Ji yu yin bian liang jian mo de zi ran yu yan li jie

Ph.D.Natural Language Understanding (NLU) is focusing on enabling machine to understand and process unstructured human language. NLU is one of the core parts of natural language processing with a wide range of applications, such as text classification, sentiment analysis, question answering, etc. However, most learning based NLU methods either rely on pre-defined, hand-crafted features, or high-quality human annotated data for training.With the risen of social media and arrival of information explosion era, the large volume and short nature of social media text data bring a lot of challenges to conversational NLU methods, making the feature engineering and data annotation very labor-intensive and domain-specific. In this thesis, we propose to model latent variables (i.e., topics and discourse) on social media text (e.g., microblog, online forum) in an unsupervised way, based on which, we design our NLU models and demonstrate their superior performance on two social media NLU tasks: short text classification and argumentation mining.First, we propose an unsupervised framework for jointly modeling topic content and discourse behavior in microblog conversations for understanding the semantics and interaction of social media messages. Concretely, we propose a neural model to discover word clusters indicating what a conversation concerns (i.e., topics) and those reflecting how participants voice their opinions (i.e., discourse). Extensive experiments show that our model can yield both coherent topics and meaningful discourse behavior. Our model can be easily extended with other neural networks. Further study shows that our topic and discourse representations can benefit the classification of microblog messages, especially when they are jointly trained with the classifier.Second, we focus on the short text classification, which is one of the most fundamental techniques in social media text understanding. To address data sparsity issue in social media short text, we propose topic memory networks for short text classification with a novel topic memory mechanism to encode latent topic representations indicative of class labels. Different from most prior work that focuses on extending features with external knowledge or pre-trained topics, our model jointly explores topic inference and text classification with memory networks in an end-to-end manner. Experimental results on four benchmark datasets show that our model outperforms state-ofthe- art models on short text classification, meanwhile generates coherent topics.Third, we focus on the online argumentation, which is a growing and challenging field in social media text understanding. We present a novel study that automatically analyzes the key factors of persuasiveness in argumentation process, beyond simply predicting who will win the debate. Specifically, we propose a novel neural model which is able to dynamically track the changes of latent topics and discourse in argumentative conversations, allowing the investigation of their roles in influencing the outcomes of persuasion. Extensive experiments have been conducted on argumentative conversations on both online forum and supreme court. The results show that our model outperforms state-of-the-art models in identifying persuasive arguments. We further analyze the effects of topics and discourse on persuasiveness, and draw some findings from our empirical results, which will help people better engage in future persuasive conversations.自然語言理解(NLU)關注于使機器能理解和處理非結構化的人類語言。NLU是自然語言處理核心部分之一，有廣泛的應用，比如文本分類，情感分析，問題解答等。然而，大多數基於學習的NLU方法或者依賴於預定義的人為製作的特徵，或者訓練所需的高質量的人工標註數據。隨著社交媒體和信息爆炸時代的到來，大量簡短的社交媒體文本給對話式NLU方法帶來了大量的挑戰，使得特徵工程和數據標註極為需要人力和領域特定。本論文中，我們提出了社交媒體文本（比如微博和在論壇）上的無監督模型隱變量（即主題和論述），基於此，我們設計了新的NLU模型並且展示了它們在兩個社交媒體自然語言理解任務上優越的性能：短文本分類和辯論挖掘。首先，我們提出了微博對話中聯合建模主題內容和論述行為的無監督模型以理解社交媒體信息的意義和交互。具體地，我們提出了一個神經模型來發現表徵一組對話設計的內容（即主題）和反應參與者如何表達他們觀點（即論述）的詞聚類。大量實驗表明我們的模型可以生成連貫的主題和有意義的論述行為。我們的模型可以比較容易地跟其它的神經網絡結合擴展。進一步的研究表明我們的主題和論述表徵能夠有助於微博信息的分類，特別是當它們與分類器聯合訓練的時候。其次，我們集中在短文本分類這個最基本的社交媒體文本理解的技術。為了解決社交媒體短文本的數據稀疏問題，我們提出了為短文本分類的主題記憶網絡，即用一個新的主題記憶機制來編碼表示類標註的隱含主題表徵。不同於大多數以往關注在用外部知識擴展特徵或者預訓練的主題，我們的模型用端到端的方式基於記憶網絡聯合探索了主題推斷和文本分類。在四個基準數據集上的實驗結果表明我們的模型在短文本分類上優於最新的模型，且生成連貫的主題。再其次，我們關注在辯論這個在社交媒體文本理解中不斷增長的且有挑戰性的領域。我們呈現了一個新的自動分析辯論階段中說服的關鍵因素的研究，而不僅僅是簡單地預測誰講在辯論中獲勝。特別地，我們提出了一個新的神經模型，來自動地動態追蹤隱含主題和論述在辯論對話中的變化，並允許探索它們在影響說服結果中的角色。大量在辯論對話場景中的實驗被執行，包括在論壇和最高法院。結果表明我們的模型在識別說服性的辯論上優於最新的模型。我們進一步分析了主題和論述在說服中的影響，並發現它們都很有用。此外，我們總結了基於實驗結果的一些發現，可以幫助人們更好地參與未來的說服性對話。Zeng, Jichuan = 曾紀川.Thesis Ph.D. Chinese University of Hong Kong 2019.Includes bibliographical references (leaves 112-141).Abstracts also in Chinese.Title from PDF title page (viewed on 21 May, 2021)