Research And Implementation Of Named Entity Recognition Method In Mass Entrepreneurship And Innovation Based On Transfer Learning

With the booming development of innovation and entrepreneurship,the data volume in the field of mass innovation and entrepreneurship keeps climbing and faces the problem of acquiring,representing and retrieving large-scale multi-domain knowledge.As a key general technology of artificial intelligence,knowledge graph can provide technical support and drive for the field of dual-innovation services in the big data environment,and Named Entity Recognition(NER)has important research significance as a pre-task of knowledge graph construction.Currently,Named Entity Recognition in generalpurpose domain can achieve good recognition effect due to sufficient corpus and deep learning technology,but in the specific domain of mass innovation and entrepreneurship,there is no large-scale labeled corpus,and it is difficult to train a Named Entity Recognition model with high accuracy through deep NER method.Transfer learning,as a generalization means to solve the distribution difference between training data and test data,can be used for knowledge migration in different domains but the same task scenario,and provides a new solution for domain named entity recognition.This paper focuses on the problem of lack of corpus for named entity recognition in the field of mass innovation and entrepreneurship,selects the investment and financing domain as the object,and studies the solution of deep transfer learning method for named entity recognition in the investment and financing domain.Multi-task transfer learning and adversarial transfer learning methods are experimented to realize the knowledge migration from source domain to target domain,respectively.The main work is as follows.(1)A multi-task migratory learning method based on domain similarity is proposed for the lack of sufficient labeled data in the investment and financing domain.The source domain data is the annotated People’s Daily news corpus,and the target domain data is a small amount of annotated investment and financing corpus,and the joint training of named entity recognition is performed on both domains.The method uses two modes of hard sharing and asymmetric sharing to expand the common text features and entity features of the target domain.On this basis,in order to screen samples with higher similarity,the learning rate of different samples is adjusted by using cosine distance to measure the similarity between the samples and the target domain at each iteration,so that the migrated features can be used with higher value.The experimental results show that this method has an advantage over the traditional named entity recognition model in terms of entity recognition accuracy and better asymmetric sharing.(2)A named entity recognition domain migration method based on adversarial learning is proposed to address the noise problem brought by the multi-task migration learning method feature migration.At this time,domain features include private features and shared features,and the minimax adversarial learning mechanism is introduced to effectively constrain and purify the shared layer features between domains and add a gradient inversion layer to achieve discriminator parameter optimization,and experiments on the investment and financing domain dataset show that the method has better performance than the multi-task migration learning method,illustrating the adversarial mechanism for suppressing the noise problem.effectiveness.(3)Based on the above research,we explored the construction method of the knowledge graph of the investment and financing domain,and integrated the named entity identification domain migration model into this system to provide the service of entity extraction.A rule-based relationship extraction template for the investment and financing domain is designed,and a small-scale knowledge graph of the investment and financing domain is obtained based on the extracted entities.