Kernel-Based Feature Fusion for Sensitive Information Filtering

摘要

Sensitive information filtering is the key technique to help people detect baneful information from the internet and insulate them for latter decisions.In this paper, based on the review about the characteristics and the hardness of sensitive information filtering task, we propose the idea of exploiting the combinational semantics of sensitive words to reinforce the filtering effect. Firstly, a new kernel, named as geometric-mean-ANOVA kernel, is introduced to generate features with specific combination degree. Further, multiple kernels across different combination degrees are fused together by composite kernel to produce the full feature space for the sensitive information filtering task. The evaluation in the real environment shows our kernel-based feature fusion methods exhibit superiority to those methods only using single kernel on both recall and precision.