A medical electronic medical record named entity recognition method based on word vector fusion

Abstract

The application discloses a medical electronic medical record named entity recognition method based on word vector fusion, and analyzes the text content of a Chinese electronic medical record, studies the influence of different forms of word vectors on named entity recognition, and proposes a text named entity recognition method based on word vector fusion, which combines a static word vector representation model and a dynamic word vector pre-training model with context semantic information, is used for solving the static word representation of medical field words and different entities with multiple semantic information in the electronic medical record text, realizes targeted extraction of medical entities, and improves the performance of the electronic medical record text named entity recognition model.

Description

Technical Field

[0001] This invention relates to the field of named entity recognition in medical electronic medical records, and in particular to a method for named entity recognition in medical electronic medical records based on word vector fusion. Background Technology

[0002] In recent years, medical institutions both domestically and internationally have systematically stored corresponding electronic medical record data for each patient during medical practices such as consultations and clinical care. Electronic medical records are generally stored in an unstructured form, which includes information such as the patient's health status, symptoms, medications, diseases, and various test indicators. Therefore, mining the text data of electronic medical records can facilitate medical institutions and experts in analyzing patients' condition information and providing treatment suggestions.

[0003] Early named entity recognition (NENT) methods for electronic medical record texts primarily relied on manually defined rules. These methods required manual rule creation and incurred high transfer costs. Later, statistical machine learning models emerged, using pre-defined features for entity recognition. These methods generally offered improvements over rule-based approaches, but still required manual feature creation. Subsequently, with the continuous development of deep learning, deep neural network models were used for named entity recognition, achieving good results. Compared to rule-based and machine learning models, the deep neural network model BiLSTM can extract more useful feature information through memory units, achieving better recognition accuracy. To further improve model performance, researchers combined various pre-trained word vector models with deep neural network models, forming a word vector-based named entity recognition model.

[0004] For named entity recognition models based on deep neural networks, there are still problems when dealing with Chinese electronic medical record text, such as the inability to accurately represent the polysemy of characters or words, and the inaccuracy of some professional medical terms that require fixed word vector representations due to the influence of contextual semantic information. For example, in the sentence "The patient underwent laparoscopic-assisted gastric body cancer root removal surgery under general anesthesia on December 15th due to gastric body cancer," "gastric body cancer" represents the disease entity in the first instance, but in the second instance, it forms a treatment-related entity together with "root removal surgery." In the sentence "The patient was given oral oxaliplatin and leucovorin after chemotherapy," "oxaliplatin" and "leucovorin" are medical terms that require fixed word vector representations in medical text. Therefore, for entities in the context of electronic medical records, it is necessary to use pre-trained models that consider contextual semantic information (such as ELMo) for recognition, and also to use fixed word vector models (such as Word2vec and GloVe) to represent certain professional terms such as drugs. Therefore, these text contents will have a certain impact on the model's prediction performance.

[0005] Therefore, it is necessary to design a medical electronic medical record named entity recognition method based on word vector fusion to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a medical electronic medical record named entity recognition method based on word vector fusion, which can solve the problems of polysemy of characters or words, as well as the need for fixed word vector representation of some professional medical terms.

[0007] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows:

[0008] A method for named entity recognition in medical electronic medical records based on word vector fusion, characterized by the following steps:

[0009] S1, Dataset and Data Preprocessing:

[0010] The existing Chinese electronic medical record dataset is preprocessed to match text words with labels one by one, and the dataset is split into training set and test set;

[0011] S2, perform data pre-training:

[0012] The output vector dimension is set to 200. Then, the text data processed in S1 is pre-trained using three word vector pre-training models: Word2Vec, Glove, and ELMo, to obtain the corresponding word vector matrices W, G, and E.

[0013] S3, performing fusion of different word vectors:

[0014] S301, merging the word vector matrices W and G:

[0015] Word vector matrices W and G are static word vector matrices of fixed dimensions; the sum and average of word vector matrices W and G are then calculated.

[0016]

[0017] Obtain the new word vector matrix N, and set the vector dimension to 200;

[0018] S302, merging the word vector matrices E and N:

[0019] The word vector matrix E is a multi-dimensional tensor matrix, which is first reduced in dimensionality:

[0020] Assume that the vector of each token is e i m is the word vector dimension; h is the number of layers in each token, and the matrix E is reduced in dimensionality. Obtain the word vector matrix after dimensionality reduction

[0021] The vector matrix N after processing S301 and Perform horizontal splicing Obtain the concatenated vector matrix C, and set the vector dimension to 400;

[0022] S4, Model Training and Entity Recognition:

[0023] The fused word vector matrix C is then introduced into the BiLSTM-CRF model, and a high-performing model, Model, is obtained by training on the training dataset. best And perform text named entity recognition prediction on the test dataset.

[0024] Preferably, the pre-training of the EMLo model in step S2 adopts a combination of CNN and two-layer bidirectional LSTM to obtain better text word vector representation; the pre-training word vector method of Word2Vec can be either Skip-gram or Continuous Bag-of-Words (CBOW).

[0025] Furthermore, in steps S2 and S3, the dimension of the word vector is set between 200 and 400; the dimension of the word vector represents the different semantic features contained in each word in the text.

[0026] The present invention has the following beneficial effects:

[0027] This invention analyzes the text content of Chinese electronic medical records to study the impact of different forms of word vectors on named entity recognition. It proposes a text named entity recognition method based on word vector fusion. This method combines a static word vector representation model and a dynamic word vector pre-training model with contextual semantic information. It is used to solve the problem of static word representation of medical terms and different entities with multiple semantic information in electronic medical record texts. This enables targeted extraction of medical entities and improves the performance of the electronic medical record text named entity recognition model. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0029] Example 1:

[0030] like Figure 1 As shown, a method for named entity recognition in medical electronic medical records based on word vector fusion is characterized by the following steps:

[0031] S1, Dataset and Data Preprocessing:

[0032] The existing Chinese electronic medical record dataset is preprocessed to match text words with labels one by one, and the dataset is split into training set and test set;

[0033] S2, perform data pre-training:

[0034] The output vector dimension is set to 200. Then, the text data processed in S1 is pre-trained using three word vector pre-training models: Word2Vec, Glove, and ELMo, to obtain the corresponding word vector matrices W, G, and E.

[0035] S3, performing fusion of different word vectors:

[0036] S301, merging the word vector matrices W and G:

[0037] Word vector matrices W and G are static word vector matrices of fixed dimensions; the sum and average of word vector matrices W and G are then calculated.

[0038]

[0039] Obtain the new word vector matrix N, and set the vector dimension to 200;

[0040] S302, merging the word vector matrices E and N:

[0041] The word vector matrix E is a multi-dimensional tensor matrix, which is first reduced in dimensionality:

[0042] Assume that the vector of each token is e i m is the word vector dimension; h is the number of layers in each token, and the matrix E is reduced in dimensionality. Obtain the word vector matrix after dimensionality reduction

[0043] The vector matrix N after processing S301 and Perform horizontal splicing Obtain the concatenated vector matrix C, and set the vector dimension to 400;

[0044] S4, Model Training and Entity Recognition:

[0045] The fused word vector matrix C is then introduced into the BiLSTM-CRF model, and a high-performing model, Model, is obtained by training on the training dataset. best And perform text named entity recognition prediction on the test dataset.

[0046] Example 2:

[0047] In step S2, the pre-training of the EMLo model uses a combination of CNN and two-layer bidirectional LSTM to obtain better text word vector representations; the pre-training word vector method for Word2Vec can be either the Skip-gram model or the Continuous Bag-of-Words (CBOW) model.

[0048] In step S2, the dimension of the word vector is set to 200; in step S3, the dimension of the word vector matrix N is set to 200, and the dimension of the word vector in the vector matrix C is set to 400. The dimension of the word vector represents the different semantic features contained in each word in the text. A word vector dimension within a certain range can contain richer text semantic information.

[0049] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The embodiments and features described in these embodiments can be arbitrarily combined without conflict. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for named entity recognition in medical electronic medical records based on word vector fusion, characterized in that: Includes the following steps: S1, Dataset and Data Preprocessing: The existing Chinese electronic medical record dataset is preprocessed to match text words with labels one by one, and the dataset is split into training set and test set; S2, perform data pre-training: The output vector dimension is set to a fixed value. Then, the text data processed in S1 is pre-trained using three word vector pre-training models: Word2Vec, GloVe, and ELMo, to obtain the corresponding word vector matrices W, G, and E. The pre-training of the ELMo model uses a combination of CNN and two-layer bidirectional LSTM to obtain better text word vector representation. The word vector pre-training method for Word2Vec can be either the Skip-gram model or the Continuous Bag-of-Words (CBOW) model. S3, performing fusion of different word vectors: S301, merge word vector matrices W and G: word vector matrices W and G are static word vector matrices with fixed dimensions; Sum and average the word vector matrices W and G: ； Obtain the new word vector matrix N, and set the vector dimensions; S302, merging the word vector matrices E and N: The word vector matrix E is a multi-dimensional tensor matrix, which is first reduced in dimensionality: Assume that the vector for each token is ; m is the word vector dimension; h is the number of layers in each token, and the matrix E is reduced in dimensionality. Obtain the word vector matrix after dimensionality reduction ; The word vector matrix N after processing S301 and Perform horizontal splicing We obtain the concatenated word vector matrix C and set the vector dimensions; S4, Model Training and Entity Recognition: The fused word vector matrix C is then introduced into the BiLSTM-CRF model, and the model is trained on the training dataset. And perform text named entity recognition prediction on the test dataset.

2. The method for named entity recognition in medical electronic medical records based on word vector fusion according to claim 1, characterized in that: In steps S2 and S3, the dimension of the word vector is set between 200 and 400; the dimension of the word vector represents the different semantic features contained in each word in the text.

Images