最小费用最大流跨领域情感分类框架

Minimum-cost and Maximum-flow Framework for Cross-domain Sentiment Classification

在对文本的情感极性进行分类的过程中,针对标注训练数据与待判断情感极性的数据来自不同领域,特征分布差异较大,导致传统监督学习方法分类准确率大幅下降的问题,提出基于最小费用最大流框架的跨领域情感分类方法.该方法将跨领域情感分类问题转化为最大化领域间传播信息量和最小化跨领域特征拟合损失量的双重最优化问题,并建立连接源领域和目标领域的费用容量网络,将信息量和损失量分别看作网络中的容量和费用权值,通过最小费用最大流框架解决上述双重最优化问题,调配训练数据特征权值.以更为直观的模型描述领域间的映射关系.实验发现,方法能有效提高跨领域情感分类的准确率.

There are some test data,and we need to determine their sentiment polarity,in case that the training data and the test data don＇t belong to the same domain,because of the huge difference of feature distribution,the performance of sentiment classification de- creases sharply when transferred from one domain to another by traditional supervised classification methods. A new cross-domain sen- timent classification method based on framework of Minimum-Cost and Maximum-Flow is proposed. We redefine the problem of cross-domain sentiment classification as a double optimization problem, the goal is maximizing the information transferred from source-domain to target-domain meanwhile minimum the loss of cross-domain feature fitting, then we build the cost-volume network to link the source-domain and the target-domain, and define the volume and cost with information and loss amount, so we can solve the double optimization problem with Minimum-Cost and Maximum-Flow framework as above-mentioned. Finally, we adjust the feature weights of train data and describe the mapping relationship between the domains more intuitively. The experiment results indicate that the proposed method can improve the performance of cross-domain sentiment classification effectively.