查询无关排序主题模型  

Ranking Topic Models without Query

作　　者：肖智博[1] 车丰 吴镝[1] 李庆丰[1] 鲁明羽[1] 

机构地区：[1]大连海事大学信息科学技术学院,大连116026

出　　处：《模式识别与人工智能》2014年第7期623-630,共8页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61370070;61272369;61301185;61300082);大连市科技计划项目(No.2011A17GX073;2013J21DW006);中央高校基本科研业务费专项资金项目(No.3132013335)资助

摘　　要：主题模型已成为机器学习和自然语言处理等领域研究的重要工具,它可发现大规模语料库中的隐含主题.随着语料库规模增大,发现的主题规模也随之增大.绝大多数主题模型以词袋模型为基础,无法描述词项间的顺序关系,使得主题之间无法按照重要性区分.文中提出查询无关排序主题模型框架,利用主题间各种关系排序主题,得到有序主题列表.主题关系从主题层面评价主题影响度,继而提出词项贡献度,从词项语义层面评价主题,削弱流行但语义空泛的排序主题.由于排序主题模型尚未有公认的评价标准,将有序主题作为特征进行多文档自动文摘生成,通过文摘效果间接评价主题排序的效果.实验结果证明有序主题模型优于非排序主题模型的结果.Topic models have become important tools in machine learning and natural language processing, which can discover hidden topics in large-scale corpus. However, as the size of the corpus grows, the scale of discovered topics grows. Most topic models are on the basis of bag-of-words model, and the orders between terms cannot be described, which makes topics undistinguishable from each other. Ranking topic models without query framework is proposed in this paper, in which topics are ranked to get ordered topic list according to their relationships. Topic relationships are used to evaluate topic influence in topic level, and term significance is used to evaluate term importance in term level and popular ranking topics with little semantics are weakened. Since there is no acknowledged evaluation criterion in ranking topic model, ranked topics are used as features to perform automatic summarization of multi-document, and the performance of ranking topic models are indirectly measured by summarization performance. The experimental results show that ranking topic models outperform topic models without ranking.

关 键 词：排序主题模型 主题模型评价 多文档自动文摘 抽取性文摘 文摘句排序 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.1[自动化与计算机技术—控制科学与工程]