Research On Tree Kernel-Based Semantic Role Labeling In Chinese

Semantic role labeling (SRL) is a particular case of shallow semantic parsing, it only labels predicate-related constituents with semantic roles in a sentence, such as agent, patient, time, place, and so on.At present, the mainstream studies of semantic role labeling focus on the feature-based method, and it can achieve high performance. However, this method also has some issues. For example: it's difficult to extract more effective features for SRL and it misses the important structural information. Current trend of SRL is to explore kernel-based method, which can effectively solve the bottleneck of those features engineering methods. This dissertation explores kernel-based SRL in Chinese and focuses on the semantic role classification.At first, we focus on how to apply current kernel-based methods to the Chinese SRL. We construct a feature-based SRL system which uses a polynomial kernel to combine features automatically. Meanwhile, we explore SRL in Chinese via tree kernel methods and explore two effect syntactic structures with respect to the characteristics of semantic role classification by extending the minimum syntactic structure. Evaluation on the Chinese PropBank shows that the tree kernel-based semantic role classification method achieves a performance of 91.53% in accuracy. We also explore composite kernel to integrate the feature-based method and the kernel-based method. The experimental results show that the accuracy is improved to 94.23%.Then, we explore the structured-fetures in Chinese SRL. Considering the dependence among the arguments of a predicate, we propose an All-Arguments Predicate Feature (AAPF) space, which can capture the dependency relation. Moreover, we introduce flat features into the kernel-based method and propose three heuristic kernel space. Experimental results on Chinese PropBank shows that our approach improves the performance of 92.54% in accuracy. Finally, we adopts composite kernel to combine tree kernel-based and feature-based approaches and the accuracy achieves 95.21%, which outperforms the state-of-the-art system.At last, we use the method described above on the Chinese nominal semantic role classification. Experimental results on Chinese NomBank shows that our method has a greater potential for further research.