Research On Agricultural Products Knowledge Graph Question Answering

Knowledge Service for agricultural domain differs from search engine service based on Internet,and it should supply convenient service with accurate and reliable answers relying on professional agricultural knowledge.Knowledge graph provides theoretical and technical support for this kind of domain knowledge service.Existing agricultural knowledge graph question answering systems concentrate on the construction of knowledge graph and development of application system,lacking of algorithm research for practice scenario.Based on knowledge graph and natural language processing,this research focus on named entity recognition and user intention recognition in question answering to provide agricultural product knowledge service.The main works are:(1)In view of the lack of open,available and annotated data set for agricultural product knowledge answering service,and considering the time-consuming and laborious method of manual cleaning and annotation to obtain agricultural texts through crawler,this paper takes the existing knowledge base of agricultural products ontology as the basis.The template-based method automatically and batch constructs the relevant agricultural product named entity recognition data set and user intent recognition data set.(2)Considering that the existing named entity recognition method based on pre-training language model,the sentence representation at the last layer of the model is regarded as the distributed representation of text,ignoring the problem that the model contains rich linguistic information.In this paper,the convolution function is used to integrate the distributed representation of sentences at all layers of the model and aggregate the linguistic information inside the model.A named entity recognition method based on progressive convolutional networks is proposed.Firstly,the text was encoded using the pre-training language model,and the distributed representation set with different emphasis was obtained.Then the representation set is fused by the progressive convolutional network to obtain the aggregate representation.Finally,conditional random field decoding is used to obtain the tag sequence.Experiments were carried out on the open data set and the named entity recognition data set of agricultural products,and the effectiveness of the method was proved.(3)In order to solve the problem that the probability of some user intentions being questioned is very small and the training data available is very small,but the system still needs to be able to respond accurately,this paper proposes a generative adversarial network assisted user intentions recognition method.Methods Based on generative adversarial network,a two-stage model training method was designed to enhance the representation and recognition ability of the model for a few classes.Experimental results on the dataset of agricultural products show that our generative adversarial net assisted method can sufficiently boost the accuracy of minority classes of user intentions and do enhance the robustness of the model.