Semantic Role Labeling Based On Deep Neural Network

The goal of natural language processing is to realize natural language communication between human and machine,so that the computer can better understand human language text,which is an important role in the field of artificial intelligence and computer.As a technical bridge between natural language and computer language,semantic analysis technique,which converts natural language text into machine-readable logical form,is a key challenge in natural language processing towards natural language understanding.Semantic role labeling(SRL)is a simple and effective shallow semantic analysis technology,which aims to analyze the relationship between predicates and corresponding semantic roles in sentences,obtain the shallow semantic representation of sentences,and thus promote the progress of other deep semantic processing tasks.In computational linguistics,semantic role labeling task is defined as analyzing the predicate-argument structure in a sentence,identifying arguments of a given predicate and labeling corresponding semantic roles to describe the semantic meaning of arguments relative to predicates.The early SRL systems were all run in the pipeline mode and were decomposed into multiple classification subtasks.However,these methods based on statistical machine learning relies heavily on feature engineering,which requires domain knowledge and a lot of feature extraction effort.In view of the shortcomings of traditional methods,this thesis explores the SRL modeling based on deep neural network.Syntactic information plays an important role in the SRL system,so this thesis proposes a modeling method based on syntactic tree,and adopts end-to-end sequential labeling model to integrate syntactic information into the neural network model of SRL.At the same time,considering that syntactic analysis relies on time-consuming and labor-intensive computational processing,this thesis also explores semantic role labeling without syntactic dependency.Based on the proposed approach,we further optimizes the syntactic integration method,proposes a more general SRL framework,which is extended to multilingual model,and deeply studies the multilingual SRL based on deep learning.Through the evaluation on the standard data set,our SRL model achieves state-of-the-art results,and is greatly improved compared with the baseline in terms of several evaluation measures.