Research On Named Entity Recognition For Motor Field

With the rise of markets such as smart robots,electric vehicles,and drones,the importance of the motor industry is self-evident.In the era of big data,the unstructured data in the Internet contains a wealth of knowledge in the field of electrical machinery.How to obtain valuable professional knowledge from massive unstructured data has attracted more and more attention.Named entity recognition technology is the key technology to solve such problems.Its purpose is to identify proper nouns from unstructured text and classify them.It is a basic work in the construction of high-level applications such as knowledge mapping and knowledge question answering.The study found that,on the one hand,the traditional named entity recognition method requires manual design of feature templates,which is cumbersome and limited in its implementation process.On the other hand,models using neural networks often face the problem of lack of labeled training data sets in specific fields,which makes the model performance unsatisfactory.In response to the above problems,this article combines the particularity of the motor text,focuses on the research of the named entity recognition method in the motor field based on deep learning,and improves the model performance from the model optimization and semantic expansion level.The main research contents are as follows:(1)Construction of named entity recognition data set in motor field.In view of the current situation that there is no large-scale named entity tagging corpus in the motor field,the crawler technology is used to collect and organize the motor-related corpus,and the motor entity is divided into four types of entities: physical objects,characteristic descriptions,problems/faults,and methods/technologies.Entity annotation specifications form the named entity recognition data set MAD-NER in the motor field.(2)Research on named entity recognition methods based on deep learning.Firstly,a two-way LSTM network is constructed based on the LSTM network,and then combined with the conditional random field algorithm,a method of named entity recognition in the motor field based on the two-way long short-term memory neural network and the conditional random field is proposed.This end-to-end model does not require manual feature selection,and can automatically extract motor-related entities in the text.In addition,the impact of different experimental parameters and Word2 Vec pre-training word vectors on the performance of the model is explored from the perspective of parameter optimization.(3)Aiming at the problem of polysemous words in specific named entity recognition and the lack of annotated corpus in the motor field,the idea of transfer learning is used to introduce the BERT pre-training language model to obtain general language knowledge from unsupervised training on large-scale unlabeled corpus.The context dynamically adjusts the word vector representation,enhances the semantic information of the text,and makes up for the small size of the motor data set.Based on this,the BERT-BiLSTM-CRF Chinese named entity recognition method in the motor field was further constructed,and experiments were carried out on the above data set.The results show that the comprehensive F1 value of the method proposed in this paper on the named entity recognition task in the motor field reached 83%,both exceed the recognition effect of other mainstream models.