高级检索

    一种面向多文本集的部分比较性LDA模型

    A Partial Comparative Cross Collections LDA Model

    • 摘要: 跨时空、跨文化文本挖掘等比较性文本挖掘(comparative text mining, CTM)旨在从多个可比的文本集中发现各文本集隐含语义结构的异同.针对当前主要的CTM模型只能分析公共话题的缺陷,提出一种部分比较性跨文本集LDA模型(partial comparative cross collections LDA model, PCCLDA)来实现跨文本集的话题分析,该模型通过层次狄利克雷过程(hierarchical Dirichlet processes, HDP) 把话题划分为公共话题和文本集特有话题,使模型能更加精确地对文本进行建模.模型采用Gibbs抽样方法进行参数推导,一系列包括Held-Out数据对数似然和模型困惑度指标在内的定量与定性的实验表明,模型不仅能够发现公共话题在不同文本集中的差异,而且能分析各文本集特有的话题;在Held-Out对数似然测度和模型困惑度指标上,PCCLDA相对当前两个主要的CTM模型具有较大的优势.

       

      Abstract: Comparative text mining like spatiotemporal and cross-cultural text mining is concerned with extracting common and unique themes from a set of comparable text collections. State-of-the-art cross collections topic models suffer from the important flaw that they can only analyze the common topics among document collections. We introduce a generative topic model PCCLDA(partial comparative cross collections LDA) for multi-collections CTM to detect both common topics and collection-special topics,and model text more exactly based on hierarchical dirichlet processes. We present a Gibbs sampling for model inference, and evaluate the model by a variety of qualitative and quantitative evaluations including model perplexity and log-likelihood measurements. PCCLDA discovers both common topics among collections and collection special topics, and also shows improvements on model perplexity and Held-Out likehood compared with two main CTM topic models.

       

    /

    返回文章
    返回