An encoder-based method and system for aligning medical entities

Abstract

This invention proposes an encoder-based method and system for medical entity alignment. The method includes: establishing a dataset of standard names and aliases for medical entities; preprocessing the dataset by deduplicating aliases and constructing training data using the standard names and deduplicated aliases; training a deep learning-based medical entity alignment model using the training data; obtaining the final medical entity alignment model upon model convergence; and using the final model to align aliases to standard names. The medical entity alignment model includes an encoder and a decoder; the encoder encodes the standard name dataset; the decoder uses a GRU to decode aliases, calculates the probability vector of the alias corresponding to the standard name, and obtains the standard name corresponding to the alias. Based on this method, an encoder-based medical entity alignment system is also proposed. This invention introduces medical text character vectors and word vectors, enriching text features and improving the accuracy of medical entity alignment.

Description

Technical Field

[0001] This invention belongs to the field of medical data processing technology, and specifically relates to an encoder-based method and system for aligning medical entities. Background Technology

[0002] Different hospitals use different names for the same medical entity, and different doctors also use different names for the same medical entity. This results in different expressions for the same medical entity. For example, the standard name for an examination procedure is 'pulmonary function test', but it can be expressed as 'pulmonary function examination', 'pulmonary ventilation function test', and 'pulmonary function measurement', etc. Therefore, aligning the aliases of medical entities with the standard names is an important step in constructing a medical knowledge graph.

[0003] Because a small amount of text plays a crucial role in the alignment of medical entities, while redundant text has little impact, aligning medical entities based on text similarity calculations suffers from significant errors due to the presence of redundant text, resulting in low alignment accuracy. While machine learning-based multi-class classification models perform well for data with a large volume of data within the same category, a small number of categories, and long entity text lengths, they are less accurate for medical entities like "check operation names" and "test operation names," where the volume of data within the same category is small, the number of categories is large, and the entity text length is short. This fails to meet the alignment requirements for medical entities. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a medical entity alignment method and system based on an encoder. It also incorporates character vectors and word vectors from medical text, greatly enriching the text's features and improving the alignment accuracy of medical entities.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] An encoder-based method for aligning medical entities includes the following steps:

[0007] Establish a dataset of standard names and aliases for medical entities;

[0008] The dataset is preprocessed, and the preprocessing process involves deduplicating the aliases and constructing training data using the standard names of medical entities and the deduplicated aliases.

[0009] The training data is used to train a deep learning-based medical entity alignment model. When the medical entity alignment model converges, a final medical entity alignment model is obtained. The final medical entity alignment model is used to align the alias to the standard name.

[0010] Furthermore, the medical entity standard names and aliases specifically include: test item names and aliases, test sub-item names and examination sub-item aliases, drug names and drug aliases, and surgical procedure names and surgical procedure aliases.

[0011] Furthermore, the process of constructing training data using standard medical entity names and deduplicated aliases includes:

[0012] Construct a standard name dataset using the standard names from the training data;

[0013] The deduplicated aliases and standard names are segmented using a word segmentation tool to form a medical dictionary; the word segmentation tool is Jieba, which adds a medical-specific dictionary.

[0014] Furthermore, the medical entity alignment model includes an encoder and a decoder;

[0015] The encoder is used to encode the standard name dataset;

[0016] The decoder uses a bidirectional gated recurrent neural network (GRU) to decode aliases. By calculating the probability vector of the standard name corresponding to the alias, the standard name corresponding to the alias is obtained.

[0017] Furthermore, the encoder input data is a standard name dataset I, which is composed of the sum of the character vectors and word vectors of the standard names;

[0018] The character vector is a vector obtained by mapping the semantics of each character to a Euclidean space, where the character vector space Z∈R. n×d ;

[0019] The word vector is a vector obtained by mapping the semantics of each word to a Euclidean space, where the word vector space C∈R. m×d ;

[0020] Where R represents the real number space, n is the number of characters in the dictionary, m is the number of words in the dictionary, and d is the dimension of the character vector and the word vector.

[0021] Furthermore, the encoder encoding process includes:

[0022] Let matrix I1 represent the data after encoder endword embedding, where I1∈R L×d ;

[0023] Let matrix I2 represent the data after encoder terminology embedding, where I2∈R Lc×d ;

[0024] Using I encoderThis represents the sum of the encoder-side character embedding data and the word embedding data; I encoder =I1+I2,I encoder ∈R L×d ; Calculate I encoder The maximum value on the embedded vector; where L represents the number of words in the encoder-side entity standard name, L c This indicates the number of words in the encoder's standard name.

[0025] Furthermore, the input to the decoder is:

[0026] The matrix I3 represents the data after the decoder endword embedding, where I3∈R l×d ;

[0027] The matrix I4 represents the data after the decoder word embedding, where

[0028] Using I decoder This represents the sum of the data after the decoder endword embedding and the data after the word embedding; I decoder =I3+I4,I decoder ∈R l×d ;l represents the number of words in the entity alias at the decoder end, l c This indicates the number of words in the entity aliases at the decoder end.

[0029] Furthermore, the specific process executed by the decoder includes:

[0030] Using the formula H1=BiGRU(I decoder Calculate the state vector H1 generated by the entity alias, where H1∈R l×2h_size BiGRU(·) represents a bidirectional gated recurrent neural network, and h_size represents the dimension of the hidden layer of the neural network.

[0031] Calculate the dimensionality-reduced state vector H2 = tanh(Linear(H1)), H2 ∈ R l×d ; where Linear(·) represents a linear layer neural network, transforming the dimension 2h_size of matrix H1 to d;

[0032] The standard name representation for calculating entity aliases is α = H2Y. T , α∈R l×K Where T represents matrix transpose;

[0033] Calculate V = Relu(α) T H2), V∈R K×d Where Relu(x) represents a non-linear activation function.

[0034] Calculate the product of V and the standard name representation Y, O = V × Y, O ∈ R. K×d , where × represents matrix multiplication at corresponding positions;

[0035] Calculate the probability vector of the standard name corresponding to the entity alias. Then P is a K-dimensional probability vector; where

[0036] Furthermore, the method also includes: calculating the loss function Loss based on the probability vector P output by the model.

[0037]

[0038] Among them, y i The symbolic function is defined such that if the i-th standard name is the standard name of the input alias, then y... i Select 1 otherwise select 0; P i This represents the probability that the standard name of the input alias is predicted to be the i-th standard name.

[0039] This invention also proposes an encoder-based medical entity alignment system, comprising an establishment module, a preprocessing module, and a training module:

[0040] The establishment module is used to establish a dataset of standard names and aliases for medical entities;

[0041] The preprocessing module is used to preprocess the dataset. The preprocessing process involves deduplicating the aliases and then constructing training data using the standard names of medical entities and the deduplicated aliases.

[0042] The training module is used to train a deep learning-based medical entity alignment model using the training data, and to obtain a final medical entity alignment model when the medical entity alignment model converges. The final medical entity alignment model is then used to align the alias to the standard name.

[0043] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects:

[0044] This invention proposes an encoder-based medical entity alignment method and system. The method includes the following steps: establishing a dataset of standard names and aliases for medical entities; preprocessing the dataset, wherein after deduplication of the aliases, training data is constructed using the standard names of the medical entities and the deduplicated aliases; training a deep learning-based medical entity alignment model using the training data; obtaining a final medical entity alignment model when the model converges; and aligning the aliases to the standard names using the final medical entity alignment model. The medical entity alignment model includes an encoder and a decoder; the encoder encodes the standard name dataset; the decoder uses a bidirectional gated recurrent neural network (GRU) to decode the aliases, obtaining the corresponding standard name by calculating the probability vector of the alias corresponding to the standard name. Based on this encoder-based medical entity alignment method, an encoder-based medical entity alignment system is also proposed. This invention also introduces character vectors and word vectors from medical text, greatly enriching the text features and improving the alignment accuracy of medical entities.

[0045] This method uses an encoder to input the text information of the standard name into a bidirectional gated recurrent neural network (GRU), calculates the correlation between the alias and the standard name, and incorporates the correlation into the model. This greatly improves the model's ability to learn key words in medical entities and increases the alignment accuracy of medical entities.

[0046] This invention can also achieve good alignment for medical entities whose aliases and standard names differ significantly. Attached Figure Description

[0047] like Figure 1 This is a flowchart of an encoder-based medical entity alignment method according to Embodiment 1 of the present invention;

[0048] like Figure 2 This is a schematic diagram of an encoder-based medical entity alignment system according to Embodiment 2 of the present invention. Detailed Implementation

[0049] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components and processing techniques and processes are omitted to avoid unnecessarily limiting the invention.

[0050] Example 1

[0051] Embodiment 1 of this invention proposes an encoder-based medical entity alignment method to solve the technical problem of low efficiency in medical entity alignment in the prior art.

[0052] like Figure 1 This is a flowchart of an encoder-based medical entity alignment method according to Embodiment 1 of the present invention.

[0053] In step S100, a dataset of standard names and aliases for medical entities is established. Specifically, the standard names and aliases for medical entities include: test item names and aliases, detailed test item names and aliases, drug names and aliases, and surgical procedure names and aliases. In this invention, this data is collected by medical personnel through viewing electronic medical records and medical textbooks.

[0054] In step S200, the dataset is preprocessed. The preprocessing process involves deduplicating the aliases and then constructing training data using the medical entity standard name and the deduplicated aliases.

[0055] In this application, the alias data is first deduplicated. Training data is constructed using aliases and standard names as samples. A standard name dataset is constructed using the standard names from the training data. A word segmentation tool is then used to segment the aliases and standard names, forming a medical dictionary. The word segmentation tool used here is Jieba, which incorporates medical-specific dictionaries.

[0056] In step S300, a deep learning-based medical entity alignment model is trained using training data. When the medical entity alignment model converges, a final medical entity alignment model is obtained. This final medical entity alignment model is then used to align the aliases to the standard names.

[0057] The medical entity alignment model includes an encoder and a decoder; the encoder is used to encode the standard name dataset; the decoder uses a bidirectional gated recurrent neural network (GRU) to decode aliases, and obtains the standard name corresponding to the alias by calculating the probability vector of the standard name corresponding to the alias.

[0058] The encoder input data is a standard name dataset I, which is composed of the sum of character vectors and word vectors of the standard names; the character vectors are vectors obtained by mapping the semantics of each character to Euclidean space, with the character vector space Z∈R. n×d The word vector is a vector obtained by mapping the semantics of each word to a Euclidean space, where C ∈ R. m×d Where R represents the real number space, n is the number of characters in the dictionary, m is the number of words in the dictionary, and d is the dimension of the character vector and the word vector.

[0059] The encoder encoding process includes: mapping each word in the entity standard name input to the encoder into a d-dimensional real vector using word embedding; and using matrix I1 to represent the word-embedded data at the encoder end, where I1∈R. L×d .

[0060] The entity standard names input to the encoder are segmented using a word segmentation tool, and each word is then mapped into a d-dimensional real vector using word embeddings. I2 represents the word-embedded data at the encoder end, where I2∈R. Lc×d ;

[0061] Using I encoder This represents the sum of the encoder-side character embedding data and the word embedding data; I encoder =I1+I2,I encoder ∈R L×d Where L represents the number of characters in the encoder-side entity standard name, L c This indicates the number of words in the encoder's standard name.

[0062] Calculate I encoder The maximum value on the embedded vector is given by y = tanh(layernorm(max(I)). enconder The standard name is represented by ))), where y∈R 1×d ;

[0063] If Y represents the standard name of all entities, then Y∈R K×d Here, L represents the number of words in the entity standard name. c This indicates the number of words in the standard name, where K represents the number of entity standard names;

[0064] layernorm(x) represents the normalization function, and tanh(x) is the nonlinear activation function. Where e represents the natural constant.

[0065] The input to the decoder is:

[0066] Each character in the entity alias input to the decoder is mapped into a d-dimensional real vector using character embedding. The decoder-side character-embedded data is represented by matrix I3, where I3 ∈ R. l×d ;

[0067] The aliases input to the decoder are segmented using a word segmentation tool. Each word is then mapped into a d-dimensional real vector using word embeddings. Matrix I4 represents the word-embedded data at the decoder end.

[0068] Using I decoder This represents the sum of the data after the decoder endword embedding and the data after the word embedding; Idecoder =I3+I4,I decoder ∈R l×d ;l represents the number of words in the entity alias at the decoder end, l c This indicates the number of words in the entity aliases at the decoder end.

[0069] The specific decoder execution process includes:

[0070] Using the formula H1=BiGRU(I decoder Calculate the state vector H1 generated by the entity alias, where H1∈R l×2h_size BiGRU(·) represents a bidirectional gated recurrent neural network, and h_size represents the dimension of the hidden layer of the neural network.

[0071] Calculate the dimensionality-reduced state vector H2 = tanh(Linear(H1)), H2 ∈ R l×d ; where Linear(·) represents a linear layer neural network, transforming the dimension 2h_size of matrix H1 to d;

[0072] The standard name representation for calculating entity aliases is α = H2Y. T , α∈R l×K Where T represents matrix transpose;

[0073] Calculate V = Relu(α) T H2), V∈R K×d Where Relu(x) represents a non-linear activation function.

[0074] Calculate the product of V and the standard name representation Y, O = V × Y, O ∈ R. K×d , where × represents matrix multiplication at corresponding positions;

[0075] Calculate the probability vector of the standard name corresponding to the entity alias. Then P is a K-dimensional probability vector; where

[0076] This process also includes: calculating the loss function Loss based on the probability vector P output by the model.

[0077]

[0078] Among them, y i The symbolic function is defined such that if the i-th standard name is the standard name of the input alias, then y... i Select 1 otherwise select 0; P i This represents the probability that the standard name of the input alias is predicted to be the i-th standard name.

[0079] This invention uses stochastic gradient descent to train the model until the loss of the deep learning-based medical entity alignment model no longer decreases with the increase of training times, indicating that the model has converged and training can be stopped. When the model converges, the final medical entity alignment model is obtained, and then the aliases are aligned to the standard names using the final medical entity alignment model.

[0080] The medical entity alignment method based on encoder proposed in Embodiment 1 of this invention introduces character vectors and word vectors of medical text, which greatly enriches the features of the text and improves the alignment accuracy of medical entities.

[0081] The encoder-based medical entity alignment method proposed in Embodiment 1 of this invention utilizes an encoder to introduce the textual information of the standard name into a bidirectional gated recurrent neural network (GRU), calculates the correlation between the alias and the standard name, and incorporates this correlation into the model. This significantly improves the model's ability to learn key words in medical entities, thereby increasing the alignment accuracy of medical entities. For example, the entity "paranasal sinus axial 128-slice spiral CT" can be accurately aligned to the entity "nasal sinus CT".

[0082] The encoder-based medical entity alignment method proposed in Embodiment 1 of this invention can also achieve good alignment for medical entities whose aliases and standard names differ significantly. For example, the entity "fetal NT screening (11-13+6W) color Doppler ultrasound" can be accurately aligned to the entity "NT real-time three-dimensional ultrasound examination".

[0083] Example 2

[0084] Based on the encoder-based medical entity alignment method proposed in Embodiment 1 of this invention, Embodiment 2 of this invention proposes an encoder-based medical entity alignment system, such as... Figure 2 This is a schematic diagram of an encoder-based medical entity alignment system according to Embodiment 2 of the present invention. The system includes an establishment module, a preprocessing module, and a training module.

[0085] The module is used to create a dataset of standard names and aliases for medical entities;

[0086] The preprocessing module is used to preprocess the dataset. The preprocessing process involves deduplicating the aliases and then constructing training data using the standard names of medical entities and the deduplicated aliases.

[0087] The training module is used to train a deep learning-based medical entity alignment model using the training data, and to obtain a final medical entity alignment model when the medical entity alignment model converges. The final medical entity alignment model is then used to align the alias to the standard name.

[0088] The establishment module specifically includes the following medical entity standard names and aliases: test item names and aliases, detailed test item names and aliases, drug names and aliases, and surgical procedure names and aliases. In this invention, these are collected by medical personnel by reviewing electronic medical records and medical textbooks.

[0089] In the preprocessing module, duplicate alias data is first removed. Training data is constructed using aliases and standard names as samples. A standard name dataset is then constructed using the standard names from the training data. Finally, a word segmentation tool is used to segment the aliases and standard names, forming a medical dictionary. The word segmentation tool used here is Jieba, which incorporates medical-specific dictionaries.

[0090] In the training module, the medical entity alignment model includes an encoder and a decoder; the encoder is used to encode the standard name dataset.

[0091] The decoder uses a bidirectional gated recurrent neural network (GRU) to decode aliases. By calculating the probability vector of the standard name corresponding to the alias, the standard name corresponding to the alias is obtained.

[0092] The encoder input data is a standard name dataset I, which is composed of the sum of character vectors and word vectors of the standard names; the character vectors are vectors obtained by mapping the semantics of each character to Euclidean space, with the character vector space Z∈R. n×d The word vector is a vector obtained by mapping the semantics of each word to a Euclidean space, where C ∈ R. m×d Where R represents the real number space, n is the number of characters in the dictionary, m is the number of words in the dictionary, and d is the dimension of the character vector and the word vector.

[0093] The encoder encoding process includes: mapping each word in the entity standard name input to the encoder into a d-dimensional real vector using word embedding; and using matrix I1 to represent the word-embedded data at the encoder end, where I1∈R. L×d .

[0094] The entity standard names input to the encoder are segmented using a word segmentation tool, and each word is then mapped into a d-dimensional real vector using word embeddings. I2 is used to represent the word-embedded data at the encoder end, where...

[0095] Using I encoder This represents the sum of the encoder-side character embedding data and the word embedding data; I encoder =I1+I2,I encoder ∈R L×d Where L represents the number of characters in the encoder-side entity standard name, L c This indicates the number of words in the encoder's standard name.

[0096] Calculate I encoder The maximum value on the embedded vector is given by y = tanh(layernorm(max(I)). enconder The standard name is represented by ))), where y∈R 1×d ;

[0097] If Y represents the standard name of all entities, then Y∈R K×d Here, L represents the number of words in the entity standard name. c This indicates the number of words in the standard name, where K represents the number of entity standard names;

[0098] layernorm(x) represents the normalization function, and tanh(x) is the nonlinear activation function. Where e represents the natural constant.

[0099] The input to the decoder is:

[0100] Each character in the entity alias input to the decoder is mapped into a d-dimensional real vector using character embedding. The decoder-side character-embedded data is represented by matrix I3, where I3 ∈ R. l×d ;

[0101] The aliases input to the decoder are segmented using a word segmentation tool. Each word is then mapped into a d-dimensional real vector using word embeddings. Matrix I4 represents the word-embedded data at the decoder end.

[0102] Using I decoder This represents the sum of the data after the decoder endword embedding and the data after the word embedding; I decoder =I3+I4,I decoder ∈R l×d ;l represents the number of words in the entity alias at the decoder end, l c This indicates the number of words in the entity aliases at the decoder end.

[0103] The specific decoder execution process includes:

[0104] Using the formula H1=BiGRU(I decoder Calculate the state vector H1 generated by the entity alias, where H1∈R l×2h_size BiGRU(·) represents a bidirectional gated recurrent neural network, and h_size represents the dimension of the hidden layer of the neural network.

[0105] Calculate the dimensionality-reduced state vector H2 = tanh(Linear(H1)), H2 ∈ R l×d; where Linear(·) represents a linear layer neural network, transforming the dimension 2h_size of matrix H1 to d;

[0106] The standard name representation for calculating entity aliases is α = H2Y. T , α∈R l×K Where T represents matrix transpose;

[0107] Calculate V = Relu(α) T H2), V∈R K×d Where Relu(x) represents a non-linear activation function.

[0108] Calculate the product of V and the standard name representation Y, O = V × Y, O ∈ R. K×d , where × represents matrix multiplication at corresponding positions;

[0109] Calculate the probability vector of the standard name corresponding to the entity alias. Then P is a K-dimensional probability vector; where

[0110] This process also includes: calculating the loss function Loss based on the probability vector P output by the model.

[0111]

[0112] Among them, y i The symbolic function is defined such that if the i-th standard name is the standard name of the input alias, then y... i Select 1 otherwise select 0; P i This represents the probability that the standard name of the input alias is predicted to be the i-th standard name.

[0113] This invention employs stochastic gradient descent to train the model until the loss of the deep learning-based medical entity alignment model no longer decreases with increasing training iterations, indicating that the model has converged, at which point training can be stopped. Upon model convergence, the final medical entity alignment model is obtained, and then this final model is used to align aliases to standard names.

[0114] The medical entity alignment system based on an encoder proposed in Embodiment 2 of this invention introduces character vectors and word vectors of medical text, which greatly enriches the features of the text and improves the alignment accuracy of medical entities.

[0115] The medical entity alignment system based on an encoder proposed in Embodiment 2 of this invention uses an encoder to introduce the text information of the standard name into a bidirectional gated recurrent neural network (GRU), calculates the correlation between the alias and the standard name, and introduces the correlation into the model, which greatly improves the model's learning of words that play a key role in medical entities and improves the alignment accuracy of medical entities.

[0116] The encoder-based medical entity alignment system proposed in Embodiment 2 of this invention can effectively align medical entities whose aliases and standard names differ significantly.

[0117] The description of the relevant parts of the encoder-based medical entity alignment system provided in Embodiment 2 of this application can be found in the detailed description of the corresponding parts of the encoder-based medical entity alignment method provided in Embodiment 1 of this application, and will not be repeated here.

[0118] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0119] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art can make other modifications or variations based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A medical entity alignment method based on an encoder, characterized in that, Includes the following steps: Establish a dataset of standard names and aliases for medical entities; The dataset is preprocessed, and after deduplication of the aliases, training data is constructed using the standard names of medical entities and the deduplicated aliases. The training data is used to train a deep learning-based medical entity alignment model. When the medical entity alignment model converges, a final medical entity alignment model is obtained. The final medical entity alignment model is used to align the alias to the standard name. The medical entity alignment model includes an encoder and a decoder; The encoder input data is a standard name dataset I, which is composed of the sum of the character vectors and word vectors of the standard names. The encoder encoding process includes: using matrix I1 to represent the data after word embedding at the encoder end, where I1∈R L×d The matrix I2 represents the data after word embedding by the encoder, where I2∈R. Lc×d ; using I encoder This represents the sum of the encoder-side character embedding data and the word embedding data; calculate I. encoder The maximum value on the embedded vector; where L represents the number of words in the encoder-side entity standard name, L c The number of words in the standard name of the encoder is represented by d, where d is the dimension of the character vector and word vector, and R represents the real number space. The input to the decoder is: using matrix I3 to represent the data after the decoder endword embedding, where I3∈R l×d The matrix I4 represents the data after the decoder's end-word embedding. Using I decoder This represents the sum of the character-embedded data and the word-embedded data at the decoder end; l represents the number of characters in the entity alias at the decoder end. c This indicates the number of words in the entity aliases at the decoder end; The decoder's specific execution process includes: utilizing H1 = BiGRU(I decoder Calculate the state vector H1 generated by the entity alias, where H1∈R l×2h_size BiGRU(·) represents a bidirectional gated recurrent neural network, and h_size represents the dimension of the hidden layer of the neural network; calculate the dimensionality-reduced state vector H2 = tanh(Linear(H1)), H2 ∈ R l×d ;in Linear(·) represents a linear layer neural network, transforming the dimension 2h_size of H1 to d; it calculates the standard name representation of entity aliases α = H2Y. T , α∈R l×K Where T represents matrix transpose; calculate V = Relu(α) T H2), V∈R K×d Where Relu(·) denotes a nonlinear activation function; the product of V and its standard name representation Y, O = V × Y, is calculated, O ∈ R. K×d ; Calculate the probability vector of the standard name corresponding to the entity alias.

2. The encoder-based medical entity alignment method according to claim 1, characterized in that, The medical entity standard names and aliases specifically include: test item names and aliases, test sub-item names and examination sub-item aliases, drug names and drug aliases, and surgical procedure names and surgical procedure aliases.

3. The encoder-based medical entity alignment method according to claim 1, characterized in that, The process of constructing training data using standard medical entity names and deduplicated aliases includes: Construct a standard name dataset using the standard names from the training data; The deduplicated aliases and standard names are segmented using a word segmentation tool to form a medical dictionary; the word segmentation tool is Jieba, which adds a medical-specific dictionary.

4. The encoder-based medical entity alignment method according to claim 1, characterized in that, The encoder is used to encode the standard name dataset; The decoder uses a bidirectional gated recurrent neural network to decode aliases. By calculating the probability vector of the standard name corresponding to the alias, the standard name corresponding to the alias is obtained.

5. The encoder-based medical entity alignment method according to claim 4, characterized in that, The character vector is a vector obtained by mapping the semantics of each character to a Euclidean space, where the character vector space Z∈R. n×d ; The word vector is a vector obtained by mapping the semantics of each word to a Euclidean space, where the word vector space C∈R. m×d ; Where R represents the real number space, n is the number of characters in the dictionary, and m is the number of words in the dictionary.

6. The encoder-based medical entity alignment method according to claim 5, characterized in that, Using I encoder This represents the sum of the data after character embedding and the data after word embedding at the encoder end, specifically: I encoder =I1+I2,I encoder ∈R L×d .

7. The encoder-based medical entity alignment method according to claim 6, characterized in that, Using I decoder This represents the sum of the data after character embedding and the data after word embedding, specifically: I decoder =I3+I4,I decoder ∈R l×d .

8. The encoder-based medical entity alignment method according to claim 7, characterized in that, Where Relu(·) represents a nonlinear activation function.

9. The encoder-based medical entity alignment method according to claim 8, characterized in that, The method further includes: calculating the loss function Loss based on the probability vector output by the model. Among them, y i The symbolic function is defined such that if the i-th standard name is the standard name of the input alias, then y... i Select 1 if the value is 1, otherwise select 0.

10. An encoder-based medical entity alignment system, used to execute an encoder-based medical entity alignment method according to any one of claims 1 to 9, characterized in that, It includes a module for building, a preprocessing module, and a training module: The establishment module is used to establish a dataset of standard names and aliases for medical entities; The preprocessing module is used to preprocess the dataset. After deduplicating the aliases, training data is constructed using the standard names of medical entities and the deduplicated aliases. The training module is used to train a deep learning-based medical entity alignment model using the training data, and to obtain a final medical entity alignment model when the medical entity alignment model converges. The final medical entity alignment model is then used to align the alias to the standard name.

Images