新聞立場辨識(News Stance Detection)幫助人們釐清新聞文章所代表的立場，進一步從不同角度去了解各種議題。在新聞立場辨識的任務中，我們必須根據給定議題及新聞文章，去判斷此篇新聞立場是中立、偏向贊成或對立。此項任務與自然語言推理(Natural Language Inference, NLI)任務類似，目標在給定兩個句子，判斷兩者之間是否無關或存在蘊涵、矛盾關係。新聞文章是由多個句子組成的篇章（Discourse），因此透過分析文章中的句子彼此的關係（稱之為篇章分析Discourse Analysis），也可能有助於新聞立場判定。 然而，在篇章分析以及自然語言推理的相關研究中，大部分的模型是在文章中「句子間」的關係（篇章關係）已知的情況下訓練；新聞立場辨識則是要找尋「文章」與「議題」的關係，因此每篇文章本身的篇章關係，以及文章中的每個句子與議題間的關係都是未知的。同時，在訓練資料中，新聞內容與議題的立場不同的資料比例很少，讓訓練模型更加不易，也是本項任務的困難之處。 在本篇論文中，我們參考了Parikh等人針對NLI任務設計的Decomposable Attention Model以及Durmus等人以階層的方式處理篇章資料的作法，提出Hierarchical Decomposable Attention Model來解決新聞立場辨識任務。實驗結果顯示，該架構確實可以達到較好的效能。針對資料不平衡的問題，我們加入新聞內容與議題立場不相關的資料，並透過實驗驗證我們的作法有助於提升模型在不相關資料的辨識能力。 ;In News Stance Detection task, we have to judge whether the stance of news is neutral, partial approval or opposition to a given query. This task is similar to the Natural Language Inference (NLI) task which accepts two sentences as input and determine whether the given two sentences are irrelevant, entailment or contradiction. A news article is a discourse composed of multiple sentences, so it may also help to determine the news stance by analyzing the relationship between the sentences in the article (called Discourse Analysis). In the related work for Discourse Analysis, most models are trained with the relationship between “sentences” (Discourse Relation) in the articles; News Stance Detection is to find relation between “query” and “article”, so the discourse relation in the articles, and the relationship between each sentence in the article and the query is unknown. At the same time, in our training data, few of news articles hold different stance toward the queries, such a data imbalance problem makes the training of models more difficult, which is also a challenge of this task. In this paper, we proposed a Hierarchical Decomposable Attention Model to solve News Stance Detection task which refer to the Decomposable Attention Model (Parikh et al. 2016) for NLI tasks and the hierarchical way to deal with discourse (Durmus et al. 2019). The experiment result showed that the performance of our architecture is better than other models. For the data imbalance problem, we added data labeled with unrelated and proved this way can improve the ability of the model to identify unrelated data.
