Research On Chinese Word Sense Disambiguation Model Based On Bidirectional Recurrent Neural Network

Word sense disambiguation is an important basic task in the field of natural language processing.There are a lot of polysemy words in Chinese,which makes it easy for computers to produce ambiguity when dealing with natural language.Word sense disambiguation aims to determine the correct meaning of a polysemous word in a specific context by calculating its context.In the 21st century,with the continuous development of deep learning,people find that deep learning shows superior performance in processing large-scale data tasks.In recent years,deep learning has been gradually used in natural language processing tasks.Chinese word sense disambiguation started late and is difficult.At present,the research on Chinese word sense disambiguation using the method of in-depth learning is still in its infancy.The lack of Chinese semantic knowledge and corpus data hinders the development of Chinese word sense disambiguation.In view of the above problems,this paper studies the word sense disambiguation method based on recurrent neural network and does the following work:Firstly,this paper summarizes the research status of word sense disambiguation model based on deep learning,compares the disambiguation effects of various disambiguation models based on neural network through experiments,and decides to use Bi-LSTM(Bidirectional Long Short Term Memory,Bi-LSTM)as the basic network model to study Chinese word sense disambiguation task.Word embedding technology is used to represent the corpus without word feature tagging with high-dimensional word vectors,so that the network can automatically capture the context feature information.Furthermore,the encoder decoder framework is combined with the Bi-LSTM to study the Chinese word sense disambiguation model.Combined with attention mechanism,a disambiguation model based on seq2seq and cross-attention mechanism is proposed to further improve the prediction accuracy of the model for polysemous words.Finally,this paper attempts to introduce the pooling layer of convolutional neural network into the model to further solve the over fitting problem caused by use of high dimensional word vectors and complex network structure.The data set of the 2021 national knowledge map and Semantic Computing Conference was used to train and test the model.Through comparative experiments,it is found that the disambiguation effect of the disambiguation model based on Bi-LSTM is the best compared with other neural network disambiguation models.Its prediction accuracy reaches 86.75%and F1 value reaches 0.8765.The disambiguation effect of the two models proposed in this paper has been improved.The best one is the disambiguation model based on seq2seq and interactive attention mechanism.Its prediction accuracy has reached 92.81%and F1 value has reached 0.9375.The experimental results prove that:(1)Among the word sense disambiguation models based on deep learning,the model using bi-LSTM has the best disambiguation effect;(2)It is feasible and effective to combine encode decode framework with Bi-LSTM to build word sense disambiguation model;(3)Cross-attention mechanism can effectively improve the disambiguation accuracy of word sense disambiguation model;(4)Introducing the pooling layer of convolutional neural network into the cyclic neural network can also effectively prevent the occurrence of over fitting.