Semi-supervised Relation Extraction Towards Industrial Cooperation

With development of Industrial Internet,analyzing the field of industrial cooperation and realizing global cooperation precisely is an inevitable trend,which has guiding significance for decision makers in both government and enterprise.Information of industrial cooperation exists widely in industry news.There will be application value and practical significance in extracting structured information of industrial cooperation from unstructured industry news with big data and artificial intelligence technologies and pulling them together as knowledge.For a new subdivision field like industrial cooperation,there will not be much labeled data set to build models for information extraction.So that building data set is a hard and key point in the task of information extraction in the field of industrial cooperation.This thesis will focus on relation extraction in the field of industrial cooperation.In this thesis I designed and structured a relation extraction frame towards the field of industrial cooperation,which can extract pairs of entities and relations from news.The frame contains a whole process from cleaning data to combine results of extraction.There are two data path in it.One is rule extraction based on pattern matching,another is model extraction based on machine learning.This frame can extract relations automatically in the field of industrial cooperation.Besides,this thesis optimized two key models in the frame.The two models are named entity recognition(NER)model to extraction relevant enterprise and relation extraction(RE)model to extract relations in the field of industrial cooperation.The NER model in this thesis combines word vector and part-of-speech tag in embedding layer,and uses multi-layer CNN and LSTM as model structure.This can make full use of local features and global features to get better performance.For relation extraction model,this thesis proposed a method called semi-supervised relation extraction constrained by context template,which can build a model from a small data set and extend it using semi-supervised learning.It can make full use of labeled data to estimate confidence of the example with pseudo label,which will improve quality of extended data set and then build a better model.In this thesis,I build a data set of industrial cooperation and experiment on it with the two model mentioned above and the result demonstrated the effectiveness of the method proposed.And I apply the method on Decision Support System of Industrial development to provide services for users.