| Named Entity Recognition(NER)is a typical sequence annotation task,by which NER can quickly and accurately identify and classify entities such as names of people,places and organizations in text,providing the basic data for entity-related knowledge extraction,semantic analysis and applications.Named entity recognition occupies an important position in natural language processing tasks.Especially in building knowledge graphs,named entity recognition is especially critical for entity recognition and classification,and provides an important information source for the construction of knowledge graphs.Among them,Chinese named entity recognition has attracted much attention in the field of named entity recognition due to the specificity of the task and the prevalence of the language.Although the continuous development of deep learning technology has made new breakthroughs in English named entity recognition,Chinese text is more complex compared to English,so Chinese named entity recognition still faces some challenges.On the one hand,named entity recognition models contain many hyperparameters,and the key hyperparameters have a large impact on the model performance,and manual adjustment of these parameters is time-consuming and labor-intensive,and the best results may not be achieved after multiple adjustments.In addition,when performing pre-training tasks for named entity recognition,increasing the model size can generally improve the performance of downstream tasks;however,in some cases,further increasing the model size makes the model degrade and leads to long model training time.To address the above problems,this paper proposes a hybrid neural network approach for named entity recognition,which first optimizes the named entity recognition method of bi-directional long and short-term memory network(BiLSTM)using multi-strategy sticky bacterium algorithm(SLSMA),and then combines with ALBERT lightweight pre-training model to enhance upstream efficiency,thus improving the accuracy of named entity recognition.The specific research in this paper is as follows:(1)To address the problem that the standard sticky bacterium algorithm randomly initializes the population,which leads to uneven population distribution,and the weak oscillation effect in the late iteration of the algorithm is easy to fall into the local optimum,resulting in slow convergence,this paper proposes a multi-strategy sticky bacterium algorithm.The initialization stage of the algorithm adopts Sobol sequence to homogenize the population density;the step size is dynamically adjusted by introducing the Lévy flight perturbation strategy in the late iteration stage to make the algorithm jump out of the local optimum.To further evaluate the performance of the SLSMA algorithm,eight test functions are used for simulation verification,and the algorithms all outperform the comparison algorithms.(2)To address the problem of uncertainty in recognition results caused by artificially set parameters of the deep learning model,this paper uses a multi-strategy sticky bacterium algorithm to optimize the hyperparameters of the BiLSTM network model,and the optimized model is combined with conditional random field(CRF)for named entity recognition experiments,which can effectively improve the accuracy and robustness of named entity recognition after comparative experimental analysis.(3)To address the problem of model degradation due to low training efficiency during pre-training,a named entity recognition technique based on fusing pre-training and SBC(SLSMA-BiLSTM-CRF)models is proposed.The method first pre-trains the text using the ALBERT pre-training model,which can fully capture the contextual features of text sequences using the powerful characterization ability of the ALBERT model,while avoiding the problem of ignoring fragment roles in traditional sequence annotation methods.The downstream model uses the SBC model to extract the contextual feature information of the text and obtain the global optimal annotation sequence,and it is experimentally demonstrated that this model effectively improves the accuracy and effectiveness of named entity recognition,which provides a new idea for the downstream model to process the text efficiently.In summary,the contribution of this paper is to provide a new technical means for Chinese named entity recognition and other natural language processing tasks,which has important research value and practical significance for the difficulties of Chinese language itself. |
