Semantic Role Labeling Based On Structure Learning

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 (SRL) is a well-defined task with a substantial body of work and comparative evaluation.Semantic role labeling is to labels the constituents with semantic roles which have direct relation with the predicate in a sentence. The semantic roles include agent, patient, time,locations and so on. At present English SRL has achieved certain results, but is almost based on supervised methods which need a large labeled corpus.However these kinds of resources are still quite limited for Chinese.In order to deal with this problem this paper presents a semi-supervised method for SRL.Structure learning algorithm is a multi-task machine learning algorithm, which extracts the common structures of multiple tasks to improve accuracy of target task. ASO is a recently proposed linear semi-supervised structure learning algorithm, which extracts the common structures via the use of auxiliary problems with large number of unlabeled data.In this paper, we build a Chinese SRL system based on ASO.We carry out experiments based on Chinese Proposition Bank corpus and improve the accuracy of system.Many SRL systems have been built on the parsing trees, in which the constituent s of the sentence structure are identified and then classified.In contrast, in this paper we use chunk as the basic labeled unit, namely the argument s of the verbs.Along with the removal of the parsing stage, the SRL task avoids the dependence on parsing,which is always the bottleneck both of speed and precision.In addition to build a suitable auxiliary problem is the key to performance of ASO algorithm. We analysis principles and methods of building auxiliary problem, and explore a number of different auxiliary problems used in a series of experiments.The results show that the structure learning algorithm can be effectively used unlabeled corpus to improve system performance.