Unsupervised Chinese Semantic Dependency Analysis

Incorporating knowledge into a statistical unsupervised model, an approach of semantic dependency analysis on Chinese is presented. Semantic units and part dependency relationships are identified based on knowledge first, and then analysis is given by training the unsupervised model with extended inside-outside algorithm. Despite F1 of the experiments is not better than that of supervised approach, it can be compared with the level of the state of the art unsupervised methods of SRL and the trouble of hand-annotated corpus is dispensed with. The experiments show the extended inside-outside algorithm can overcome the shortcomings of the original one such as expensive training costs, local maximum and unsatisfactory similarity with results given by linguists.