Research On Patient Sentiment Analysis Based On LSTM And LDA Model

With the development of Internet medical care,a large number of patient message texts have appeared on the platform.Through the text mining of patient messages,the patient's emotional information is analyzed,which is of great significance to patients and hospitals.Two aspects of the patient's message text were studied in this paper,first of all,the research on emotional polarity classification based on deep hierarchical neural network,Furthermore,the classification of emotional topics based on LDA model.The main innovations are shown below:(1)A two-channel fusion layer based on CNN model and BLSTM model is proposed.In the traditional fusion method,the word vector trained by the CNN model and the LSTM model is simply vector spliced.The disadvantage is that the emotion results analyzed by the model in the sentence training of multi-feature information are often influenced by the non-characteristic direction information.The dual-channel fusion layer introduces a weight matrix,so that when the improved LSTM model is trained,the emotion information in the direction of the feature vector classified by the CNN model is amplified,and the emotion information in the direction of the nonfeature vector is weakened.Experiments show that the two-way fusion layer is more accurate for the emotional classification of sentences with multi-feature information.(2)A deep hierarchical network model is proposed.The CNN model and the BLSTM model are different in the direction of emotional polarity classification.The CNN model is not good at processing time series information,but it performs well in the emotional polarity classification of short text,while the BLSTM model is good at processing time series information,but can't handle it well.Emotional polarity classification of short text.The improved model is divided into two layers of regional CNN layer and BLSTM word layer,which retains time series information and feature information in the corpus.Finally,the splicing of word vectors through the two-way fusion layer shows that the new model has an increase of 7.84% compared with the improved model,a recall rate of 3.35%,and an F1 value of 2.45%.(3)A word vector replacement layer is proposed.The classification of topic models for short texts has the problem of poor context dependency and insufficient vocabulary.The experiment compares the word embedding model with the word bag model,and proposes to use the word embedding model to train the word vector space that conforms to the theme of the whole article.The purpose is to supplement The vocabulary of short text sentiment topics is classified,and at the same time,the problem of poor context dependency of short texts is solved.(4)Proposed an improved LDA model.When the short text classification is performed for the LDA model,the sampling vocabulary of the Gibbs sampling layer is single.It is proposed that the Gibbs sampling layer samples the cosine distance nearest the word vector from the word vector with a certain probability ?,and adjusts the parameters.The optimal probability ? is obtained.Experiments show that the confusion of improving the LDA model is reduced by 1.42%,and the subject consistency is increased by 3.75%.