Cross-Collection Emotion Tagging For Online News
Published 2016 · Computer Science
With the rapid development of Internet and social media, online news has become an important type of information that attracts millions of readers to express their emotions. Therefore, it is of great significance to build an emotion classifier for online news. However, it largely relies on the collection with sufficient labeled news to build an emotion classifier and the manually labeling work can be quite labor intensive. Moreover, different collections may have different domains such as politics or entertainment. Even in the same domain, different collections require different classifiers, since they have different emotion labels and feature distributions. In this paper, we focus on the task of cross-collection emotion tagging for online news. This task can be formulated as a transfer learning problem which utilizes a source collection with abundant labeled data and a target collection with limited labeled data within the same domain. We proposed a novel method to transfer knowledge from source collection to help build an emotion classifier for target collection. Experimental results on four real datasets show that our method outperforms competitive baselines.