Research On Question Answering System Based On Knowledge Graph Of Real Information Environment

With the rapid development of information technology and the explosive growth of information,people's demand for accurate acquisition and convenient interaction of real information is becoming more and more urgent.In order to improve the accuracy of information promotion and reduce the consumption of manpower and time,this thesis constructs the real information knowledge graph,and completes the answer reply of the system to the natural language text in the real environment from three aspects: knowledge graph construction,semantic disambiguation and intention recognition.First,the knowledge graph based on real information environment is built.Since the data used is usually text information,using such data as a database is not only a huge memory footprint,but also takes a lot of time to retrieve information from a large amount of data.In this thesis,a BERT(Bidirectional Encoder Representation from Transformers)model based on pre-training is applied to enhance the entity information,and finally the entity label sequence of real information is output through the conditional random field through entity recognition.The entity attributes and relations are defined,and the relevant real information is represented by the way of knowledge graph,and the real information environment knowledge graph is built.Secondly,a joint semantic disambiguation method based on BERT pre-training model is proposed.This paper mainly studies the semantic disambiguation of keywords in the text to avoid the accuracy of downstream tasks being affected.It is difficult to improve the accuracy of disambiguation because the short text contains little effective information and sparse semantic expression.In this thesis,capsule network with attention mechanism and bidirectional long short-term memory network are used to extract important text semantic information,and semantic disambiguation is regarded as text classification,so as to complete the task of semantic disambiguation.Finally,a slot gate intention recognition method combining multi-information fusion and embedding is proposed.On the basis of disambiguation data,the user's questions are identified with intent.For intention recognition in colloquial Chinese essay in this context information lack enough specification,syntax,and easy to cause fuzzy semantic information and other issues,this method will character vector,vector,keywords attribute information,and solid knowledge base vector for mosaic,constitute a semantic information integration vector,and will get the semantic information fusion of vector as input of word embedded,make model to extract semantic features in depth at the same time,reduce the model in dealing with a short text segmentation error on downstream due to affect the accuracy of intention recognition tasks,and through the slot gating mechanism is used to identify the joint of slot and intention,slot door mechanism are introduced to study the explicit relationship between slot,the slot filling can be adjusted according to the results of the study to the intention of the prediction,So as to achieve better intent recognition effect.