Research On Joint Entity Extraction On Specific Domain Based On Deep Learning

The rapid development of the Internet has brought an era of information explosion.Various kinds of text data are produced by the network every day.How to extract useful information from these text has become an urgent problem to solve.Entity and relation information are the most basic information in text.Now we have made many researches on entity and relation extraction.Just like the traditional NER tasks on person name,place name and organization name are already mature,but according to difference between industries,entity extraction tasks in specific domain still have much work to do.In addition to the information brought by the entity,the semantic relationship information between entities is also important for analysis.Only joint entity extraction can provide information that text reflect better.There are some difficulties in joint entity extraction on specific domain.Firstly,only word2 vector encoding is too simple,for task on specific domain,only word vector lacks effective domain-related information;Secondly,on multiple targets task like joint entity extraction.Different words should be analyzed differently according to the relation with those targets.Based on the above two problems,we designs a multi-category domain dictionary coding and introduce a multi-coding fusion method to introduce more domain information.Besides,we also design a model cal ed Recurrent Attention Network Dependent On Multi-target to improve the efficiency of relationship extraction.For joint entity extraction task we research in this paper,we divides into three parts,namely domain text classification,entity extraction and relationship extraction.To solve the problem of lack of domain information,we combines several types of coding,including word embedding,part-of-speech embedding,and building domain dictionary coding by collecting domain dictionaries,to improve embedding layer and improve the final classification and extraction accuracy.For the problem of multiple targets in relation extraction,we designed a model cal ed MTD-RAM NET which introducing position feature PF,and input Attention mechanism,combining the recurrent attention mechanism to achieve differential grasp on text information.Finally,we design some comparative experiments to validate the effectiveness of the proposed embedding construction method and the relation extraction network on joint entity extraction.