Research On Knowledge Base Construction Method For Scientific And Technical Information Analysis

With the development of the Internet,the explosive growth of data has made it difficult for traditional methods to extract useful information from tons of data.The knowledge base has powerful semantic processing capacity and is also good at development and organization.These capabilities can effectively solve this problem.The knowledge base aims to describe the relationships between entities that exist in the real world and build a huge semantic network map.The key technology of constructing the knowledge base is the extraction of entities and their relationships.The research on constructing methods of knowledge Base is important facing the field of science intelligence.In this thesis,an Encoder-Decoder-based CWATT-BiLSTM-LSTMd model for entity extraction is proposed firstly.In this process,BiLSTM(bidirectional long-term and short-term memory network)model was used for encoding,and improvements have been made to the embedding layer.The decoding layer uses the LSTMd(long-short-term memory network)model with Attention mechanism.The attention mechanism can effectively improve the performance of the model,and LSTMd decoding layer can simulate the problem of tag dependencies.A series of experiments have been conducted based on the dataset in the field of science and technology intelligence,and compared with some existing models,the result shows that the model proposed in this thesis has a good improvement on the F-value.On the basis of entity extraction,the relationship extraction method is improved for the solving limitations of remote monitoring.Remote monitoring assumes that sentences with the same entities pair describe the same relationship.This method can only classify at the bag level,cannot extract the mapping between entities and sentences,and is vulnerable to noisy tags.Therefore,a RL-Tree LSTM model is proposed in this thesis.The model is for relationship extraction at the sentence level using reinforcement learning.The model can be divided into a selector and a classifier.The selector selects the correct sentence and sends it to the classifier.The classifier extracts the relationship between the entities in the sentence and provides a feedback reward to the selector.The selector and the classifier are trained together to optimize the instance selection and classification process.Comparing Experiments were conducted on the datasets that used for entity extraction,compared with some excellent models,and the performance of the selectors was independently verified experimentally.Experiments show that the RL-Tree LSTM model has excellent performance on remotely supervised data and improves the F value of relationship extraction.In summary,the method proposed in this thesis can effectively extract entities and their relationships and performance of the knowledge base is also improved.