Research On Spoken Language Understanding Based On Knowledge Graph And Semantic Graph Technology

Spoken Language Understanding is a research hotspot in the dialogue system,which requires accurate identification of user's intent and accurate extraction of relevant requirements from the spoken sentence.The result of SLU directly determines the subsequent decision of the dialogue system.Therefore,the accuracy of the intention classification and the requirement identification directly affect the operation effect of the dialogue system.The intent detect over a sentence is considered a classification problem,and traditional research usually uses a classifier to model and classify sentence information.Extracting the user's detailed requirements for specific intents,called slot filling,is considered a serialization labeling problem,that is,the words of the original sentence are labeled with slot label information.Current research usually focuses on sentence grammatical structure and word order,looking for a global optimal label distribution scheme.Based on the existing research and learning,this dissertation proposes a method based on knowledge map and semantic graph for oral comprehension.The main contents are as follows:(1)Existing entity detection schemes typically map individual words in the text to resources in knowledge graph,fail to detect phrase combinations and map to composite semantic resources.In view of this situation,this dissertation proposes a measure method to measure the closeness of phrases and knowledge map resources,and based on this method,proposes a longest matching resource mapping scheme to solve this problem.In order to reduce the time and space overhead of the algorithm,this dissertation adopts states compression dynamic programming,proposes a resource mapping optimization algorithm,and deletes the invalid state to further improve the operating efficiency of the algorithm.The experimental results show that the proposed resource mapping algorithm can effectively map the combined phrases to the composite semantic resources,and the optimization algorithm has a higher efficiency than the baseline algorithm.(2)For the current SLU researches,focusing on the current sentence structure and word order,this dissertation proposes a combination of knowledge graph and sentence semantic parse for SLU.In this solution,entity in the original sentence would be detected by related resources in knowledge graph.Then an RNN is used to convert the sentence with the entity mark into the order sequence,constructs the semantic graph based on the order sequence for semantic parse,and obtains the logical form of the sentence.We use two bidirectional RNNs to independently model the logical form and the sentence itself to extract the respective features for joint intent detection and slot filling.Finally,a SLU model that combines the logical form of sentences with the semantic features of sentences is proposed to achieve the goal of joint intent detection and slot filling.After the experiment,the effectiveness of the solution was verified and highest score of the slot filling was obtained.