Research On Vietnamese Semantic Role Tagging Based On Dependency Relations

Semantic parsing has always been one of the most important goals in the field of natural language processing research.The importance of shallow semantic parsing is self-evident when deep semantic parsing is difficult at present.Semantic role labeling is a common expression of shallow semantic parsing.Its task is to find the semantic parameters of the predicate of a sentence,and mark them out in the form of semantic roles.Through semantic role labeling,relevant information on the semantic level such as the agent,the receiver,and the time of occurrence of the event described by a certain sentence can be specified.This paper studied the semantic role labeling of Vietnamese based on dependency.The first stage discussed and solved the problem of the quality of the corpus.The second and third stages adopted two methods respectively to complete the labeling work.The research results achieved in this paper were as follows:(1)This paper proposed an error detection method for Vietnamese dependency treebank combining rule and treebank conversion.Aiming at the problem that the existing Vietnamese dependency treebank was not high quality and could not be used as the experimental corpus of semantic role labeling,this paper proposed an error detection method for Vietnamese dependency treebank combining rule and treebank conversion.This method fully fused Vietnamese features and grammatical characteristic,built a rule base using a traversal algorithm,and converted dependency trees into phrase trees based on the Xia's transformation algorithm.Based on whether it could be converted successfully and whether the contrast phrase types were consistent,the errors in the dependency treebank were detected and the detected errors were corrected.The experimental results showed that the proposed method could effectively improve the quality of Vietnamese dependency treebank,thus solving the problem of experimental corpus of Vietnamese semantic role labeling based on dependency.(2)This paper proposed an effective method of integrating multiple features to solve the Vietnamese semantic role labeling based on dependency.Aiming at the current research on the Vietnamese semantic role labeling based on dependency was lacking,this paper proposed an effective method of integrating multiple features to solve the Vietnamese semantic role labeling under the premise of high-quality corpus provided in the previous stage.This method took the dependency as the basic unit of labeling,fully fused the Vietnamese features and grammatical characteristic,used the Xue's pruning algorithm for preprocessing,used a greedy strategy algorithm to select rich and effective features,and trained the model based on the maximum entropy method.Finally,experiments were performed based on two corpus respectively,and multiple sets of comparative experiment were set up to compare and analyze the factors that affected system performance from different perspectives.Experimental results showed that the proposed method could effectively solve the problem of Vietnamese semantic role labeling based on dependency.(3)This paper proposed a hybrid method based on BiLSTM-CRF hybrid model to solve Vietnamese semantic role labeling.Concerning the limitations of the existing traditional machine learning methods and aiming at the deep learning method could automatically learn and construct the required features,the paper proposed a hybrid method for solving Vietnamese semantic role labeling based on the BiLSTM-CRF hybrid model.This method combined deep learning technique with traditional machine learning method,combined neural network model with linear model,used dependency as the basic unit of labeling,used the word vector tool word2 vec to convert the text to word vector sequences based on the skip-gram model,automatically extracted contextual features of dependency through Bi LSTM,used the sigmoid activation function and the CRF to train model,and finally realized the semantic role labeling.The experimental results showed that the proposed method could automatically capture various features related to dependency and solve the problem of Vietnamese semantic role labeling based on dependency better than other methods.