Bi-perceptron 分類中文網頁新聞

Abstract

新闻因为其高频的访问属性,在中国大陆已经成为了各大互联网公司追逐的领 域。而新闻分类一直是新闻自动处理中的一项主要议题。有很多的监督式学习方法 可以应用在这一领域,其中,支持向量机在离散的特征空间上表现尤为出色。本文 中提出 bi-perceptron 的思路去解决基本的二分类问题,希望在某些方面达到甚至 超过支持向量机的效果。 Bi-perceptron 学习法是一种分而治之的思想,本文中提出这种思想,并实现了 一种基本的解决方法。我们将分类转化成数据划分,基本分类,分类合成三个步骤 并比较了不同划分及合成方法。另外,本文从中文网络新闻的基本处理流程出发, 分析了分词方法,关键词提取数量,基本分类器规则化方法,数据划分个数等对分 类结果的影响。最终,我们也给出了在时间和空间上都比较好的 bi-perceptron 学习 方法。

Mobile news, due to its natural attributes of high frequency, has become a popular area pursued by many commercial companies in China. News categorization is an important technology in news automatic process. Many supervised learning methods can be applied in this area, where Support Vector Machine(SVM) achieves the state-of-art performance with discrete features. This paper provides the idea of bi-perceptron learning to solve the binary-class classification problem in the hope of achieving comparable or even better results than SVM. Bi-perceptron learning is a divide-and-conquer idea. We proposed this idea in this paper and realized a basic approach of it. We divided the classification problem into three steps: data partition, base classification and aggregation and compared different partition and aggregation methods. Moreover, we analyzed the effect of word segmentation methods, keywords number, the regularization of base classifiers and partition number on the categorization performance. Finally, we find an approach of bi-perceptron learning that is perfect in both time and memory consumption.