Research And Implementation On Joint Labeling Of Syntactic And Semantic

The development of computers has aroused the desire for artificial intelligence. The rapid expansion of the Internet change the way people understand the information. It is hoped that the computer can communicate with people using natural language, read the information on the Internet, and help people to find the information exactly. But how can a computer do these? First of all, let the computer"learn"the language of humanity. This is the job of natural language understanding - an important branch of artificial intelligence. However, this is a very arduous task. Scholars began to focus on simple, but practical and smaller tasks after many failures. Semantic role labeling (SRL) is one of representative tasks. It is a shallow semantic parsing. The results can be used for a deeper semantic analysis in order to achieve the goal of natural language understanding gradually. Semantic role labeling is very important as it is a bridge between syntactic and semantic. However, this bridge also needs the support of syntactic research, such as syntactic dependency parsing (SDP).Syntactic dependency parsing and semantic role labeling has been studied many years and many good results have been obtained. But there still exists much space for improvement. This thesis studies from a different angle how to design a joint labeling system to mutual-improve syntactic and semantic labeling performance.The main contributions of this thesis are as follow:Firstly, we design and implement a new probabilistic based joint labeling model. This model generates a few groups of candidate results and then evaluates each result. The result with the best evaluation is the final result.Secondly, we modified the feature used by conventional approach of SRL to compatible with dependency-based representation.Thirdly, we design and implement a new iterative joint labeling model. The result of SRL is used as a feedback to SDP module and even to SRL module itself. Such information can be used to improve the result.Finally, for both SDP module and SRL module, we design new iterative feature for them to extend the original features in order to use the feedback information.