基於有向圖與爭論導向摘要的網路辯論之爭論元素辨識(Identifying Argument Components in Online Debates through Directed Graph and Argument-oriented Summarization)

基於有向圖與爭論導向摘要的網路辯論之爭論元素辨識 Identifying Argument Components in Online Debates through Directed Graph and Argument-oriented Summarization 魏奇安 Chi-An Wei 國立成功大學資訊工程研究所 Institute of Computer Science and Information Engineering National Cheng Kung University kevinwei3@gmail.com 高宏宇 Hung-Yu Kao 國立成功大學資訊工程研究所 Institute of Computer Science and Information Engineering National Cheng Kung University hykao@mail.ncku.edu.tw 摘要 辨識觀點在「辯論語意挖掘」的領域中扮演著重要的角色。在任意文章中辨識出論點， 不僅可以用在被立場分類做為特徵來源增加辨識的準確性，同時也可以做為一篇文章表 明其支持或反對一個目標的理由。在這項研究中，已經有許多方法嘗試解決這一項任務， 較廣泛提起的是文字分類方法和自動化摘要的技巧。然而，將辨識觀點轉化成文字分類 的問題不僅著重在特徵的選擇和使用，同時也因為將文章中一個一個句子拆出視為獨立 的句子，所以也失去了文章整體的關係；自動化摘要的技巧雖然將文章中的所有句子一 起處理評估出其中一句適合做為文章代表的句子，然而現有的一些方法多使用詞袋模型 以及通用的自動化摘要方法，不僅沒有表示的特徵過於稀疏，同時亦沒有考量及引入辯 論文章的特性。 本研究主要在可以保留文章為一體的自動化摘要方法上進行深刻的觀察及調整，使其不 僅可以以辯論強度為導向摘要文章，並建構一有向圖來表示這些文章。在這些增強和調 整下，不僅讓原本無辯論導向的自動化摘要方法可以更好的排序出辯論的句子。最終， 本研究提出的方法可以在辯論文章中識別論點句子的任務上提升 8%的辨識準確率。 The 2019 Conference on Computational Linguistics and Speech Processing ROCLING 2019, pp. 378-387 ©The Association for Computational Linguistics and Chinese Language Processing 378 Abstract Identifying argument components has become an important area of research in argument mining. When argument components are identified, they can not only be used for stance classification but also can provide reasons for determining an article is supporting or opposing about a specific target. Previous research mainly used text classification and summarization techniques to solve this task. However, by transforming the task to a classification problem, not only rely heavily on choosing and using bag-of-words features, but also lose the article entity information due to extract the sentences out of the article and treat as an individual training instance. In the other hand, although summarization techniques handle on entire article and try to figure out which sentence can best represent the core concept of the article, in identifying argument components still heavily relies on bag-of-words feature representation and lack of argument-oriented features to concern about argument components characteristics. In our study, we dive down to the core of the summarization method, not only makes it based on argument strength to summarize articles and identify argument components, but also proposed a directed graph construction approach. Experiments show that our proposed method outperforms 8% better than those without argument-oriented methods. 關鍵詞：辯論語意挖掘、辨識論點、自動化摘要、有向圖 Keywords: Argument Mining, Argument Components, Summarization, Directed Graph. 1. Introduction In “Argument Components Extraction (ACE),” researchers try to get which sentence can be treated as arguing points, as known as “Argument Components (AC).” According to [1], which summarized tons of state-of-art knowledge about AM, summarized that how can a sentence become an argument component. The argument components in argumentative articles are usually formed by five types of sentences: “Claim” are the sentences that represent the statement being argued, “Data” are the facts or evidence used to prove the claim, “Warrant” are the sentences that make a connection between data and claim, “Backing” and “Rebuttal” are the sentences that support and against the warrant, respectively. ACE is what our study is targeting on, to find out which are the key sentences that make 379 people explain why they support or against something. Furthermore, the results can feedback to SC task, since we know the key sentences that lead an article to support or against. We are using the dataset released by [2], which collected posts under the domain “Abortion,” “Gay Rights,” “Obama” and “Marijuana” from an online debate forum. They manually labeled argument components in each post under each domain. To accomplish the goal of ACE, previous methods commonly use bag-of-words or feature-based approach to represent the sentence …

Content truncated to 500 words