 |
    |
Tree-Structured Conditional Random Fields for Semantic Annotation*
Jie Tang1, Mingcai Hong1, Juanzi Li1, and Bangyong Liang2
1Department of Computer Science, Tsinghua University, 12#109, Tsinghua University, Beijing, 100084. China
j-tang02@mails.tsinghua.edu.cn
hmc@keg.cs.tsinghua.edu.cn
ljz@keg.cs.tsinghua.edu.cn
2NEC Labs China, 11th Floor, Innovation Plaza, Tsinghua Science Park, Beijing, 100084, China
liangbangyong@research.nec.com.cn
Abstract. The large volume of web content needs to be annotated by ontologies (called Semantic Annotation), and our empirical study shows that strong dependencies exist across different types of information (it means that identification of one kind of information can be used for identifying the other kind of information). Conditional Random Fields (CRFs) are the state-of-the-art approaches for modeling the dependencies to do better annotation. However, as information on a Web page is not necessarily linearly laid-out, the previous linear-chain CRFs have their limitations in semantic annotation. This paper is concerned with semantic annotation on hierarchically dependent data (hierarch-ical semantic annotation). We propose a Tree-structured Conditional Random Field (TCRF) model to better incorporate dependencies across the hierarchic-ally laid-out information. Methods for performing the tasks of model-parameter estimation and annotation in TCRFs have been proposed. Experimental results indicate that the proposed TCRFs for hierarchical semantic annotation can significantly outperform the existing linear-chain CRF model.
*Supported by the National Natural Science Foundation of China under Grant No. 90604025.
LNCS 4273, p. 640 ff.
lncs@springer.com
© Springer-Verlag Berlin Heidelberg 2006