北京大学学报(自然科学版)

一种结合有监督学习的动态主题模型

蒋卓人1,陈燕1,高良才2,汤帜2,刘晓钟3   

  1. 1. 大连海事大学交通运输管理学院, 大连 116026; 2. 北京大学计算机科学技术研究所, 北京 100080;3. Department of Information and Library Science, Indiana University Bloomington, Bloomington, IN 47405;
  • 收稿日期:2014-06-30 出版日期:2015-03-20 发布日期:2015-03-20

A Supervised Dynamic Topic Model

JIANG Zhuoren1, CHEN Yan1, GAO Liangcai2, TANG Zhi2, LIU Xiaozhong3   

  1. 1. College of Transportation Management, Dalian Maritime University, Dalian 116026; 2. Institute of Computer Science & Technology, Peking University, Beijing 100080; 3. Department of Information and Library Science, Indiana University Bloomington, Bloomington,IN 47405;
  • Received:2014-06-30 Online:2015-03-20 Published:2015-03-20

摘要: 针对传统主题模型存在的不足, 提出一种新的结合有监督学习的动态主题模型(Supervised Dynamic Topic Model, S-DTM)。该模型不仅能够随时间的变化对语言进行动态建模, 而且结合有监督学习技术, 在主题变分推理中加入标签约束, 从而建立主题与标签之间的映射关系, 提高主题的表达解释能力。通过在一个跨越25年“以自然语言处理领域的中文期刊论文为主导”的中文语料库上的实验, 证明该模型相较于静态的有监督主题模型和无监督的动态主题模型, 具有更好的语义解释概括能力, 能更准确地反映文档的主题结构, 更精确地捕捉主题?词汇概率分布的动态演化。

关键词: 有监督学习, 动态主题模型, 变分推理

Abstract: An innovative Supervised Dynamic Topic Model (S-DTM) is developed for overcoming the limitation of tranditional topic models. S-DTM models the time-varying language dynamics and is combined with supervised learning technology by adding label restriction in topic variational inference. It makes the topic-label mapping and improves the interpret ability of topics. A set of experiments is conducted on a twenty-five-year-spanning Chinese journal paper corpus that is mainly focusing on natural language processing. Experiment results show that compared with static supervised topic model and unsupervised dynamic topic model, S-DTM has a better semantic interpretation performance, reflects the topic structure of a document more accurately, captures the dynamic evolution of the term-distribution of topics more precisely.

Key words: supervised learning, dynamic topic model, variational inference

中图分类号: