Chinese Dependency-based Semantic Role Labeling

As a research focus in natural language understanding area, the purpose of semantic analysis is to transfer the mankind's natural language into formal language that computer can understand. Due to the difficulty in deep semantic analysis, previous work mainly focuses on shallow semantic analysis, a simplified alternative to deep semantic analysis. Given a sentence and a predicate (either a verb or a noun) in it, the task of shallow semantic analysis is to recognize and map all the word sequences in the sentence into their corresponding semantic arguments or non-argument. The semantic roles include agent, patient, locative, temporal, etc. As a particular case of shallow semantic analysis, semantic role labeling (SRL) has been drawing more and more attention due to its importance in deep natural language processing applications, such as information extraction, question answering, and machine translation. According to the predicate types, SRL could be divided into SRL for verbal predicates (verbal SRL, in short) and SRL for nominal predicates (nominal SRL, in short).Most previous work on SRL focused on constituent-based parsing trees and achieved good performance. However, automatic parsing is crucial to the performance of SRL and becomes a bottleneck to constituent-based SRL. So, some researchers begin to explore dependency-based SRL. Moreover, previous research has shown that the study on Chinese SRL is much less than that on English SRL. One major reason is the lack of appropriate labeling corpus. But ,at the present stage , the Chinese corpus (Chinese PropBank and Chinese NomBank) have been issued. So, it's possible to study on Chinese SRL. This paper implements an Chinese dependency-based SRL system. The contribution of this work includes:Firstly, this paper has addressed the issue of predicate identification (PI) and predicate classification (PC). Predicate labeling is an essential component in real SRL applications and its performance directly determines the performance of SRL. This paper implements a maximum entropy classifier-based system to resolve PI and PC for Chinese verbal predicates on CoNLL2008 and CoNLL2009 corpus. In addition to maximum entropy classifier method, we also propose a Tree Kernel-based method for nominal predicates labeling on the transferred Corpus from Chinese NomBank.Secondly, we have developed a feature-based SRL system. And the emphasis of this paper is on exploring various features and detailing their influence on SRL. Respectively, for verbal predicates and nominal predicates, this paper selects different feature set and analyzes each feature's contribution to the SRL system. The experiments on nominal predicates show our system can achieve 71.37/86.20/78.09 on precision, recall, and labeled F1 score.Finally, this paper has made a preliminary investigation of the tree kernel-based SRL for Chinese nominal predicates, with focus on how to properly express the structural representation between predicates and arguments and let the input tree contain less noise information. The experiments show that the tree kernel-based method performs similar to the feature-based method. As a seminal work on tree kernel-based SRL, it exhibits an important reference value to the future work in this literature.The major contributions of this paper lie in systematic and in-depth research on semantic role labeling in Chinese from the dependency tree structure: 1)the proposal of two different methods in verbal and nominal predicates labeling; 2)the feature-based SRL system and the analysis of the various features'contributions; 3)the proposal of tree kernel-based method on Chinese SRL. Our research significantly improves the performance of SRL and thus exhibits an important reference value to the future work in semantic parsing.