Pattern Recognition Based On Neural Network In Natural Language Processing

With the process of social informatization,the Internet has a large amount of data,and the demand for processing text data is increasing day by day,making Natural Language Processing(NLP)become one of the fields of great research value.As the basic task of text data mining,text classification is widely used in recommendation system,spam recognition,voice assistant and other tasks.How to construct a classification model with strong inductive learning ability,high interpretability and flexible scenarios has always been a challenge.In recent years,Neural Symbolic Learning,which is regarded as the way to solve this difficulty,has increasingly become the frontier of research.Based on the company's research project[1],study methods to improve the rule system,the project team proposed the "Neural Rule Engine"(NRE)model.The method used in NRE model has good application value for upgrading existing rule system and building neural rule system that does not rely on large amount of data.The innovations and main work of the NRE model are as follows:(1)Unlike previous researchers,symbolic knowledge was introduced into neural networks.The NRE model adopted a new fusion strategy—using neural networks to improve the effects of rules to enhance the learning ability and interpretability of the model.The NRE model abstracts the unified basic operation module from all the rules,and then generates the execution order and parameters of the module through the parser,and finally executes the module to generate the output in order.(2)For the module introduced into the neural network,use random windows and regular matches to automatically generate the training data set of the module,so that no training data needs to be constructed manually.After the initial module training,the training effect of the module is improved through intensive learning.For the module implemented by the neural network,it is difficult to use a single digit as the input information to represent the difficulty of the distance information between the fixed points of the sequence,and a sequence with distance information is adopted as the network input to effectively solve the problem.(3)The parser adopts the encoding-decoding architecture and introduces the Attention mechanism to predict the generation order and parameters of the modules.In the experiment,comparing the strategies with or without rule modularization,the experiment proves that the modular strategy is very effective for effective coding of rules.(4)In the experimental part,the four modules and parsers introduced into the neural network are tested in separate modules,and the influence of each module on the generalization ability of the NRE model is analyzed.Under the premise of maintaining high accuracy,the recall rate of NRE is 19.31% higher than that of the rule system in the Chinese crime case classification data set,and 5.32% in the English relation classification data set Sem Eval-2010 Task 8.The experimental results show that the NRE model can effectively improve the generalization ability of the rule,and can maintain high accuracy while significantly improving the recall rate.