一种基于局部词位置相对定位的非概率主题模型  

A NON-PROBABILISTIC TOPIC MODEL BASED ON RELATIVE POSITIONING OF LOCAL WORDS

作　　者：张新豪[1] 陈知行 Zhang Xinhao;Chen Zhixing(Modern Education Technology Center,Huanghe Science and Technology College,Zhengzhou 450063,Henan,China;School of Automation,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]黄河科技学院现代教育技术中心,河南郑州450063 [2]北京理工大学自动化学院,北京100081

出　　处：《计算机应用与软件》2020年第9期215-220,262,共7页Computer Applications and Software

基　　金：河南省科技厅科技攻关项目(182102310944);北京市自然科学基金项目(1183027)。

摘　　要：为了克服n-Gram等概率主题模型在捕获词局部性时存在向量特征空间激增和稀疏性等问题,提出一种非概率主题模型。定义一个局部上下文,实现对词的相对定位进行建模;采用一个平滑核来估计局部上下文,每个核带宽检查一个唯一的局部分辨率范围;通过应用贪婪坐标下降法和损失函数的因式分解以及投影梯度下降法来求解所构建的模型,从而生成高度区分的特征。实验结果表明,该模型相比于目前先进的多数概率主题模型,不但能够高效地发现局部主题和文档表示形式,分类精度也有较大提高。In order to overcome the problem of vector feature space growing rapidly and sparsity when n-Gram and other probabilistic topic models capture word locality,this paper proposes a non-probabilistic topic model.A local context was defined to model the relative positioning of words;a smoothing kernel was used to estimate the local context,and each kernel bandwidth examined a unique range of local resolutions;the formulated model was solved by applying greedy coordinate descent method and the factorization of the loss function as well as projected gradient descent method,so as to generate highly discriminating features.The experimental results show that compared with most advanced probabilistic topic models,our model can not only efficiently discover local topics and document representations,but also improve the classification accuracy.

关 键 词：信息检索 主题模型 局部上下文 稀疏性 平滑核 分类精度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]