Product Information Mining Based On Deep Learning

Product information refers to the message,intelligence,data and knowledge of products.As an integral part of economic information,it plays an important role in the whole business process.Automatic mining of semantic items from free texts for statistical analysis and knowledge discovery has become a hot research topic in Natural Language Processing field.Today,Deep learning technology is widely used in the field of Natural Language Processing and has made breakthroughs in many aspects.The research on product information mining based on deep learning aims at mining related terms,related components and other related concept words of a product by utilizing the deep learning technologies such as deep neural network.The main work includes:1.Mining related terms of a product;2.Mining related components of a product;3.Mining other related concepts of a product.First,for mining related terms of a product,we adopt a method that combined the bidirectional long short term memory model(BiLSTM)with conditional random fields(CRF).The former takes word embedding as input.It considered the context information of keywords and made Bi-directional encoding;The latter considered the transfer probability of the previous state node to the current state node,and the global optimal decoding is carried out by using viterbi algorithm.Experimental results show that the F1 value of our model is 0.7%higher than that of CRF.Second,for mining related components of a product,we proposes an automatic extraction method based on the combination of statistics and knowledge.In order to ensure the relevance of extraction,the word frequency pruning method is used to filter out the interference words,and the context vectors are generated by Bi LSTM,and the semantic correlation is calculated.In order to ensure the accuracy of the extraction,HowNet is introduced to verify whether the extraction results are part concept words according to the classification of the word first.The experiment shows that the recognition result is 3.7%higher than the existing method F1 value.Then,for mining related other concept of a product,we mainly completed related named entities mining,including the person name,location name and organization name.The methods of mining are similar to the related term mining methods,but in terms of label category contrast is more complex,so we adopts bidirectional gated recurrent unit neural network model(BiGRU),and joined the attention mechanism,relationship between the classification of the candidate entities.The experimental results show that the above model recognition results are 6.3% higher than that of the non classified the F1.Finally,the product related information mining system is designed,which can be visualized to the above tasks.