词关联规则层状稀疏主题编码模型

Hierarchical Sparse Topic Coding Model Based on Words Relationship

模型规则化可以通过给模型加入先验知识,而避免模型过拟合,并且能够使模型参数稀疏,选择出最有代表性的模型参数.具有稀疏性的主题特征能够更有效的表示文本语义信息,通过WordNet和Word2Vec可以得到相似词集,将相似词集的编码向量归属为相同组,构造相似词组语义约束,并将其表示为层状结构化先验信息,论文因此而实现了两种层状稀疏规则化方法,应用于主题编码模型.实验表明采用层状规则化的稀疏编码模型可以提高主题模型编码效果,学习到主题一致性、分类结果更好的文本主题信息.

Regularization technology in the machine learning by adding the prior knowledge to the model,can not only address overfitting,but also result the sparse parameters,and gets the most representative parameters. The topic model with sparse features can more effectively represent the text semantic information. The sets of similar words can be obtained by using WordNet and Word2 Vec. After the semantic constraints of the similar words are constructed,the coding of those words can be regarded as in the same group. Then the hierarchical structured prior information can be constructed. The paper describes two class sparse coding regularization methods and applies them to topic coding. Experiments showthat those models with hierarchical regularization can improve the effect of the topic coding,and can get better subject consistency and better classification results.