Low-resource-constraint medical term standardization processing method based on rag retrieval

Abstract

The invention discloses a low-resource-constraint medical term standardization processing method based on rag retrieval, and relates to the technical field of character standardization processing, and the method comprises the steps: obtaining an original word of a medical term to be subjected to standardization processing, inputting the original word into a pre-training model for processing, and obtaining a standardization processing result corresponding to the original word; an RAG module is integrated in the pre-training model, and the RAG module is used for querying information associated with the original words in real time through an external knowledge base according to the original words and performing decision fusion processing on the information associated with the original words to generate a standardized processing result. The decision fusion processing of the RAG module can integrate various information sources, and a more comprehensive and more accurate standardized result is generated.

Description

Technical Field

[0001] The present invention relates to the technical field of text standardization processing, and in particular to a low-resource-constrained medical term standardization processing method based on RAG retrieval. Background Art

[0002] Medical entities can exhibit numerous morphological and grammatical variations in written records, often using different word orders or synonyms, making their standardization challenging. Current terminology standardization models rely heavily on hardware resources for training, which in turn requires significant capital investment.

[0003] 1. Rule-based methods, using string matching or dictionary lookup, which rely heavily on handcrafted rules and domain knowledge.

[0004] Disadvantages: Time-consuming and labor-intensive, unable to handle flexible and changing scenarios;

[0005] Second, a two-step framework based on recall followed by classification: The original medical entity word in the current query is similarly calculated with the standard word library, a certain number of predicted standard words are returned, and the final result is determined by 0-1 (yes / no) classification through the classification model.

[0006] Disadvantages: This method only considers literal similarity in similarity calculation, resulting in unsatisfactory results for original words with low literal coverage (i.e., significant differences in literal representation) between the original word and the corresponding standard word. It also performs poorly in one-to-many scenarios.

[0007] Due to the complex process and large model parameters, it requires greater computing power and higher hardware resources. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the deficiencies of the existing technology and specifically provide a low-resource-constrained medical terminology standardization method based on rag retrieval, as follows:

[0009] 1) In the first aspect, the present invention provides a low-resource-constrained medical terminology standardization method based on rag retrieval, the specific technical solution is as follows:

[0010] Obtaining the original words of the medical terms to be standardized, and inputting the original words into the pre-trained model for processing to obtain the standardized processing results corresponding to the original words;

[0011] The pre-training model integrates a RAG module, which is used to query the information associated with the original word in real time through an external knowledge base based on the original word, and perform decision fusion processing on the information associated with the original word to generate a standardized processing result.

[0012] The beneficial effects of the low-resource-constrained medical terminology standardization processing method based on rag retrieval provided by the present invention are as follows:

[0013] First, standardizing medical terminology through a pre-trained model can fully leverage the model's powerful language understanding and generation capabilities, effectively improving the accuracy and efficiency of standardization. Second, the integrated RAG module can query external knowledge bases in real time to obtain information associated with the original terms. This enables the standardization process to incorporate the latest medical knowledge, further improving the accuracy and reliability of the results. In addition, the decision fusion processing of the RAG module can integrate multiple information sources to generate more comprehensive and accurate standardization results, helping to address the ambiguity and complexity of medical terminology. Finally, the proposed solution performs well in the task of medical terminology standardization, with high performance and generalization capabilities.

[0014] Based on the above solution, the present invention can also be improved as follows.

[0015] Furthermore, the pre-trained model is trained by combining the Transformer architecture with an attention optimization processing strategy and a result constraint processing strategy.

[0016] Furthermore, the result constraint processing strategy is specifically as follows:

[0017] Constructing a prefix tree, wherein a root node of the prefix tree represents a blank character, and child nodes other than the root node in the prefix tree represent characters;

[0018] In the current time step, according to the child nodes, at least one candidate character is determined and used as the input data for the next time step;

[0019] Repeat the determination of candidate characters, and stop determining candidate characters when the length of the entity name composed of the candidate characters meets a preset requirement.

[0020] Furthermore, it also includes:

[0021] The standardized processing result is displayed, and the user's result confirmation instruction is obtained through the interactive platform to complete the processing of the original word.

[0022] 2) In a second aspect, the present invention further provides a medical terminology standardization processing system with low resource constraints based on rag retrieval, the specific technical solution of which is as follows:

[0023] The acquisition unit is used to: acquire the original word of the medical term to be standardized, and input the original word into the pre-training model for processing to obtain the standardized processing result corresponding to the original word;

[0024] The pre-training model integrates a RAG module, which is used to query the information associated with the original word in real time through an external knowledge base based on the original word, and perform decision fusion processing on the information associated with the original word to generate a standardized processing result.

[0025] Based on the above solution, the present invention can also be improved as follows.

[0026] Furthermore, the pre-trained model is trained by combining the Transformer architecture with an attention optimization processing strategy and a result constraint processing strategy.

[0027] Furthermore, the result constraint processing strategy is specifically as follows:

[0028] Constructing a prefix tree, wherein a root node of the prefix tree represents a blank character, and child nodes other than the root node in the prefix tree represent characters;

[0029] In the current time step, according to the child nodes, at least one candidate character is determined and used as the input data for the next time step;

[0030] Repeat the determination of candidate characters, and stop determining candidate characters when the length of the entity name composed of the candidate characters meets a preset requirement.

[0031] Furthermore, it also includes:

[0032] The display unit is used to display the standardization processing result and obtain the user's result confirmation instruction through the interactive platform to complete the processing of the original word.

[0033] 3) In a third aspect, the present invention further provides an electronic device, comprising a processor, wherein the processor is coupled to a memory, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor so that the electronic device implements any of the above methods.

[0034] 4) In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores at least one computer program, and the at least one computer program is loaded and executed by a processor to enable a computer to implement any of the above methods.

[0035] It should be noted that the beneficial effects achieved by the technical solutions of the second to fourth aspects of the present invention and the corresponding possible implementation methods can be found in the above-mentioned technical effects of the first aspect and its corresponding possible implementation methods, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0037] Figure 1 Schematic diagram of a process flow of a medical term standardization method with low resource constraints based on rag retrieval according to an embodiment of the present invention;

[0038] Figure 2 Schematic diagram showing a comparison before and after attention optimization of a low-resource-constrained medical terminology standardization method based on rag retrieval according to an embodiment of the present invention;

[0039] Figure 3 This is a structural framework diagram of an electronic device of the present invention. DETAILED DESCRIPTION

[0040] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0041] like Figure 1 As shown, a low-resource-constrained medical term standardization processing method based on rag retrieval in an embodiment of the present invention includes the following steps:

[0042] S1, obtaining the original words of the medical terms to be standardized, and inputting the original words into the pre-training model for processing to obtain the standardized processing results corresponding to the original words;

[0043] The pre-training model integrates a RAG module, which is used to query the information associated with the original word in real time through an external knowledge base based on the original word, and perform decision fusion processing on the information associated with the original word to generate a standardized processing result.

[0044] The beneficial effects of the low-resource-constrained medical terminology standardization processing method based on rag retrieval provided by the present invention are as follows:

[0045] First, standardizing medical terminology through a pre-trained model can fully leverage the model's powerful language understanding and generation capabilities, effectively improving the accuracy and efficiency of standardization. Second, the integrated RAG module can query external knowledge bases in real time to obtain information associated with the original terms. This enables the standardization process to incorporate the latest medical knowledge, further improving the accuracy and reliability of the results. In addition, the decision fusion processing of the RAG module can integrate multiple information sources to generate more comprehensive and accurate standardization results, helping to address the ambiguity and complexity of medical terminology. Finally, the proposed solution performs well in the task of medical terminology standardization, with high performance and generalization capabilities.

[0046] Example 1: In order to overcome a series of problems existing in the prior art, this solution reduces resource consumption by optimizing the attention mechanism, uses beam search to constrain the generated results, and finally combines rag retrieval to increase candidate results to assist model decision-making to achieve the goal.

[0047] 1) Training data format

[0048] The training data format constructed by matching the original words and standard words one by one is as follows:

[0049]

[0050] 2) Model structure:

[0051] Based on the classic Transformer architecture, attention optimization and result constraints are performed. The encoder inputs the original words, and the decoder outputs the standard words for training.

[0052] Attention optimization: To adapt to limited hardware resources, the attention mechanism is optimized and windowing is used to reduce the amount of computation.

[0053] The traditional self-attention mechanism can see and integrate all other characters, that is, each character must interact with all other characters in the sequence, so its time and space complexity is as high as O(n 2 )

[0054] To this end, this application sets up a sliding window to address the limited training resources in small resource scenarios. The total number of characters in each window is limited to w, that is, the current character is limited to seeing only w / 2 characters on each side. Therefore, its time complexity is O(n*w). The smaller the window setting, the greater the degree of reduction in complexity compared to the traditional self-attention mechanism.

[0055] The specific process of sliding window attention optimization:

[0056] 1. Initialization

[0057] Input sequence: Assume that the input sequence is X=[x1,x2,...,x n ], where n is the length of the sequence.

[0058] Window size: Set the size of the sliding window to w, and the total number of characters in the window to w.

[0059] Window step: Usually, the window step can be set to 1, that is, the window slides one character position each time, so that every character in the sequence can be covered by the window.

[0060] Processing at each time step:

[0061] For each character x in the sequence i (i ranges from 1 to n), and the process is as follows:

[0062] Determine the window range:

[0063] Calculate the current character x i The window has left border left and right border right.

[0064] Left border That is, the left border of the window cannot be less than 1 to prevent it from crossing the border.

[0065] Right border That is, the right boundary of the window cannot exceed the length n of the sequence to prevent out-of-bounds.

[0066] This determines the x i The sliding window range is centered, and the characters in the window are [x left ,x left+1 ,...,x right ].

[0067] 2. Calculate attention weight:

[0068] In the window range, for each character x in the window j (j from left to right), calculate the current character x i with x j The attention weight between .

[0069] The calculation formula of attention weight can be calculated using the traditional self-attention mechanism, for example: Among them, score(x i ,x j ) is a measure of x i and x j Functions of similarity or correlation between them, common ones include dot product and additive attention.

[0070] This gets the current character x i Attention weight distribution with other characters within the window range.

[0071] 3. Weighted summation to obtain context representation:

[0072] According to the calculated attention weight, the characters in the window are weighted and summed to get the current character x i The context representation c i .

[0073] The formula is:

[0074] This context indicates that c iThe information of other characters within the window range is integrated for subsequent decoding or other tasks.

[0075] 4. Cyclic process

[0076] Starting from the first character x1 of the sequence, each character in the sequence is processed in turn according to the processing steps under each time step mentioned above until the last character x1 in the sequence is processed. n .

[0077] During the loop, each time a character is processed, the window slides one position to the right according to the set step size (usually 1), and then continues to process the next character until the entire sequence is processed.

[0078] 5. Conditions for stopping calculation

[0079] When the last character x in the sequence n After processing is completed, the loop ends and the calculation stops.

[0080] At this point, each character in the entire sequence has its corresponding context representation, which can be used for subsequent decoding or other tasks.

[0081] Because the Transformer model framework is the result of multiple layers of superposition, the upper layers of the model have a broader perception field than the lower layers and can obtain more information. Therefore, it is capable of modeling a global representation that integrates all sequence information.

[0082] 3) Beam search performs result constraints:

[0083] During the decoding process, to ensure that the generated entity name belongs to the set of standard names, a beam search method is used. This method defines the constraints by using a prefix tree, where each node of the tree represents a character in the vocabulary and the child nodes represent all possible subsequent characters.

[0084] specific:

[0085] 1. Build a prefix tree: Based on a specific range, a prefix tree is constructed. The root node of the tree represents the empty character, and each node's child node represents the next character. By traversing all entity names in the standard vocabulary, each character of the entity name is added to the prefix tree one by one.

[0086] A standard vocabulary is a collection of standardized and codified terms within a specific field or discipline. It is often used to ensure accurate representation and consistent understanding of information, reducing ambiguity and misunderstanding. In medicine, standard vocabularies are particularly important, as they relate to clinical documentation, research, data sharing, and the interoperability of medical information systems.

[0087] 2. Decoding: During decoding, a beam search algorithm is used to generate candidate entity names. Initially, a blank character is used as the root node of the prefix tree. At each time step, several candidate characters are generated based on the children of the current prefix tree node. Then, based on the model's output scores, the top candidate characters are selected as the input for the next time step.

[0088] 3. Constrained filtering: When generating candidate entity names, only candidate characters that belong to nodes in the prefix tree are selected according to the definition of the prefix tree. This ensures that the generated entity names must belong to a specific range.

[0089] 4. Repeat steps 2 and 3 until a complete entity name is generated or the maximum length is reached.

[0090] By using this approach, you can ensure that the generated entity names are not made up.

[0091] To facilitate understanding of the above content, the above content is explained through the following examples:

[0092] Example 2, decoding process:

[0093] 1. Input and encoding:

[0094] Input: The encoder input is a raw word sequence, such as a text sequence containing errors or non-standard expressions.

[0095] Encoding: The encoder (usually a Transformer encoder) encodes the input sequence into a series of hidden states (context representations) that contain semantic information about the input sequence.

[0096] 2. Decoding target:

[0097] Decoding: The goal of the decoder is to gradually decode the hidden states generated by the encoder into a sequence of standard words. Here, "standard words" refer to words or phrases that conform to a predefined set of standard words.

[0098] 3. Beam search and result constraints:

[0099] Beam Search: During the decoding process, beam search is a commonly used decoding strategy used to generate multiple candidate decoding paths and select the optimal path as the final output. It maintains a fixed-size candidate set (called a "beam") and selects the most likely candidate words at each step to continue decoding.

[0100] Result Constraints: To ensure that the generated entity names fall within the set of standard names, a prefix tree (Trie) is used to define the constraints. Each node in the prefix tree represents a character in the vocabulary, and its child nodes represent all possible subsequent characters. This ensures that the characters generated at each step in the decoding process conform to the constraints of the standard vocabulary set.

[0101] Specific steps:

[0102] 1. Initialization:

[0103] 1.1 Input hidden state: The decoder receives the hidden state generated by the encoder as input.

[0104] 1.2 Bunch size: Set the bundle size B, which represents the number of candidate paths retained in each decoding step.

[0105] 1.3 Prefix Tree: Build a prefix tree where each node represents a character in the vocabulary and the child nodes represent all possible subsequent characters. The prefix tree is used to constrain the decoding process, ensuring that the generated sequence belongs to the canonical set of words.

[0106] 2. Decoding process:

[0107] 2.1 Initial candidates: Starting from the initial state of the decoder, generate a candidate set for the first character. These candidate characters must be children of the root node of the prefix tree.

[0108] 2.2 Expanding candidate paths:

[0109] At each step, for each candidate path in the current bundle, the decoder generates a set of candidates for the next character.

[0110] The selection of each candidate character must comply with the constraints of the prefix tree, that is, the current character must be a child node of the current path in the prefix tree.

[0111] Each candidate path is scored by the decoder's scoring function (based on the previous tranformer model), and the B candidate paths with the highest scores are selected as the new bundle.

[0112] 2.3 Loop decoding: Repeat the above steps of expanding the candidate path until the stopping condition is met (such as reaching the maximum decoding length, generating an end marker, etc.).

[0113] 3. Stop conditions:

[0114] 3.1 Maximum length: When the length of the decoding path reaches the preset maximum length, decoding stops.

[0115] 3.2 End Marker: When the decoding path generates an end marker (such as <eos>), stop decoding.

[0116] 3.3 Unique Path: If there is only one candidate path left in the bundle, decoding can also be stopped.

[0117] 4. Output results:

[0118] The path with the highest score is selected from the final cluster as the decoding result, that is, the generated standard word sequence.

[0119] For example:

[0120] Assume that the input sequence is "Pu Gan Min", and the standard word set is "Feng Amine" and "Chlorpheniramine".

[0121] 1. Encoding: The encoder encodes the input sequence "Push Sensitivity" into hidden states.

[0122] 2. Initialization: cluster size B = 3. Prefix tree is constructed, containing all words in the standard word set.

[0123] 3. Decoding process:

[0124] 3.1 Initial candidates: The decoder generates a candidate set for the first character, such as "phenol", "chest", "lock", etc., but only "phenol" meets the constraints of the prefix tree.

[0125] 3.2 Expanding candidate paths:

[0126] Step 1: Generate candidate paths such as "phenol-amine" and "phenol-benzene". Only "phenol-amine" meets the constraints of the prefix tree.

[0127] Step 2: Generate candidate paths such as "Phenol amino coffee" and "Phenol amino non". Only "Phenol amino coffee" meets the constraints of the prefix tree.

[0128] And so on, a path that meets the standard word set is gradually generated.

[0129] 4. Stop condition: Decoding stops when the generated path length reaches the maximum length or an end marker is generated.

[0130] 5. Output result: The path with the highest score, "Phenylamidine", is finally selected as the output.

[0131] In this way, the decoder is able to generate output sequences that conform to the standard word set, while utilizing beam search and prefix tree constraints to ensure decoding efficiency and accuracy.

[0132] 4) Integrated rag

[0133] After obtaining the trained transformer model, the RAG module is integrated.

[0134] That is, during prediction, the encoder inputs the original word, and before the decoder outputs candidate standard words, a retrieval component (calling an API interface) is added to query the external knowledge base in real time and incorporate relevant information into the final decision-making process. This retrieval component can dynamically query relevant literature or databases based on the original word input to provide more accurate standard word suggestions.

[0135] Specific methods:

[0136] Based on the existing Transformer architecture, a retrieval step is added before the Decoder outputs the candidate standard words.

[0137] Retrieval Results: When the decoder receives a raw word as input, it first uses the retrieval component to query relevant literature or databases to obtain a list of possible standard term candidates. The information from the retrieval component is weighted to influence the final standard term selection. This ensures that even for terms not seen in the original training data, external information can be used to make more accurate predictions.

[0138] Scoring and Filtering: The final step is to score all candidate criteria words and select the most suitable answers based on the scoring results. The scoring mechanism considers multiple factors, including the relevance of the search results (as returned by the API, in the range [0, 1]) and the model's own prediction probability [0, 1].

[0139] For example, when processing an original word "Pu Gan Min", the retrieval component will query the external knowledge base based on "Pu Gan Min" and return a series of possible standard words (such as "Feng Amine", "Chlorpheniramine", etc.) and their relevance scores (such as: 0.84, 0.22).

[0140] It should be noted that the scoring mechanism needs to take into account multiple factors:

[0141] 1. Relevance of search results

[0142] Relevance score returned by the API: The relevance score for each candidate standard word returned by the search component, typically a value in the interval [0, 1]. A higher relevance score indicates that the search result is more relevant to the original input word.

[0143] 2. The predicted probability of the model itself

[0144] Transformer model prediction probability: The Transformer model predicts the probability of each candidate standard word, which is also a value in the interval [0,1]. The higher the prediction probability, the more confident the model is that the candidate word is the correct standard word.

[0145] 3. Source reliability

[0146] Source of search results: Consider the source of the search results, such as whether they come from authoritative databases, literature, etc. Search results from authoritative sources can be assigned higher weights.

[0147] Comprehensive scoring formula:

[0148] S total =α·S retrieval +β·S model +γ·S source

[0149] Among them, α, β, and γ are weight coefficients, which are adjusted according to specific tasks and data.

[0150] S retrieval is the relevance score of the retrieval result, S model is the predicted probability of the model, S source is the source reliability score

[0151] In this way, the scoring mechanism can take multiple factors into consideration and more accurately select the most suitable standard words.

[0152] This information is passed to the Decoder part of the Transformer model, which refers to this additional information when generating candidate standard words.

[0153] Using the optimized attention mechanism and scoring system, the results generated internally by the model and the information retrieved are comprehensively considered to determine the final standard word mapping.

[0154] Furthermore, the pre-trained model is trained by combining the Transformer architecture with an attention optimization processing strategy and a result constraint processing strategy.

[0155] Furthermore, the result constraint processing strategy is specifically as follows:

[0156] Constructing a prefix tree, wherein a root node of the prefix tree represents a blank character, and child nodes other than the root node in the prefix tree represent characters;

[0157] In the current time step, according to the child nodes, at least one candidate character is determined and used as the input data for the next time step;

[0158] Repeat the determination of candidate characters, and stop determining candidate characters when the length of the entity name composed of the candidate characters meets a preset requirement.

[0159] Furthermore, it also includes:

[0160] The standardized processing result is displayed, and the user's result confirmation instruction is obtained through the interactive platform to complete the processing of the original word.

[0161] like Figure 2 As shown, the present invention also provides a medical terminology standardization processing system with low resource constraints based on rag retrieval, and the specific technical solution is as follows:

[0162] The acquisition unit is used to: acquire the original word of the medical term to be standardized, and input the original word into the pre-training model for processing to obtain the standardized processing result corresponding to the original word;

[0163] The pre-training model integrates a RAG module, which is used to query the information associated with the original word in real time through an external knowledge base based on the original word, and perform decision fusion processing on the information associated with the original word to generate a standardized processing result.

[0164] Furthermore, the pre-trained model is trained by combining the Transformer architecture with an attention optimization processing strategy and a result constraint processing strategy.

[0165] Furthermore, the result constraint processing strategy is specifically as follows:

[0166] Constructing a prefix tree, wherein a root node of the prefix tree represents a blank character, and child nodes other than the root node in the prefix tree represent characters;

[0167] In the current time step, according to the child nodes, at least one candidate character is determined and used as the input data for the next time step;

[0168] Repeat the determination of candidate characters, and stop determining candidate characters when the length of the entity name composed of the candidate characters meets a preset requirement.

[0169] Furthermore, it also includes:

[0170] The display unit is used to display the standardization processing result and obtain the user's result confirmation instruction through the interactive platform to complete the processing of the original word.

[0171] In the above embodiments, although the steps are numbered S1, S2, etc., these are only specific embodiments given by the present invention. Those skilled in the art may adjust the execution order of S1, S2, etc. according to actual conditions, which is also within the scope of protection of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0172] It should be noted that the beneficial effects of the medical terminology standardization processing system with low resource constraints based on rag retrieval provided by the above embodiment are the same as the beneficial effects of the medical terminology standardization processing method with low resource constraints based on rag retrieval provided by the above embodiment, and will not be repeated here. In addition, when implementing its functions, the system provided by the above embodiment only uses the division of the above functional modules as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to actual conditions to complete all or part of the functions described above. In addition, the system and method embodiments provided by the above embodiment belong to the same concept. The specific implementation process is detailed in the method embodiment, and will not be repeated here.

[0173] like Figure 3 As shown, an electronic device 300 according to an embodiment of the present invention includes a processor 320, which is coupled to a memory 310. The memory 310 stores at least one computer program 330. The at least one computer program 330 is loaded and executed by the processor 320 to enable the electronic device 300 to implement any of the above methods. Specifically:

[0174] The electronic device 300 may vary significantly due to different configurations or performance, and may include one or more processors 320 (Central Processing Units, CPUs) and one or more memories 310, wherein the one or more memories 310 store at least one computer program 330, which is loaded and executed by the one or more processors 320 to enable the electronic device 300 to implement the low-resource-constrained medical terminology standardization processing method based on rag retrieval provided in the above embodiment. Of course, the electronic device 300 may also have components such as a wired or wireless network interface, a keyboard, and an input / output interface for input and output. The electronic device 300 may also include other components for implementing device functions, which will not be detailed here.

[0175] A computer-readable storage medium according to an embodiment of the present invention stores at least one computer program, and the at least one computer program is loaded and executed by a processor to enable a computer to implement any of the above methods.

[0176] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, or the like.

[0177] In an exemplary embodiment, a computer program product or computer program is also provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform any of the above methods.

[0178] It should be noted that the terms "first," "second," and so on, in the specification and claims of this application are used to distinguish similar objects and to define a specific order or precedence. Where appropriate, the order used for similar objects can be interchanged, such that the embodiments of the present application described herein can be implemented in an order other than the order shown or described.

[0179] Those skilled in the art will appreciate that the present invention may be implemented as a system, method, or computer program product. Therefore, the present disclosure may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the present invention may be implemented as a computer program product embodied in one or more computer-readable media containing computer-readable program code.

[0180] Any combination of one or more computer-readable media can be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or device.

[0181] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.< / eos>

Claims

1. A low-resource-constrained medical terminology standardization method based on rag retrieval, characterized in that: include: Obtaining the original words of the medical terms to be standardized, and inputting the original words into the pre-trained model for processing to obtain the standardized processing results corresponding to the original words; The pre-training model integrates a RAG module, which is used to query the information associated with the original word in real time through an external knowledge base based on the original word, and perform decision fusion processing on the information associated with the original word to generate a standardized processing result.

2. A low-resource-constrained medical term standardization method based on rag retrieval according to claim 1, characterized in that: The pre-trained model is trained by combining the Transformer architecture with an attention optimization processing strategy and a result constraint processing strategy.

3. The method for standardizing medical terms with low resource constraints based on rag retrieval according to claim 2, characterized in that: The specific result constraint processing strategy is: Constructing a prefix tree, wherein a root node of the prefix tree represents a blank character, and child nodes other than the root node in the prefix tree represent characters; In the current time step, according to the child nodes, at least one candidate character is determined and used as the input data for the next time step; Repeat the determination of candidate characters, and stop determining candidate characters when the length of the entity name composed of the candidate characters meets the preset requirement.

4. The method for standardizing medical terminology with low resource constraints based on rag retrieval according to claim 1, characterized in that: Also includes: The standardized processing result is displayed, and the user's result confirmation instruction is obtained through the interactive platform to complete the processing of the original word.

5. A low-resource-constrained medical terminology standardization processing system based on rag retrieval, characterized by: include: The acquisition unit is used to: acquire the original word of the medical term to be standardized, and input the original word into the pre-training model for processing to obtain the standardized processing result corresponding to the original word; The pre-training model integrates a RAG module, which is used to query the information associated with the original word in real time through an external knowledge base based on the original word, and perform decision fusion processing on the information associated with the original word to generate a standardized processing result.

6. A medical term standardization processing system with low resource constraints based on rag retrieval according to claim 5, characterized in that: The pre-trained model is trained by combining the Transformer architecture with an attention optimization processing strategy and a result constraint processing strategy.

7. The low resource constraint medical term standardization processing system based on rag retrieval according to claim 6 is characterized in that: The specific result constraint processing strategy is: Constructing a prefix tree, wherein a root node of the prefix tree represents a blank character, and child nodes other than the root node in the prefix tree represent characters; In the current time step, according to the child nodes, at least one candidate character is determined and used as the input data for the next time step; Repeat the determination of candidate characters, and stop determining candidate characters when the length of the entity name composed of the candidate characters meets the preset requirement.

8. The low-resource-constrained medical terminology standardization processing system based on rag retrieval according to claim 5 is characterized in that: Also includes: The display unit is used to display the standardization processing result and obtain the user's result confirmation instruction through the interactive platform to complete the processing of the original word.

9. An electronic device, characterized in that: The electronic device includes a processor coupled to a memory, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor so that the electronic device implements the method according to any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable a computer to implement the method according to any one of claims 1 to 4.

Images