Research On Dependency-based Semantic Role Labeling

Semantic parsing is a fundamental as well as a tough issue to natural language understanding.Due to the difficulty in semantic parsing,the well-defined and easily-evaluated semantic role labeling(SRL)maps a natural language sentence into a formal representation of its meaning and has been drawing more and more attentions.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 makes it a bottleneck to constituent-based SRL.Meanwhile,dependency parsing receives more and more attention due to its characteristic.Therefore,this paper focuses our research on dependency-based SRL by using dependency parse trees.The contribution of this work includes:Firstly,this paper has addressed the issue of predicate identification(PI) and sense classification(SC).Since predominated SRL systems are predicate-driven,it makes predicate labeling as an essential component in real SRL applications.This paper proposes a machine learning-based method to resolve PI and SC.The experiments on CoNLL2008 test dataset show our best system can achieve F1 score of 89.9%and 82.1%for PI and SC respectively.To our knowledge,it is the best achieved performance by using the same dataset.Secondly,this paper has developed a dependency-based SRL system.The system architecture is divided as three consecutive phrases:pruning,argument identification and argument classification.This paper proposes and compares different pruning strategies which play an important role in SRL.It also explores various features and details their influence on SRL.Meanwhile,automatic dependency parsing and predicate labeling are also applied in SRL as to develop a fully automatic SRL system.The experiments on CoNLL 2008 dataset show our best SRL system can achieve F1 score of 80.94%by using automatic dependency trees and automatic predicates labeling.This performance is comparable to the best one achieved on CoNLL 2008 test.Finally,this paper has made a preliminary investigation of the dependency-based SRL for Chinese.According to the characteristic of Chinese language,this paper explores various Chinese-specific features which benefits the performance.As a seminal work on dependency-based SRL for Chinese,it exhibits an important reference value to the future work in this literature.