A Named Entity Recognition Method and System Based on Pinglattice Enhanced Linear Transformer

Abstract

The invention discloses a named entity recognition method and system based on a flat grid enhanced linear converter, and belongs to the field of natural language processing named entity recognition. Firstly, the text sequence sample is obtained and the label category of the entity is marked, and the text sequence is converted into a grid structure; then the recognition model is trained by minimizing the negative log likelihood loss function to the named entity recognition model; in the process of named entity recognition, For the text sequence to be recognized, it is used as the input of the trained named entity recognition model after text preprocessing, and the recognition result is obtained according to the maximum prediction score. The invention introduces lexical information based on a more efficient flat grid structure for lexical enhancement, provides prior knowledge and entity lexical boundary information for the model, and improves the recognition accuracy of the model for entity boundaries and entity types. The context information is modeled by using a linear converter, which reduces the complexity of the model, significantly improves the efficiency of the model operation, and has higher practical value.

Description

technical field [0001] The invention relates to the field of natural language processing named entity recognition, in particular to a method and system for named entity recognition based on a pin grid enhanced linear converter. Background technique [0002] In recent years, with the rapid development of artificial intelligence technology, text data such as news content, personal blogs, intelligent customer service dialogue content, electronic medical records, etc. have gradually occupied a higher and higher proportion, which contains a lot of useful information and huge potential. value. Named entity recognition refers to the recognition of entity nouns with specific meanings in the text, such as place names, institution names, person names and proper nouns, etc. Named entity recognition is one of the indispensable steps in processing advanced tasks of natural language processing such as machine translation, question answering system, knowledge graph construction and inform...

AI Technical Summary

Problems solved by technology

Some existing classic algorithms, such as Lattice LSTM, Soft-Lexicon, LR-CNN, LGN, etc., are very complex. Among them, Lattice LSTM only obtains lexical information ending with characters, and cannot be parallelized on the CNN model. ;Compared with Lattice LSTM, Soft-Lexicon introduces more vocabulary boundary information. Its weight is a static statistical information, which reduces the amount of calculation. compatible, but its accuracy and operation speed still need to be improved; LR-CNN adds high-level features as input feedback to adjust the weight of vocabulary to solve the problem of vocabulary conflicts, effectively improving the accuracy, but also increasing the complexity of the model , the operation speed is slow; similarly, as a graph-structured algorithm LGN, since the order of the input sequence is more important for named entity recognition, the named entity recognition based on the graph neural network still needs RNN as the underlying encoder to capture the order. Also requires a complex structure, and the training speed of the model is very slow

[0005] It can be seen that most of the existing named entity recognition models focus on the recognition accuracy. In order to improve the recognition accuracy, it is often necessary to sacrifice the model operation speed, and the improvement of the recognition accuracy is also very limited, which is very unfavorable in practical applications.

Images