摘要： 跨领域的文本分类，是指利用有标记领域的知识去帮助另一个概率分布不同的，未标记领域的知识进行分类的问题。从多视图学习的视角提出一个新的跨领域文本分类的方法(MTV算法)。通过在核空间典型相关分析中引入与标记相关的信息，MTV算法可以得到一个判别性能更优的公共子空间。在多个情感类文本数据上的实验表明，MTV算法可以大大提升传统监督式学习算法面对领域迁移时的分类性能，并且在引入判别式的核空间典型相关分析后，进一步优化性能。

关键词: 跨领域学习, 迁移学习, 情感分类, 文本分类, 核空间典型相关分析

Abstract: Cross domain text classification aims at leveraging the knowledge in labeled source domains to predict the unlabeled data in another domain, where the distribution is different from the source domains. This paper proposes a new algorithm for cross domain classification in a multi-view learning perspective, named MulTi-View(MTV) transfer classification. By incorporating discriminative information related to labels, a more discriminative common low dimensional space is extracted. In MTV perspective, MTV algorithm provides a clear and neat way to transfer knowledge between domains. By applying KCCA, MTV also gives a closed form solution. Experiments show that MTV algorithm can significantly improve the performance of traditional supervised learning methods on extensive sentiment text datasets.

Key words: cross domain learning, transfer learning, sentiment classification, text classification, kernel canonical correlation analysis

中图分类号: 

• TP391