使用輔助向量的雙邊特徵分群以改善中文新聞的立場偵測分類

Abstract

為了紓解媒體偏頗以及閱聽者選擇性偏好的現象，本篇論文專注於發展一智慧程式，用以分辨中文爭議性議題新聞之立場。我們提出一個簡單且有效率的方法，能夠考量無標記新聞資料庫的資訊、以及訓練資料之資訊，以合併相似的特徵。在我們提出的方法中，特徵會先根據初始訓練過程被分為兩邊，接著使用word2vec工具為每一個特徵產生輔助向量，最後使用高速的社群偵測演算法將意義上相近的特徵合併。實驗結果顯示，在大多數的情況下，我們提出的解決方案比直接使用原始特徵、以及使用常見的降維演算法還要好。

In order to relieve media bias problem and selective preference problem, we aim at developing an intelligent system to classify the stance of Chinese news article on several controversial topics. We proposed a simple and efficient approach which can incorporate the information of unlabeled news corpus and the information of training data to merge similar features. In our approach, features were divided into two sides according to initial training process, and word2vec tool was utilized to produce auxiliary vectors for each feature. Finally, fast community detection algorithm was applied for clustering similar features. Experimental results show that our approach outperforms raw features and common dimensionality reduction techniques in most cases.