應用階層可解構式注意力模型於新聞立場辨識任務

A Hierarchical Decomposable Attention Model for News Stance Detection

"新聞立場辨識任務的目的為判斷一篇新聞對於某個議題的立場是中立、贊成或反對。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。本論文以新聞立場檢索競賽提供的資料作為參考，但其大部分的新聞文章都屬於支持特定議題的立場，造成資料在不同類別的分佈不平衡。本篇論文提出Hierarchical Decomposable Attention Model來解決新聞立場辨識任務，我們以句子為單位分割新聞文章，並基於Decomposable Attention的原理找出文章中的每個句子與特定議題的關係。針對資料不平衡的問題，我們建立反義的議題，並手動標記新聞對反義議題的立場，以改善模型效能。實驗結果顯示，我們提出的模型效能優於其他模型。"

The goal of News Stance Detection task is to detect whether the stance of a news article is neutral, approval or opposition with respect to a given query. The task is similar to Natural Language Inference (NLI) task, which aims to determine if one given statement (a premise) semantically entails another given statement (a hypothesis). Since most news articles hold neutral stances with respect to the given query, the training data is often unbalanced. In this paper, we proposed a Hierarchical Model based on the Decomposable Attention Model for NLI tasks to compare individual sentences with the given query and jointly predict the stance of the complete article. For the data imbalance problem, we heuristically create opposite queries and label supporting news articles from unrelated ones of the original query to identify unrelated news articles. The experiment result showed that the performance of our architecture is better than other models.