【摘要】 传统向量空问模型(VSM)特征间无关联,且不能动态增量训练,不适合主题和焦点实时变化的Internet新闻信息,为此提出了一种改进的文本实时分类模型——动态向量空间模型(DVSM)。通过对VSM的特征提取策略进行改进,提出了特征聚合和增量训练算法。通过将对分类有相同贡献的文本特征词聚合,使用它们共同的分类贡献向量特征模式作为文本特征向量的基本维;采用增量动态训练改变对分类贡献已改变的特征词在文本向量的特征模式中的位置,适应Internet新闻信息的实时特性。使用静态训练集和动态训练集进行的DVSM与传统VSM的对比实验表明,采用特征聚合和动态训练的DVSM在Internet新闻实时分类中优势效果明显优越。更多还原

【Abstract】 Traditional Vector Space Model does not consider the relationship between features, and is not suitable for dynamic training. Focus on the Internet news with dynamically changing topics and focus, a Dynamic VSM (DVSM) is proposed. Multiple discriminating features with similar contribution to classification are combined into one pattern, which is used as the basic feature dimension. When new samples need to be learned, the changed discriminating features are moved between patterns with dynamic incremental training method for the real-time characteristics of Internet. Comparison experiments using static and dynamic training sets respectively show that DVSM outperforms the traditional model significantly in Internet News Real Time Categorization.更多还原