一套基於詞排名的抽象式文件摘要模型訓練法

WordRank: A Word Ranking based Training Strategy for Abstractive Document Summarization

文件摘要一直是個經典且重要的研究議題，旨在將給定的一篇文章濃縮成簡潔精鍊的小段落。關鍵字(Keyword)在文章內扮演著承上啟下的重要角色，它們通常乘載著文章的主題、重點與核心概念。於是，過去許多研究提出基於關鍵字的文件摘要模型。然而，這些模型通常由一個萃取關鍵字模型，以及一個將關鍵字做為導引的摘要生成模型。這樣的設計不僅增加流程的複雜度，可能遭遇錯誤傳遞的問題，也將導致多餘的資源消耗。有鑑於此，本研究致力於提出一套基於詞排名的抽象式摘要模型訓練法，著眼於將關鍵字萃取與文件摘要兩者合而為一。為此，模型不僅可以自動地標示出文章內的關鍵字，亦可根據這些關鍵字產生文章的抽象式摘要。實驗結果顯示，使用基於詞排名的模型訓練法後，確實可以有效地提升摘要的成效，並且在關鍵字擷取的任務裡，也可獲得很好的成績。

Document summarization has always been a classic and important research topic, aiming to condense a given article into a few concise paragraphs. Keywords, which usually convey the theme, focus, and core concept of the content, play an essential role in the document. Therefore, many studies in the past have proposed keyword-based document summarization models. However, these models usually consist of a keyword extractor and a keyword-based summarizer. Such a design not only increases the complexity of the process but also may encounter an error propagation problem and will also lead to redundant resource consumption. In view of this, this research dedicates to proposing a word ranking based training strategy for abstractive document summarization, which mainly focuses on combining keyword extraction and document summarization. On top of the training strategy, the resulting model can automatically select keywords in the document and generate an abstractive summary based on these keywords. The experimental results show that using the proposed training strategy can indeed effectively improve the quality of the abstractive summarization and achieve good results in the keyword extraction task.