Research On Chinese Semantic Role Labeling Based On Multi-strategy

In recent years,Chinese information processing technology based on "big data" and "artificial intelligence" technology has been popularized and applied.Relevant research in the field of Chinese information processing has gradually deepened to the level of sentence understanding.Semantic role labeling(SRL)is the key link of sentence semantic understanding,and has been widely used in the fields of automatic question answering,machine translation,text understanding and other fields.SRL is to automatically identify and label the semantic role corresponding to the target vocabulary for the specified vocabulary in the sentence sequence.It usually consists of two sub-tasks: argument recognition and role classification.At present,research on Chinese SRL has achieved Great progress,but there are still some problems that need to be solved urgently,such as: poor model adaptability,low feature expansion efficiency,reliance on tagging accuracy for syntactic analysis,etc.These problems lead to development bottlenecks and cannot meet current intelligent information processing applications in previous research,the main methods to improve the performance of SRL focused on model selection and feature refinement,ignoring the limitations of a single labeling method and the complementarity between different methods.Therefore,this thesis proposes a fusion of multiple strategies Chinese SRL method.The study takes multiple strategies as the starting point,and introduces in detail the basic theory of SRL,the three main mainstream labeling methods proposed by the predecessors,and the labeling model under the guidance of different strategies.First,the Chinese SRL based on the linear sequence strategy is discussed.Constructed and trained a multi-feature SRL model based on conditional random fields,and used the word-sentence-sentence multi-level feature training method to preliminarily verify the scope of multi-level features.Next,the semantics of the phrase and the dependent syntactic strategies were compared based on the characteristics of role tagging,a hierarchical tree tagging strategy combining phrases and dependent syntax is proposed.Phrase-dependent double syntactic features are introduced during model training,recognition and classification tasks are performed simultaneously according to the hierarchical tree model of the double syntax.The labeling results of the double syntax tree model are compared and analyzed.Finally,an optimization model of Bi-LSTM semantic role annotation based on deep learning strategy is proposed,and a maximum pooling processing method is introduced in the model post-processing layer.By comparing the labeling characteristics of the three strategies,exploring the complementarity between the strategies and introducing the modular fusion technology,a step-by-step,multi-strategy,and modular fusion labeling method is designed.This method focuses on the corpus expansion in the fusion mode,features arrangement and module combination,based on the public Chinese syntactic tagging corpus,using Penn's Chinese corpus tagging strategy to build a sentence corpus that can add and delete custom features;in the feature orchestration,a corpus self-expansion mechanism is introduced for semi-autonomous dual Expansion of syntactic features,flexibly filter multi-level features such as morphology,syntax,etc.,to improve the semantic robustness of the labeling model;in the four basic processing links of SRL,fully leverage the strengths of the three major labeling strategies to achieve multi-module mutual assistance combination more accurate and high-quality multi-level SRL optimization model was completed.Experimental results show that the modular integration of step-by-step and multi-strategy effectively improves the labeling performance of each step of the labeling.The performance of the branch,identification and post-processing stages has been improved by approximately 3.2%,1.3%,and 0.2%,and the overall performance has been improved by approximately 1.5%.