Login
Computer Science @ Rochester
Thursday, April 29, 2004
11:00 AM
CSB 209
Ph.D. Thesis Proposal
Shan He
University of Rochester
Cross Domain Semantic Role Identification
Accurate recognition of predicate-argument relationship in a sentence has been shown to be important to multiple natural language research areas, like information extraction and machine translation. And this problem can be addressed by a procedure called semantic labeling, which links syntactic constituents of a sentence to semantic roles. Semantic labeling can be done by supervised or unsupervised learning approaches. For unsupervised approaches, there is no a practical framework yet. And supervised approaches have the problems of data sparseness and domain limitation. The main concern in a new domain is how to handle new predicates. In this proposal, we propose two approaches to cope with this problem. One is to combine both labeled and unlabeled data by a Co-Training algorithm. The other approach is to generalize class knowledge about head words and target predicates to give reasonable estimates for unseen words. Both basic ideas and potential experiments are proposed for these two approaches.