| China is a big agricultural country and agricultural informatization plays an important role in the development of agricultural modernization.In recent years,the explosive growth of Internet information resources makes it impossible for agricultural producers to obtain target information accurately and efficiently.To solve the above problems,help the agricultural producers for the grape industry obtain demand information quickly and accurately,this study used question-answering of natural language processing to research knowledge about grapes,with deep learning model learning sentence rich semantic information and extracting sentence features,the research is carried out from three aspects of question classification,similar question matching and answer selection.Because the agricultural technical service in China is defective,this study is of great significance to improve grape yield and quality.The main work of this paper is as follows:(1)Research on the question classification algorithm based on BERT embedded Softmax model.For the collected grape knowledge questions,there are short and sparse short text points,while the traditional text classification algorithm is mainly applied to the long text with rich features,and the accuracy is low if it is applied to the short text.In this study,BERT pretraining language model with strong semantic ability was used to express the sentence level vector of preprocessed grape knowledge questions,and the acquired feature vector of the questions was input into the Softmax regression model,according to the question content,the sentences were divided into four categories.By adjusting the model parameters,and comparing with the traditional Text CNN and Fast Text classification model,the classification method in this paper has improved by 4.22% and 4.97% respectively on the F1 value,reaching 71.39%,and the classification effect is better.(2)Research on the method of calculating the similarity of questions by weighted fusion.Against the low accuracy of traditional text similarity calculation method considering single factor,the first to use three kinds of calculation methods to calculate similarity from different angles,based on the deformation of Damerau-Levenshtein distance from the view of cooccurrence word,considering the influence of word frequency on similarity,based on How Net knowledge base righteousness to the original point of view,the term take the maximum of questions keywords similarity and all word similarity separately,based on Word2vec-LSTM model studying question semantic from the perspective of deep learning;Then the three similarity values are checked by the test threshold;Finally,fusing the three similarity used weighted factor if at least two kinds of questions whose calculated values meet the threshold,then the most similar question is obtained by sorting the calculated results.Experiments show that the F1 value of MCWFS method is 73.67%,which is 9.25%,7.34% and 4.57% higher than Jaccard algorithm,Word Net method and Word2 vec method respectively.(3)The answer selection model based on Atten-Bi LSTM-2D was constructed.For the character of long length and rich information of grape answers,using Bi LSTM model learning context information,and join the attention mechanism,makes the model focus on more useful information,on this basis,join the 2D neural network,the matching features between text pairs are extracted hierarchically to calculate the relevancy of question and answer sentences and sort them.Experiments show that the MRR of the method in this paper is improved by 16.6%,12.37% and 8.08% compared with LSTM,Bi LSTM and Attention-Bi LSTM.When the value of TOP-N is 20,the maximum value of NDCG is 58.63%. |
PDF Full Text Request