Research On Word Representation Optimization Combining Dependency Relation And Quantification Of Semantic Contribution

Word representation is a basic problem in the field of natural language processing.Distributed word representation maps vocabulary into low-dimensional dense real-valued vectors,which can effectively capture the semantic association between vocabulary.Distributed word representation models usually use the co-occurrence characteristics of the target word and context for iterative learning,so the accuracy of word representation is closely related to context.However,the existing word representation model still has the following deficiencies in the construction and representation of the context:(1)The problem of context selection in words represent learning.Existing word representation models often use inherent context windows or randomness strategies when constructing contexts,ignoring syntactic associations between words,resulting in insufficient semantic correlation between context and target words.(2)The problem of quantifying the contribution of context words to target words in words represent learning.Most word representation models treat context words equally,ignoring the difference in semantic contribution of words in context representation,resulting in inaccurate context representation.In order to solve the shortcomings of the context in the existing word representation model,this paper combines the dependency relations and a semantic contribution quantification method to optimize the structure and representation of context in the word representation model.The research content is as follows:(1)Contextual optimization strategy based on dependency relationship and quantification of semantic contribution.For the problem of context selection in the word representation model,we propose a context optimization strategy based on dependency relations.The specific strategy is to integrate the dependency into the structure of the context and measure the difference in dependency syntax to improve the semantic relevance of the context and the target word.For the problem of quantifying the contribution of context words to target words in the word representation model,we propose a method for calculating semantic contribution based on word order and part of speech.Firstly,a weight matrix is designed according to the part of speech,and then the weight matrix and word order weighting function are used to reasonably represent the semantic contribution of words.On this basis,we propose a word embedding model EDW based on dependency relations and the quantification of semantic contributions.(2)Context re-optimization based on multi-order dependency.In order to further optimize the structure and representation of the context in the word representation model,we propose a multi-level dependency relationship representation strategy based on(1).We integrate the multi-order dependency relationship representation into the objective function of the word representation model to improve the effectiveness of the model in capturing semantic components in the dependency relationship.On this basis,we propose a word embedding model EMDW based on multi-level dependencies and the quantification of semantic contributionsWe evaluate the EDW and EMDW through word similarity,word analogy,and text classification tasks,and prove that the context optimization strategy proposed in this paper can effectively improve the accuracy of word representation.Finally,we designed and implemented a semantic search engine for scientific and technological resources based on the above research results,and applied it to ZuoChuangZhiTui precise matching platform to help users search for scientific and technological resources more accurately and comprehensively.