表示法學習技術於節錄式語音文件摘要之研究

Abstract

在現今日常生活裡，大量的多媒體內容與日俱增促使自動語音文件摘要成為一項重要的研究議題。當中最為廣泛地被探究的是節錄式語音文件摘要(Extractive Spoken Document Summarization)：其目的是根據事先定義的摘要比例，從語音文件中選取一些重要的語句，用以代表原始語音文件的主旨或主題。另一方面，表示法學習(Representation Learning)是近期相當熱門的一個研究議題，多數的研究成果也證明了這項技術在許多自然語言處理(Natural Language Proceeding, NLP)的相關任務上，可以獲得優良的成效。有鑑於此，本論文主要探討使用詞表示法(Word Representations)及語句表示法(Sentence Representations)於節錄式語音文件摘要任務上。基於詞表示法及語句表示法，本論文提出三種新穎且有效的排序模型(Ranking Models)。除了文件中的文字資訊外，本論文更進一步地結合語音文件上的各式聲學特徵，如韻律特徵(Prosodic Features)等，以期望可獲得更好的摘要成效。本論文的語音文件摘要實驗語料是採用公視廣播新聞(MATBN)；實驗結果顯示，相較於其它現有的摘要方法，我們所發展的新穎式摘要方法能夠提供顯著的效能改善。The rapidly increasing availability of multimedia associated spoken documents on the Internet has prompted automatic spoken document summarization to be an important research subject. Thus far, the majority of existing work has focused on extractive spoken document summarization, which selects salient sentences from an original spoken document according to a target summarization ratio and concatenates them to form a summary concisely, in order to convey the most important theme of the document. On the other hand, there has been a surge of interest in developing representation learning techniques for a wide variety of natural language processing (NLP)-related tasks. However, to our knowledge, they are largely unexplored in the context of extractive spoken document summarization. With the above background, this thesis explores a novel use of both word and sentence representation techniques for extractive spoken document summarization. In addition, three variants of sentence ranking models built on top of such representation techniques are proposed. Furthermore, extra information cues like the prosodic features extracted from spoken documents, apart from the lexical features, are also employed for boosting the summarization performance. A series of experiments conducted on the MATBN broadcast news corpus indeed reveal the performance merits of our proposed summarization methods in relation to several state-of-the-art baselines.