Research On Semantic Role Labeling Technology Based On Deep Learning

Internet information is in the state of explosive growth and the proportion of unstructured data is increasing,so how to extract important knowledge from largescale data becomes particularly important.Through the analysis and processing of unstructured text,the knowledge expressed by it can be presented in a structured form,so as to construct a large-scale knowledge graph,make computer understand what human beings want to express more accurately,and speed up the processing ability of it.As an important goal in natural language processing,the task of event identification and extraction plays an important role in knowledge graph.However,due to the complexity and difficulty of dealing with this issue,current progress is not obvious.Since shallow semantic analysis is an important part of the task of event identification and extraction,relevant researchers begin to study it.Semantic Role Labeling,which is an important core technology of constructing event knowledge bases,can meet the needs of shallow semantic analysis.It takes the sentence as the processing unit and the predicate it contains as the core part,annotating the components in the sentence with the corresponding roles according to the relationship between these components and the predicate.A large number of studies have shown that semantic role labeling can effectively improve the performance of natural language processing applications,such as abstract generation,question answering system and machine translation.For abstract generation,semantic role labeling can fill in ”who”,”when”,”where”,”what” and other questions.For the question answering system,we can find semantic role of question first and then match the result of the template question answering to get the corresponding answer.For machine translation,the source text can be annotated with semantic roles to get a tree containing basic semantic information and then put it into the translation model.English is the most frequently used language in the world.Compared with other languages,its grammar and expression form are simpler,and the technology of semantic role labeling of it is relatively mature.In comparison,there are few corpus available in Russian and Chinese,and the semantic representation is more complex.As a result,they are relatively backward in task research.Therefore,we focus on the improvement of semantic role labeling technology in these two languages.For Russian,the main research contents include:(1)adopting the gate recurrent unit and attention mechanism to extract potential context information in the sentence,which can enrich the semantic features of words and then address the problem of low accuracy of the current model;(2)applying model fusion technology which can integrate multiple basic models to obtain the final fusion model and then further improve the accuracy of model labeling results.For Chinese,the main research contents include:(1)adding self-attention mechanism and highway bidirectional long short term memory to solve the long-term dependence problem on the basis of the existing model;(2)conducting comparative experiments on various factors affecting the model to find the best practice parameters.The experimental results show that the performance of these two models can be significantly improved,and the generalization ability of the models are greatly enhanced.