Chinese Semantic Role Labeling Based On Feature Vectors

Nowadays there has been an increasing interest in shallow semantic parsing of natural language, which is becoming an important component in all kinds of natural language process applications. As a particular case, semantic role labeling is a well-defined task with a substantial body of work and comparative evaluation. A semantic role labeling system based on feature vectors is experimented on Chinese PropBank. Each key step in semantic role labeling is accomplished in a deeper research fashion. Maximum Entropy model and Support Vector Machine were used to classify semantic roles. We achieve a higher performance compare to others.Firstly, a baseline system based on six basic features is constructed. Effect of each step ( pruning,identification,classification and post-processing ) is investigated simply. Also, influence on system performance with different implementation of base features is compared.Secondly, based on baseline system, studying on feature engineering and pruning algorithm for semantic role labeling is performed. Best performance on hand-crafted and auto parser result are achieved respectively. Different feature sets are used for identification and classification. The result showed that not many and useful features which were combined effectively, can lead to higher performance. Also two pruning algorithms which performed well in English SRL are used in the system. The result shows that the two pruning are also useful is Chinese SRL. After applied the above feature sets and pruning algorithm, best system performance are achieved ( P/R/F1: 92.53 /92.28/92.40%).Finally, an integrated system which integrated of Chinese parser and semantic parser is established. It provides an operational interface.