意見持有者辨識之研究

A Study on Identification of Opinion Holders

意見持有者辨識是從意見句中擷取出表述意見的人或組織，本研究將意見持有者辨識分為作者意見辨識及意見持有者標記兩部分，作者意見辨識使用支援向量機處理，意見持有者標記使用條件隨機域處理。本研究提出的方法應用在NTCIR7 MOAT繁體中文語料的效能達到F值0.734，是採取機器學習方法的參賽隊伍中效能最佳者，也相當接近目前最佳系統的效能。對於意見持有者辨識語料中標記歧異的情形，本研究加以分析，並提出使用此語料來訓練模型的方法。

The identification of opinion holders aims to extract entities that express opinions in opinion sentences. In this paper, the task of opinion holder identification is divided into two subtasks: the identification of author's opinions and the labeling of opinion holders. Support vector machine is adopted to identify author's opinions, and conditional random field model (CRF) is utilized to label opinion holders. The proposed method achieves an F-score 0.734 in NTCIR7 MOAT task at traditional Chinese side. The proposed method achieves the best performance among participants who adopted machine learning methods, and also this performance was close to the best performance in this task. In addition, the ambiguous markings of opinion holders are analyzed, and the best way to utilize the training instances with ambiguous markings is proposed.