An intelligent business platform

Abstract

The present invention relates to the technical field of natural language data processing, and in particular to an intelligent agent business middle station, including a record collection module, a word embedding module, a vector analysis module, a secondary embedding module and a business response module, which first obtains medical records and establishes a vocabulary set, then obtains the initial word vector of each word in the vocabulary set based on the initial word embedding matrix, obtains the target window size and the target learning rate according to the inner product of different initial word vectors, performs secondary word embedding on the words in the vocabulary set, and finally performs business response according to the secondary word embedding matrix. Compared with the prior art, the present invention uses medical records for secondary word embedding, and also analyzes the initial word vector of each word in the medical records to obtain a suitable target window size and target learning rate, so as to more reasonably control the effect of secondary word embedding, and solves the problem that the existing word embedding technology is not accurate enough in hospital application scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language data processing, and in particular to an intelligent business middle platform. Background Art

[0002] Natural language processing technology is widely used in the field of intelligent medical care. It can realize functions such as intelligent medical guidance and preliminary consultation to improve the patient's medical experience. Word embedding is an important part of natural language processing technology. It cleverly captures the subtle connections between words by converting them into vector form. The quality of word embedding actually determines the quality of natural language processing. In specific scenarios such as hospital medical care, existing word embedding technology may find it difficult to fully understand the special connections between certain words in the hospital environment.

[0003] For example, the words "examination" and "floor" may not be closely related in daily life, but in a hospital, they may mean the location of the examination department. Or, some drug names are transliterated, and in word embedding technology, these drug names may not directly reflect their pronunciation or spelling in the original language, and will not be associated with words such as diseases and symptoms. For another example, different regions or dialects may have different ways of saying the same thing. This special contextual relationship may not be so easy to capture for traditional word embedding models.

[0004] Therefore, people need a method that can optimize the existing word embedding technology according to the actual situation of hospital medical care, so as to solve the problem that the existing word embedding technology is not accurate enough in hospital application scenarios. Summary of the invention

[0005] Therefore, the present invention provides an intelligent business middle platform to solve the problem that the word embedding technology in the prior art is not accurate enough in hospital application scenarios.

[0006] The present invention provides an intelligent agent service middle platform, including:

[0007] A record collection module is used to obtain medical records and establish a vocabulary set based on the medical records;

[0008] The word embedding module is used to obtain the initial word embedding matrix and obtain the initial word vector of each word in the vocabulary based on the initial word embedding matrix;

[0009] The vector analysis module is used to obtain the target window size and target learning rate based on the inner product of different initial word vectors;

[0010] A secondary embedding module is used to perform secondary word embedding training on the words in the vocabulary set based on the initial word vector, the target window size and the target learning rate to obtain a secondary word embedding matrix;

[0011] The business response module is used to perform business response according to the secondary word embedding matrix.

[0012] In a preferred implementation, obtaining medical records and establishing a vocabulary set based on the medical records includes:

[0013] Obtain medical records and identify content objects of the medical records;

[0014] Extracting text from the medical records to obtain a first text set;

[0015] Acquire the associated text of the content object to obtain a second text set;

[0016] Perform word segmentation on the first text set and the second text set to obtain an initial word set;

[0017] The initial vocabulary set is screened to obtain a vocabulary set.

[0018] In a preferred implementation, the target window size and target learning rate are obtained according to the inner products of different initial word vectors, including:

[0019] Using the inner product as the distance measurement method between the initial word vectors, cluster the initial word vectors to obtain the clustering results;

[0020] According to the clustering results, the number of clusters, the number of noise points, the size difference eigenvalues ​​of all clusters and the average inner product of all clusters are obtained;

[0021] The target window size is obtained according to the number of clusters, the number of noise points and the size difference eigenvalues ​​of all clusters;

[0022] The target learning rate is obtained based on the number of clusters, the number of noise points, and the average inner product of all clusters.

[0023] In a preferred implementation, the target window size is obtained according to the following formula:

[0024]

[0025] Among them, S is the target window size, N c is the number of clusters, N n is the number of noise points, C is the size difference eigenvalue of all clusters, k is the preset adjustment coefficient, and e is a natural constant.

[0026] In a preferred implementation, the target learning rate is obtained according to the following formula:

[0027]

[0028] Among them, σ is the target learning rate, I is the average inner product of all clusters, and α and β are weight coefficients.

[0029] In a preferred implementation, based on the initial word vector, based on the target window size and the target learning rate, a secondary word embedding training is performed on the words in the vocabulary set to obtain a secondary word embedding matrix, including:

[0030] Establish an initial embedding matrix, and establish a word embedding training model based on the initial embedding matrix;

[0031] Based on the target window size, multiple target words are selected from the vocabulary set;

[0032] Divide the target vocabulary into first target vocabulary and second target vocabulary;

[0033] Input the second target vocabulary into the word embedding training model to obtain the output result;

[0034] According to the difference between the output result and the first target vocabulary, based on the target learning rate, the word embedding training model is optimized to obtain an optimized initial embedding matrix, wherein the initial embedding matrix after multiple optimizations is a secondary word embedding matrix.

[0035] In a preferred implementation, establishing an initial embedding matrix includes:

[0036] According to the clustering results, calculate the cluster center of each cluster;

[0037] Taking the number of initial word vectors in each cluster as the weight, perform weighted summation on multiple cluster centers to obtain the dimension preference vector;

[0038] A low-dimensional space is established based on the dimensions corresponding to the preset number of elements with the largest values ​​in the dimension preference vector;

[0039] Project the initial word vector into a low-dimensional space to obtain the initial embedding matrix.

[0040] In a preferred implementation, the word embedding training model is a CBOW model, the word embedding training model includes an embedding layer, and the embedding layer includes multiple initial embedding matrices.

[0041] In a preferred implementation, performing a service response according to the secondary word embedding matrix includes:

[0042] Get the target initial text;

[0043] Perform context prediction on the target initial text based on the initial word embedding matrix to obtain the first vocabulary probability vector;

[0044] Perform context prediction on the target initial text based on the secondary word embedding matrix to obtain the second vocabulary probability vector;

[0045] The first vocabulary probability vector and the second vocabulary probability vector are combined to obtain the context vocabulary of the target initial text.

[0046] In a preferred implementation, performing a service response according to the secondary word embedding matrix includes:

[0047] Get the guidance text;

[0048] Identify key words in the guidance text;

[0049] Based on the secondary word embedding matrix, obtain the location words related to the keywords;

[0050] Predict the number of patients at different locations in a hospital based on the number distribution of words with the same location.

[0051] The beneficial effects of adopting the above embodiment are:

[0052] The present invention provides an intelligent agent business middle station, including a record collection module, a word embedding module, a vector analysis module, a secondary embedding module and a business response module, which first obtains medical records, and establishes a vocabulary set based on the medical records, then obtains an initial word embedding matrix, and based on the initial word embedding matrix, obtains the initial word vector of each word in the vocabulary set, obtains the target window size and the target learning rate according to the inner product of different initial word vectors, and performs secondary word embedding training on the words in the vocabulary set based on the initial word vector, based on the target window size and the target learning rate, obtains the secondary word embedding matrix, and finally responds to the business according to the secondary word embedding matrix. Compared with the prior art, the present invention uses medical records for secondary word embedding to obtain a secondary word embedding matrix that is more suitable for the actual application scenario of the hospital, and enables subsequent business to respond based on the secondary word embedding matrix, thereby improving the accuracy of natural language processing. In addition, the present invention also analyzes the initial word vector of each word in the medical record, evaluates the representation of the initial word vector, and then obtains a suitable target window size and target learning rate, so as to more reasonably control the effect of secondary word embedding, and solves the problem that the existing word embedding technology is not accurate enough in the scenario of hospital application. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A system architecture diagram of the intelligent service center provided by the present invention;

[0054] Figure 2 A diagram of the steps of the method executed by the intelligent agent service center provided by the present invention;

[0055] Figure 3 for Figure 2 Specific step diagram of step S203;

[0056] Figure 4 for Figure 2Specific step diagram of step S204;

[0057] Figure 5 for Figure 4 Specific step diagram of step S401 in FIG. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Combination Figure 1 and Figure 2 As shown, a specific embodiment of the present invention discloses an intelligent agent service middle platform, including:

[0060] The record collection module 110 is used to execute step S201, obtain medical records, and establish a vocabulary set based on the medical records;

[0061] The word embedding module 120 is used to execute step S202, obtain an initial word embedding matrix, and obtain an initial word vector for each word in the vocabulary set based on the initial word embedding matrix;

[0062] The vector analysis module 130 is used to execute step S203, and obtain a target window size and a target learning rate according to the inner products of different initial word vectors;

[0063] The secondary embedding module 140 is used to execute step S204, perform secondary word embedding training on the words in the vocabulary set based on the initial word vector, the target window size and the target learning rate, and obtain a secondary word embedding matrix;

[0064] The business response module 150 is used to execute step S205 and perform business response according to the secondary word embedding matrix.

[0065] In the above content, medical records refer to text records generated during the medical consultation process, such as consultation records, pre-filled records, registration records, medical case studies, test reports, or prescriptions, etc. Medical records can be records of a patient's complete medical consultation process or records of multiple patients, depending on the actual situation.

[0066] Word embedding is a technique that converts words into numerical vectors that capture the semantic and syntactic relationships between words. The word embedding matrix is ​​a large two-dimensional array in which each row represents the vector representation of a word, i.e., the word vector. This matrix is ​​usually learned by the model during training, with each column corresponding to a dimension and each row corresponding to a word in the vocabulary. In the word embedding process, the window (or context window) refers to the range of words around the current word considered when training the model. The window size determines how many surrounding words the model considers at the same time to predict the current word. The learning rate is a hyperparameter used in machine learning to control the step size of model parameter updates.

[0067] In this embodiment, the initial word embedding matrix is ​​the word embedding matrix used by the trained and general natural language processing model, which can be obtained by deconstructing the general large model. The purpose of this intelligent business middle station is to perform secondary word embedding based on the initial word embedding model to obtain an optimized secondary word embedding matrix that is more in line with the hospital application scenario, provide a more reasonable word embedding matrix for other modules, and realize a smart medical system.

[0068] Compared with the prior art, the present invention uses medical records for secondary word embedding to obtain a secondary word embedding matrix that is more suitable for actual hospital application scenarios, and enables subsequent services to respond based on the secondary word embedding matrix, thereby improving the accuracy of natural language processing. In addition, the present invention also analyzes the initial word vector of each vocabulary in the medical record, evaluates the representation of the initial word vector, and then obtains a suitable target window size and target learning rate, so as to more reasonably control the effect of secondary word embedding, thereby solving the problem that the existing word embedding technology is not accurate enough in hospital application scenarios.

[0069] It should be noted that, unlike the existing word embedding technology, the purpose of the present invention is only to optimize some specific words in medical scenarios, and to optimize based on the existing word embedding matrix. At the same time, in practice, most medical records are also in the form of tables, documents, etc. Therefore, in this embodiment, only specific words need to be extracted and a vocabulary set needs to be established. The order relationship between words can be ignored, and the connection between different words can be enhanced without obtaining complete text fragments.

[0070] Furthermore, in a preferred embodiment, the above step S201, obtaining medical records and establishing a vocabulary set based on the medical records, specifically includes:

[0071] Obtain medical records and identify content objects of the medical records;

[0072] Extracting text from the medical records to obtain a first text set;

[0073] Acquire the associated text of the content object to obtain a second text set;

[0074] Perform word segmentation on the first text set and the second text set to obtain an initial word set;

[0075] The initial vocabulary set is screened to obtain a vocabulary set.

[0076] The above process is a preferred way to establish a vocabulary set, wherein the content object refers to the meaningful object described in the medical record, including behavioral objects such as blood drawing and urine test, and entity objects such as drug names. Identification of content objects can be achieved using any existing technology, such as a language recognition model based on an initial word embedding matrix, or by manually setting keywords for mechanical matching.

[0077] Related texts are other preset texts related to the content object, such as the specific location of the department, the instructions for the medicine, the description of the condition and symptoms, and other preset texts that need to be optimized. In this embodiment, related texts are introduced to improve the comprehensiveness of the secondary word embedding. Screening is to eliminate some words that are common in medical records but do not need to be optimized, such as dates, document numbers, names, etc.

[0078] It is understandable that in practice, due to the different qualities of the initial word embedding matrix, the initial word embedding matrix used by the user may have good semantic description quality. Therefore, blindly performing secondary word embedding may cause overfitting, making the originally accurate word vector inaccurate.

[0079] So, combined Figure 3 As shown, in a preferred embodiment, the above step S203, obtaining the target window size and the target learning rate according to the inner products of different initial word vectors, specifically includes:

[0080] S301, clustering the initial word vectors using the inner product as the distance measurement method between the initial word vectors to obtain a clustering result;

[0081] S302, according to the clustering results, obtain the number of clusters, the number of noise points, the size difference eigenvalues ​​of all clusters and the average inner product of all clusters;

[0082] S303, obtaining a target window size according to the number of clusters, the number of noise points, and the size difference characteristic values ​​of all clusters;

[0083] S304: Obtain a target learning rate according to the number of clusters, the number of noise points, and the average inner product of all clusters.

[0084] The distribution of the initial word vector can represent the initial association of each word in the vocabulary set. This embodiment controls the target window size and target learning rate for this factor to adjust the strength of the subsequent secondary word embedding. When the initial word vector itself represents a good degree of association (for example, the two words "fever" and "cold" have good associations), the size of the target window and the learning rate can be reduced, so that the secondary word embedding pays more attention to details, slows down the convergence speed, and prevents overfitting. On the contrary, when the initial word vector itself represents a poor degree of association (for example, "blood drawing" and "second floor" have poor associations, and in this embodiment, it is hoped that the association between the two will be enhanced), the size of the target window and the learning rate can be increased, so that the secondary word embedding pays more attention to the overall situation, and the convergence speed and learning strength are improved.

[0085] It is understandable that, for word vectors, the word vectors corresponding to words with similar meanings or related words are closer to each other, and the inner product between the two is larger. The present invention defaults to the hope to associate all words in the vocabulary set, which is equivalent to the secondary word vectors corresponding to all words in the vocabulary set are as concentrated as possible after the secondary word embedding, and the inner product is larger. Therefore, in practice, the average inner product of all word vectors and other related parameters can be calculated to roughly represent the original degree of association of words in the vocabulary set. However, in practice, there are many words, and the meanings of association between different words are also different. This general calculation method has poor accuracy.

[0086] Therefore, this embodiment uses clustering to more accurately analyze the distribution characteristics of the initial word vectors, and according to the meaning and role of the window and learning rate itself, four indicators, namely the number of clusters, the number of noise points, the size difference eigenvalues ​​of all clusters, and the average inner product of all clusters, are used for analysis to obtain a more scientific and reasonable target window size and target learning rate. Among them, the size difference eigenvalues ​​of all clusters refer to the differences in the number of vectors in all clusters, which can be represented by statistical characteristics such as the variance and standard deviation between the number of vectors in different clusters.

[0087] Specifically, the fewer the number of clusters, the better the correlation of the initial word vectors. At this time, the target window size and the target learning rate should be reduced. Similarly, the more noise points, the more words with poor overall correlation. At this time, there are more word vectors that need to be corrected. The target window size should be increased so that when the secondary word embedding matrix is ​​optimized later, the words with poor correlation can be linked to more words. The larger the size difference eigenvalue of the cluster, the greater the difference in the number of vectors between the clusters, and the worse the correlation of the initial word vectors. Similar to the case with more noise points, the target window size should be increased at this time. The larger the average inner product of all clusters, the better the aggregation between the word vectors. At this time, the learning rate needs to be reduced to avoid overfitting and maintain the original accuracy of the initial word vector.

[0088] Specifically, the above process can be expressed by the following formula: The target window size is obtained according to the following formula:

[0089]

[0090] Among them, S is the target window size, N c is the number of clusters, N n is the number of noise points, C is the size difference eigenvalue of all clusters, k is the preset adjustment coefficient, and e is a natural constant.

[0091] The target learning rate is obtained according to the following formula:

[0092]

[0093] Among them, σ is the target learning rate, i is the average inner product of all clusters, and α and β are weight coefficients.

[0094] Further, combined with Figure 4 As shown, in a preferred embodiment, the above step S204, based on the initial word vector, based on the target window size and the target learning rate, performs secondary word embedding training on the words in the vocabulary set to obtain a secondary word embedding matrix, specifically includes:

[0095] S401, establishing an initial embedding matrix, and establishing a word embedding training model according to the initial embedding matrix;

[0096] S402, selecting a plurality of target words from the vocabulary set based on the target window size;

[0097] S403, dividing the target vocabulary into first target vocabulary and second target vocabulary;

[0098] S404, inputting the second target vocabulary into the word embedding training model to obtain an output result;

[0099] S405. According to the difference between the output result and the first target vocabulary, based on the target learning rate, the word embedding training model is optimized to obtain an optimized initial embedding matrix, wherein the initial embedding matrix after multiple optimizations is a secondary word embedding matrix.

[0100] In practice, a continuous text segment is usually used for word embedding, and the word vector of a word in the text segment is used to update the word vector of its adjacent words, and finally the desired result is obtained by iteration. In this embodiment, it is hoped that all the words in the vocabulary set are connected. Therefore, in this embodiment, it is only necessary to ensure that the number of target words meets the target window size, and the first target word is used as the context of the second target word. Based on the second target word, the word embedding training model is used to predict what the context of the second target word should be. Only based on the difference between the output result and the first target word, the word vector of the second target word is corrected to achieve secondary word embedding.

[0101] It is understandable that the initial embedding matrix can be obtained by random number generation, or the initial word embedding matrix can be directly used as the initial embedding matrix. The present invention provides a more preferred method, specifically, combining Figure 5 As shown, in a preferred embodiment, in the above step S401, establishing an initial embedding matrix specifically includes:

[0102] S501, calculating the cluster center of each cluster according to the clustering result;

[0103] S502, taking the number of initial word vectors in each cluster as a weight, performing weighted summation on multiple cluster centers to obtain a dimension preference vector;

[0104] S503, establishing a low-dimensional space based on the dimensions corresponding to the preset number of elements with the largest values ​​in the dimension preference vector;

[0105] S504: Project the initial word vector into a low-dimensional space to obtain an initial embedding matrix.

[0106] It is understandable that in practice, the number of words that need to be re-determined is not as large as the usual number of words. Therefore, this embodiment essentially reduces the dimensionality of the initial word embedding matrix. On the one hand, it enables the secondary word embedding to start based on the original semantic features of the vector. On the other hand, it greatly reduces the number and significantly improves the speed of the secondary word embedding of this intelligent agent business middle station, making this middle station more flexible and real-time.

[0107] In practice, any existing method such as principal component analysis can be used to reduce the dimension of the initial word embedding matrix. In this embodiment, the clustering results in the previous text are used for analysis, which further improves the reusability of the data.

[0108] Further, in a preferred embodiment, in the above process, the word embedding training model is a CBOW model, the word embedding training model includes an embedding layer, and the embedding layer includes multiple initial embedding matrices. The CBOW model is a basic model for training word embedding matrices, and in this embodiment, the target vocabulary itself does not have the characteristics of word order, so the CBOW model can also be used to obtain a better training effect.

[0109] For example, when the target window size is 5, the first target vocabulary is 1 and the second target vocabulary is 4. The four second target vocabulary are input into the CBOW model to obtain the output result, and then the loss function is established according to the difference between the output result and the first target vocabulary, and the back propagation is performed in the CBOW model based on the loss function, and finally the word vector corresponding to the second target vocabulary in the initial embedding matrix in the embedding layer is optimized. In practice, the SKIP-GRAM model can also be used. At this time, it is just the opposite of the CBOW model. The first target vocabulary is 4 and the second target vocabulary is 1. It can be understood that the CBOW model and SKIP-GRAM are both existing technologies that can be understood by people in this field, so this article will not explain too much.

[0110] Further, in a preferred embodiment, the above step S205, performing a service response according to the secondary word embedding matrix, specifically includes:

[0111] Get the target initial text;

[0112] Perform context prediction on the target initial text based on the initial word embedding matrix to obtain the first vocabulary probability vector;

[0113] Perform context prediction on the target initial text based on the secondary word embedding matrix to obtain the second vocabulary probability vector;

[0114] The first vocabulary probability vector and the second vocabulary probability vector are combined to obtain the context vocabulary of the target initial text.

[0115] In the above process, the target initial text is the text needed by the system where this middle station is located when performing natural language processing. The significance of this embodiment is to combine the initial word embedding matrix and the secondary word embedding matrix to jointly perform context prediction and semantic understanding to obtain more accurate language processing results and improve the patient experience.

[0116] Furthermore, in a preferred embodiment, the above step S205, performing a service response according to the secondary word embedding matrix, specifically further includes:

[0117] Get the guidance text;

[0118] Identify key words in the guidance text;

[0119] Based on the secondary word embedding matrix, obtain the location words related to the keywords;

[0120] Predict the number of patients at different locations in a hospital based on the number distribution of words with the same location.

[0121] In the above process, the guidance text is the text generated when the user registers, or when consulting or asking a preliminary question in the mobile app. The significance of this embodiment is to analyze these texts, and use the association between keywords and some location words in these texts to quickly and in real time predict the number of patients who may arrive in different areas of the hospital in the next period of time, so that scheduling can be carried out in advance, relieving the pressure on staff, reducing queuing, and improving the patient's medical experience.

[0122] The present invention provides an intelligent agent business middle station, including a record collection module, a word embedding module, a vector analysis module, a secondary embedding module and a business response module, which first obtains medical records, and establishes a vocabulary set based on the medical records, then obtains an initial word embedding matrix, and based on the initial word embedding matrix, obtains the initial word vector of each word in the vocabulary set, obtains the target window size and the target learning rate according to the inner product of different initial word vectors, and performs secondary word embedding training on the words in the vocabulary set based on the initial word vector, based on the target window size and the target learning rate, obtains the secondary word embedding matrix, and finally responds to the business according to the secondary word embedding matrix. Compared with the prior art, the present invention uses medical records for secondary word embedding to obtain a secondary word embedding matrix that is more suitable for the actual application scenario of the hospital, and enables subsequent business to respond based on the secondary word embedding matrix, thereby improving the accuracy of natural language processing. In addition, the present invention also analyzes the initial word vector of each word in the medical record, evaluates the representation of the initial word vector, and then obtains a suitable target window size and target learning rate, so as to more reasonably control the effect of secondary word embedding, and solves the problem that the existing word embedding technology is not accurate enough in the scenario of hospital application.

[0123] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0124] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent agent business middle station, characterized in that: include: A record collection module is used to obtain medical records and establish a vocabulary set based on the medical records; The word embedding module is used to obtain the initial word embedding matrix and obtain the initial word vector of each word in the vocabulary based on the initial word embedding matrix; The vector analysis module is used to obtain the target window size and target learning rate based on the inner product of different initial word vectors; A secondary embedding module is used to perform secondary word embedding training on the words in the vocabulary set based on the initial word vector, the target window size and the target learning rate to obtain a secondary word embedding matrix; The business response module is used to perform business response according to the secondary word embedding matrix.

2. The intelligent agent service middle platform according to claim 1, characterized in that: Obtain medical records and build a vocabulary based on the medical records, including: Obtain medical records and identify content objects of the medical records; Extracting text from the medical records to obtain a first text set; Acquire the associated text of the content object to obtain a second text set; Perform word segmentation on the first text set and the second text set to obtain an initial word set; The initial vocabulary set is screened to obtain a vocabulary set.

3. The intelligent agent service middle platform according to claim 1, characterized in that: According to the inner product of different initial word vectors, the target window size and target learning rate are obtained, including: Using the inner product as the distance measurement method between the initial word vectors, cluster the initial word vectors to obtain the clustering results; According to the clustering results, the number of clusters, the number of noise points, the size difference eigenvalues ​​of all clusters and the average inner product of all clusters are obtained; The target window size is obtained according to the number of clusters, the number of noise points and the size difference eigenvalues ​​of all clusters; The target learning rate is obtained based on the number of clusters, the number of noise points, and the average inner product of all clusters.

4. The intelligent agent service middle platform according to claim 3 is characterized in that: The target window size is obtained according to the following formula: Among them, S is the target window size, N c is the number of clusters, N n is the number of noise points, C is the size difference eigenvalue of all clusters, k is the preset adjustment coefficient, and e is a natural constant.

5. The intelligent agent service middle station method according to claim 4 is characterized in that: The target learning rate is obtained according to the following formula: Among them, σ is the target learning rate, I is the average inner product of all clusters, and α and β are weight coefficients.

6. The intelligent agent service middle station method according to claim 3 is characterized in that: Based on the initial word vector, the target window size and the target learning rate, the words in the vocabulary set are trained for secondary word embedding to obtain a secondary word embedding matrix, including: Establish an initial embedding matrix, and establish a word embedding training model based on the initial embedding matrix; Based on the target window size, multiple target words are selected from the vocabulary set; Divide the target vocabulary into first target vocabulary and second target vocabulary; Input the second target vocabulary into the word embedding training model to obtain the output result; According to the difference between the output result and the first target vocabulary, based on the target learning rate, the word embedding training model is optimized to obtain an optimized initial embedding matrix, wherein the initial embedding matrix after multiple optimizations is a secondary word embedding matrix.

7. The intelligent agent service middle platform according to claim 6, characterized in that: Create an initial embedding matrix, including: According to the clustering results, calculate the cluster center of each cluster; Taking the number of initial word vectors in each cluster as the weight, perform weighted summation on multiple cluster centers to obtain the dimension preference vector; A low-dimensional space is established based on the dimensions corresponding to the preset number of elements with the largest values ​​in the dimension preference vector; Project the initial word vector into a low-dimensional space to obtain the initial embedding matrix.

8. The intelligent agent service middle platform according to claim 6, characterized in that: The word embedding training model is a CBOW model. The word embedding training model includes an embedding layer, and the embedding layer includes multiple initial embedding matrices.

9. The intelligent agent service middle platform according to claim 1, characterized in that: Business responses are performed based on the secondary word embedding matrix, including: Get the target initial text; Perform context prediction on the target initial text based on the initial word embedding matrix to obtain the first vocabulary probability vector; Perform context prediction on the target initial text based on the secondary word embedding matrix to obtain the second vocabulary probability vector; The first vocabulary probability vector and the second vocabulary probability vector are combined to obtain the context vocabulary of the target initial text.

10. The intelligent agent service middle platform according to claim 1, characterized in that: Business responses are performed based on the secondary word embedding matrix, including: Get the guidance text; Identify key words in the guidance text; Based on the secondary word embedding matrix, obtain the location words related to the keywords; Predict the number of patients at different locations in a hospital based on the number distribution of words with the same location.

Images