Automatic Labeling Of Chinese Frame Semantic Roles Based On The Dependency Features

The research work of the paper, automatic annotation of the Chinese framework semantic role, is come from the National Natural Science Foundation of "Automatically Extract of Chinese Framework for Semantic Dependency Graph(No.60970053)".Chinese FrameNet semantic role labeling has great significance for the semantic analysis of the Chinese framework, because extract the Chinese FrameNet semantic dependency graph automatically is a new approach of semantic analysis of Chinese sentences. Recently, the research of semantic roles labeling are mostly based on statistical machine learning models, and studies have shown that the main factor which restricting the semantic role labeling performance is the feature selection in statistical machine learning. Based on the Chinese FrameNet, the paper discussed the features selection and Chinese framework semantic role labeling problem by the Tree Conditional Random Fields (T-CRF). The main contents include:(1) Discuss the baseline framework for semantic role labeling model based on the word, POS, and combinations of the features. We select the word, part of speech, and combinations thereof as the basic features, construct three kinds of baseline features template according to the different sizes of the window. The best F values of the baseline model on the framework of "invention","view" and "owner" are55.36%,52.26%,66.19%.(2) Study the annotation model of the Chinese framework semantic role based on the dependency features. Add the parent-child dependencies node, parent-child dependencies relationships, and combinations thereof into the basic feature. Construct21kinds of extended features templates according to the combination of features and the size of the windows. The best F values of the framework of "invention","view" and "owner" are58.30%,55.29%,67.24%.(3) The contribution of the features are discussed for the Chinese framework semantic role labeling. The experimental results show that, it can enhance the F value of the Chinese framework semantic role labeling, when increasing the dependency syntactic level features based on the features of word, POS, and combinations thereof. After comparing the influence of dependency features on the framework semantic roles of different lengths, we found that it has a better labeling result for a longer framework semantic role; and the child nodes are more important than the parent node, dependency relationship are more important than the dependency node.The main contribution of this paper is a comprehensive study of the role of the parent-child node, the parent-child dependencies and combinations thereof of the dependency syntactic features in the Chinese FrameNet semantic role annotation. These conclusions will provide important features selection basis for further Chinese FrameNet semantic roles label study which is large-scale and open.