Non-ferrous Metal Retrieval Key Technology Research Entity In The Field

There are a number of nonferrous data in the internet, most of these data are formed as structured, semi-structured or unstructured. How to obtaining these data fastly, conveniently, and exactly which is very useful in the nonferrous metal industry and even the whole commodity market. At the present, there is no special entity retrieval system in the field of nonferrous metal. In this paper, according to the characteristics of the non-ferrous metals, combining entity retrieval key difficult problem in the information retrieval, proceeded a serious of studies that focus on key technologies of entity recognition, entity relation extraction, entity evidence document recognition, entity ranking etc, which used in the process of information retrieval. The distinctive achievements are as follow:(1) In the field of nonferrous, Directed at the complex and strong nested structure characteristics of several entity such as products, organizations, placename etc, it proposed a kind of nonferrous metal industry entity recognition model based on Deep Neural Network (DNN) framework. The model treats this entity retrieval work as the entity sequence labeling work, in order to use the characteristics of close combination and the special field, Chinese characters are made as the input of this model, and the word embedding technology is used to map the input of Chinese character into word embedding vector firstly, and bunches the vectors initial combination feature vector as the input of DNN. The pre-training by a plurality of DNN model hidden layer (text window reduction Autoencoderr model) automatically extracts optimal feature vector as the training features of the nonferrous metals entity classifier. The experiments of entity recognition in the field of nonferrous metal show that this model is better than the conditional random field model or neural network model.(2) It proposed a model of Chinese entity relation extraction based on DBN (deep belief networks) in the field of nonferrous metal on the correlation recognition problem of all kinds of entities such as products, minerals etc in the same one document page in the field of nonferrous metal. Firstly, the relation instances are represented into word embedding vectors as the input of DBN in the field of nonferrous metal, and then uses the DBN model training to obtain the most stable feature vector as the input of BP (back propagation) network for relation extraction supervised. Finally uses the standard corpus annotation using BP neural network inverse optimization training nonferrous metal field entity relation classifier. The experimental results show that the method we proposed has an preferably effect on the retrieval task of relationships of the same kind of entity, production and sales, and subordinations in the field of nonferrous.(3) An evidence document recognition undirected graph model on nonferrous metal solid was proposed. At first, this method analyzes words, URL links,nonferrous metal entity metadata in all kinds of nonferrous metal solid evidence files such as independent page features,and the association between links and contents of the candidate nonferrous metal solid evidence documents, then constructs a nonferrous metal solid evidence document identification of undirected graph model by putting the independent page features and the relation of pages into an undirected graph,and finally uses the weight of the features of the gradient descent method in learning model, and the Gibbs sampling method recognizes the nonferrous metal solid evidence document. The experimental result shows that the method we proposed has an preferably effect.(4) It proposed a nonferrous metal entity ranking model based on deep learning. Firstly, the nonferrous metals entity queries, nonferrous metal entity candidate documents, nonferrous metal entity metadata, and nonferrous metal solid theme networks are mapped to the same semantic space respectively through a nonlinear transformation of deep architecture net. Then calculated the similarity among the concept vector that the low dimensional semantic space the query,non-ferrous metal entity metadata,non-ferrous metal entity relation and candidate document mapped in. Finally, fused the candidate document and the semantic similarity of the three vectors as the final ranking score. We can have a conclusion from the experiment that our model for the field of nonferrous metal solid ordering tasks has an preferably effect.