the correctness of the "deep learning-based layered article abstraction extraction method". In this study, we also designed a method evaluation process and conducted experiments for each of the four sub-steps. The results show that this technique is capable of generating layered article abstraction of sufficient quality, although it is not yet stable. It is also proved that the research direction of this method is correct and effective.