A method and device for extracting and standardizing discharge medical order information
By optimizing the generative model through alternating reinforcement training and multi-turn dialogue mechanism, the efficiency and accuracy problems of traditional NER models in the extraction and standardization of discharge medical order information are solved, and efficient and accurate information extraction and standardization processing are achieved.
Patent Information
- Application Number
- CN202411067456.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-06
AI Technical Summary
Traditional NER models suffer from problems such as complex information standardization, text length limitations, low efficiency of multi-model stacking, and errors in semantic similarity standardization when extracting discharge instructions from medical texts. Furthermore, generative models have low extraction accuracy and insufficient standardization.
An alternating augmentation training method is adopted, combining extraction and generative models. Initial entity terms are labeled with seed terms, and the model is iteratively trained and standardized using a multi-turn dialogue mechanism to construct a loss function to optimize the generative model.
It improves the extraction accuracy and standardization efficiency of generative models, reduces human intervention, makes full use of contextual information, is suitable for long text processing, and simplifies post-processing steps.
Smart Images

Figure CN118586404B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of natural language processing, and in particular relates to a method and device for extracting and standardizing discharge medical order information. Background Art
[0002] In the field of medical information processing, especially for extracting information about discharge orders from electronic medical records, a common method is to use the traditional named entity recognition (NER) model. The traditional NER model recognizes and extracts key information such as the time, location, and content of the follow-up visit from the text to achieve structured processing of discharge orders. However, the traditional NER model has the following shortcomings:
[0003] First, traditional NER models mostly use an extractive approach, that is, extracting relevant information directly from the text. However, the terms in medical texts are often not standardized, and the extracted information needs to be further standardized in medical terms, which increases the complexity of subsequent processing. Taking the extraction of medical order actions as an example, after the traditional method extracts the medical order content such as "recheck blood routine test in 3 days", it is necessary to build additional time and medical order action content standardization classifiers, convert "3 days later" into a standard date format, and convert "blood routine test" into standard medical test terms, in order to obtain structured medical order information. Secondly, some key information may not appear directly in the text, but requires a certain reasoning process. For example, the medical order mentions "3 years ago", but the specific date information in the "year-month-day" format is expected. It is difficult for traditional NER models to directly implement this time reasoning. For example, when it is necessary to determine whether a symptom is mentioned in the medical order, the traditional NER model needs to build an additional classifier, and cannot directly output the "yes / no" label information. In addition, the text length that traditional NER models can process is usually limited to 256-512 characters, while medical documents, especially hospitalization and discharge records, often have a text length of more than 1,000 characters, which also poses challenges to the application of traditional methods.
[0004] In view of the technical defects of the above-mentioned extractive models, after extracting various entity terms from medical texts, another model is needed to standardize the entity terms and convert them into standard entity terms to facilitate subsequent query, tracking and analysis. However, this method has some technical problems: first, multiple models need to be stacked to complete the task for entity term extraction and standardization, which is inefficient; second, when standardizing terms, the context of the original text is often ignored, and errors are easily made when standardizing the extracted entity terms alone. In addition, existing work usually only uses word vector models (such as BERT) to calculate the semantic similarity between the terms to be standardized and the standard terms, and selects the standard terms with the highest similarity as the standardization result. Relying solely on semantic similarity for standardization often leads to errors.
[0005] In the field of medical information processing, generative models are also used for text extraction and standardization. Although the generative model will not miss entity terms when extracting text like a separate NER model, the accuracy of extracting entity terms is low and the degree of standardization is low, which cannot meet application requirements. Summary of the invention
[0006] In view of the above, the purpose of the present invention is to provide a method for extracting and standardizing discharge medical order information, to improve the accuracy of entity terminology extracted by the generative model through alternating reinforcement training, and to improve the accuracy and efficiency of the standardized processing of the generative model through multi-round dialogue mechanism training.
[0007] To achieve the above-mentioned purpose of the invention, an embodiment provides a method for extracting and standardizing discharge medical order information, comprising the following steps:
[0008] Based on the seed terms, the initial entity terms are matched and annotated from the discharge instructions text as extractive samples;
[0009] A training framework including extractive and generative models was constructed, and the following extraction tasks were iteratively performed in an alternating reinforcement training process: the extractive model was trained with extractive samples, and entity terms were extracted from the medical discharge order text using the trained extractive model as generative samples; the generative model was trained with generative samples, and entity terms were extracted from the medical discharge order text using the trained generative model as the next round of extractive samples;
[0010] Based on a multi-round dialogue mechanism, the generative model that has been trained through alternating reinforcement is trained through multi-round dialogue for standardization tasks, including: using the generative model to extract entity terms to be standardized from the discharge order text, and matching candidate standard entity terms for the entity terms to be standardized, inputting the entity terms to be standardized, the candidate standard entity terms, and the discharge order text into the generative model for standardization to obtain standardized entity terms, and constructing a loss function to train the generative model;
[0011] A generative model trained with multi-round dialogue enhancement is used to extract and standardize discharge order information using a multi-round dialogue approach.
[0012] Preferably, the matching and labeling of initial entity terms from the discharge medical order text as extractive samples based on seed terms includes:
[0013] Based on the existing seed terms in the medical term dictionary, entities matching the seed terms are found in the discharge instructions text through keyword matching and marked as initial entity terms, and the initial entity terms are used as extraction samples.
[0014] Preferably, the extractive model adopts a Med-BERT model, which uses the BERT model as a text encoder and adds a token classification layer on this basis to perform entity classification on each token to achieve noun entity recognition and extraction.
[0015] Preferably, the generative model includes a Tranformer Decoder structure.
[0016] Preferably, matching the to-be-standardized entity term with the candidate standard entity term includes:
[0017] The similarity between the entity terms to be standardized and the terms in the standard term library is calculated, and the top K terms with high similarity are selected as matching candidate standard entity terms. When calculating the similarity, the BERT model term encoding is converted into a semantic vector, and the similarity between the two semantic vectors is calculated based on the semantic vector.
[0018] Preferably, the constructing a loss function to train a generative model comprises:
[0019] A cross entropy loss is constructed based on the extracted entity terms to be standardized and the true values of the entity terms to be standardized. At the same time, a cross entropy loss is constructed based on the generated standardized entity terms and the true values of the standardized entity terms. The generative model is trained based on the two cross entropy losses.
[0020] Preferably, the method also includes: when performing enhanced training based on a multi-round dialogue mechanism, adding unique token characters for identification to the entity terms to be standardized, candidate standard entity terms, discharge order texts, and standardized entity terms, so as to accurately select the entity terms to be standardized and the standardized entity terms when constructing the loss function.
[0021] To achieve the above-mentioned purpose of the invention, the embodiment further provides a device for extracting and standardizing discharge medical advice information, including:
[0022] An initial entity term extraction module, which is used to match and annotate initial entity terms from discharge medical order texts as extraction samples based on seed terms;
[0023] An alternating reinforcement training module is used to construct an alternating reinforcement training framework including an extractive model and a generative model, and iteratively perform the following reinforcement training process to implement the extraction task: using the extractive samples to train the extractive model, and using the trained extractive model to extract entity terms from the discharge medical order text as generative samples, using the generative samples to train the generative model, and using the trained generative model to extract entity terms from the discharge medical order text as the next round of extractive samples;
[0024] A multi-round dialogue enhancement training module is used to perform multi-round dialogue enhancement training of a generative model that has undergone alternating enhancement training on a standardization task based on a multi-round dialogue mechanism, including: using a generative model to extract entity terms to be standardized from a discharge order text, and matching candidate standard entity terms for the entity terms to be standardized, inputting the entity terms to be standardized, the candidate standard entity terms, and the discharge order text into the generative model for standardization processing to obtain standardized entity terms, and constructing a loss function to train the generative model;
[0025] An application reasoning module is used to extract and standardize discharge order information using a multi-round dialogue method using a generative model that has been trained through multi-round dialogue enhancement.
[0026] To achieve the above-mentioned purpose of the invention, an embodiment further provides a computing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned discharge medical order information extraction and standardization method.
[0027] In order to achieve the above-mentioned purpose of the invention, the embodiment also provides a computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, the above-mentioned discharge medical order information extraction and standardization method is implemented.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] Initial entity terms are constructed based on seed terms. This process is completely dependent on seed terms and requires little or no manual intervention, which improves the accuracy of initial entity terms and the extraction efficiency.
[0030] Based on the training framework that includes extractive and generative models, enhanced training for extraction tasks can effectively improve the accuracy and recall of generative models without time-consuming and labor-intensive manual annotation, thereby achieving more comprehensive and accurate information extraction. This method fully utilizes the advantages of seed terms and alternating model training, greatly reduces the reliance on manual annotation, and improves the efficiency of data preparation.
[0031] Multi-round dialogue enhanced training based on a multi-round dialogue mechanism implements the extraction process and standardization through multi-round dialogues based on a generative model, without the need for multiple models. It is highly efficient and combines contextual information to improve the accuracy of standardization. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0033] Figure 1 is a flow chart of a method for extracting and standardizing discharge medical advice information provided by an embodiment;
[0034] Figure 2 is a flow chart of alternating reinforcement training based on a training framework provided by an embodiment;
[0035] Figure 3 is a flow chart of multi-round dialogue enhancement training provided by an embodiment;
[0036] Figure 4 is a schematic diagram of the structure of a device for extracting and standardizing discharge medical advice information provided by an embodiment;
[0037] Figure 5 It is a schematic diagram of the structure of a computing device provided in an embodiment. DETAILED DESCRIPTION
[0038] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.
[0039] like Figure 1 As shown, the embodiment provides a method for extracting and standardizing discharge medical advice information, comprising the following steps:
[0040] S1, matching and annotating initial entity terms from discharge order texts based on seed terms as extractive samples.
[0041] In the embodiment, based on the seed terms in the medical term dictionary, through keyword matching, entities matching the seed terms are found in the discharge order text as initial entity terms, and the initial entity terms are used as extraction samples. Among them, the matching entities include medical institutions, tests, examinations, medications, and time names, etc. This process is completely based on seed terms and almost no manual intervention such as manual annotation is required. In special cases, only a small amount of manual annotation is required to filter the labeling results, which greatly improves the extraction efficiency.
[0042] S2, build a training framework that includes an extractive model and a generative model, and perform alternating reinforcement training on the extraction task based on the training framework.
[0043] In the embodiment, based on the initial entity terms, cyclic iterative enhanced data collection is performed based on the initial entity terms, and the accuracy and recall rate of the generative model are continuously improved by alternating training of extractive models such as NER models and generative models.
[0044] In the training framework, the NER model uses the Med-NER model. The generative model contains a Tranformer Decoder structure, such as qwen0.5b, Qwen 1.8b, etc. When using this automatically generated initial entity term to train the NER model, although the accuracy of extractive models such as the NER model is very high, the recall rate is low. When using the initial entity term to train the generative model, although the recall rate of the generative model is high, that is, it will not miss information, the accuracy rate is low and does not meet the application requirements. In order to make up for the shortcomings of the NER model in recall rate and the shortcomings of the generative model in accuracy, the alternating reinforcement training method is adopted.
[0045] like Figure 2 As shown, the alternating reinforcement training process of the extraction task is iterated: the extractive samples are used to train the extractive model, and the trained extractive model is used to extract entity terms from the discharge order text as generative samples, the generative samples are used to train the generative model, and the trained generative model is used to extract entity terms from the discharge order text as the next round of extractive samples. It should be noted that the extractive samples used in the first round of training come from the initial entity terms. It should also be noted that the method used to train the NER model is a conventional method, and the method used to train the generative model is also a conventional method, without limitation.
[0046] By training the extractive model and the generative model alternately in a cycle, the performance of both models can be continuously improved. In each round of iteration, the extraction results of the extractive model are used to train the generative model, and the extraction results of the generative model are used to train the extractive model. In this process, a small amount of manual annotation can be used to filter the results of automatic labeling, further improving data quality, and the trained generative model has higher extraction accuracy.
[0047] S3, based on the multi-round dialogue mechanism, performs multi-round dialogue enhancement training on standardized tasks on the generative model after alternating reinforcement training.
[0048] In the embodiment, based on step S2, a multi-round dialogue mechanism is proposed to extract and standardize entity terms. Before specific application, multi-round dialogue enhancement training is performed.
[0049] like Figure 3 As shown in the figure, the multi-round dialogue enhancement training process includes:
[0050] The generative model trained with alternating reinforcement is used to extract the medical institutions, tests, examinations, medications, and time names of follow-up visits and reexaminations from the discharge instructions as the entity terms to be standardized in the first round of dialogue.
[0051] To match the entity terms to be standardized with the candidate standard entity terms, specifically, the similarity between the entity terms to be standardized and the terms in the standard term library is calculated, and the top K terms with high similarity are screened as matching candidate standard entity terms. When calculating the similarity, the BERT model is used to convert the term encoding into a semantic vector, and the similarity between the two semantic vectors is calculated based on the semantic vector.
[0052] Input the entity terms to be standardized, the candidate standard entity terms, and the discharge instructions text as the context into the generative model after alternating reinforcement training for standardization to obtain the standardized entity terms as the second round of dialogue;
[0053] A loss function is constructed to train the generative model. Specifically, a cross entropy loss is constructed based on the extracted entity terms to be standardized and the true values of the entity terms to be standardized, and a cross entropy loss is constructed based on the generated standardized entity terms and the true values of the standardized entity terms. The generative model is trained based on the two cross entropy losses.
[0054] During specific training, the generative model involves four categories of terms: discharge instructions, entity terms to be standardized, candidate standard entity terms, and standardized entity terms. In order to make the model better distinguish these categories of terms, special unique token characters used for identification are used to segment the terms. The example of a special unique token character segmentation result is:
[0055] <raw_input> Discharge instructions text< / raw_input> ;
[0056] <raw_output> Entity terms to be standardized< / raw_output> ;
[0057] <candidates> K candidate standard entity terms< / candidates> ;
[0058] <output> Standardized entity terminology< / output> .
[0059] During specific training, the above text and various terms are input into the generative model.<raw_input> Discharge instructions text< / raw_input> and <candidates> K candidate standard entity terms< / candidates> The tokens between are masked so that the generative model can better distinguish different parts of the input data, improving the performance of the generative model to avoid participating in the loss calculation.<raw_output> Entity terms to be standardized< / raw_output> and <output> Standardized entity terminology< / output> Parameter loss function calculation, specifically, each predicted character in the entity term to be standardized generated by the generative model in the first round of dialogue is cross-entropy calculated with each character in the true value of the entity term to be standardized to construct a cross-entropy loss. Each predicted character in the standardized entity term generated by the generative model in the second round of dialogue is cross-entropy calculated with each character in the true value of the standardized entity term to construct a cross-entropy loss. Then, the generative model is trained based on the two cross-entropy losses. The cross-entropy loss is used to optimize the generative model, which can better learn the standardized results.
[0060] In the above training process, the problems in the traditional method are effectively solved by integrating the process, utilizing contextual information, optimizing the loss function, and introducing special tokens, thereby improving the efficiency and accuracy of term standardization.
[0061] S4, using the generative model trained with multi-round dialogue enhancement, extracts and standardizes discharge order information using a multi-round dialogue approach.
[0062] In the embodiment, when applying reasoning, the first round of dialogue is performed from the input of the discharge order text into the generative model after multiple rounds of dialogue enhancement training to obtain the entity terms to be standardized, and the entity terms to be standardized are matched with the standard terminology library to obtain K candidate standard entity terms, and then the entity terms to be standardized, the candidate standard entity terms, and the discharge order text are input into the generative model together for a second round of dialogue to achieve standardization processing and obtain standardized entity terms.
[0063] In the above reasoning process, the extraction and standardization processes are integrated into a generative model, which improves the overall extraction efficiency. By splicing together the original discharge order text, the entity terms to be standardized in the first round of dialogue, and the candidate standard entity terms, the generative model can make full use of contextual information and improve the accuracy of standardization. The generated results have been standardized and no further complex post-processing steps are required. The extracted structured information can be directly stored in the knowledge base for subsequent patient follow-up, medical data analysis and other applications.
[0064] like Figure 4 As shown, the embodiment also provides a discharge medical order information extraction and standardization device 40, including an initial entity term extraction module 41, an alternating reinforcement training module 42, a multi-round dialogue enhancement training module 43, and an application reasoning module 44, wherein the initial entity term extraction module 41 is used to match and annotate initial entity terms from the discharge medical order text as extraction samples based on seed terms; the alternating reinforcement training module 42 is used to construct a training framework including an extraction model and a generative model, and perform alternating reinforcement training of the extraction task based on the training framework; the multi-round dialogue enhancement training module 43 is used to perform multi-round dialogue enhancement training of the standardization task on the generative model after the alternating reinforcement training based on the multi-round dialogue mechanism; the application reasoning module 44 is used to use the generative model after the multi-round dialogue enhancement training to adopt a multi-round dialogue method to extract and standardize the discharge medical order information.
[0065] It should be noted that the device for extracting and standardizing discharge medical order information provided in the above-mentioned embodiment should be illustrated by the division of the above-mentioned functional modules when extracting and standardizing discharge medical order information. The above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the terminal or server is divided into different functional modules to complete all or part of the functions described above. In addition, the device for extracting and standardizing discharge medical order information provided in the above-mentioned embodiment belongs to the same concept as the embodiment of the method for extracting and standardizing discharge medical order information. The specific implementation process is detailed in the embodiment of the method for extracting and standardizing discharge medical order information, which will not be repeated here.
[0066] The embodiment further provides a computing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the device is used to implement the above-mentioned discharge medical order information extraction and standardization method, which specifically includes the following steps:
[0067] S1, matching and annotating initial entity terms from discharge instructions text based on seed terms as extractive samples;
[0068] S2, build a training framework that includes an extractive model and a generative model, and perform alternating reinforcement training on the extraction task based on the training framework;
[0069] S3, based on the multi-round dialogue mechanism, conducts multi-round dialogue enhancement training on standardized tasks for the generative model after alternating reinforcement training;
[0070] S4, using the generative model trained with multi-round dialogue enhancement, extracts and standardizes discharge order information using a multi-round dialogue approach.
[0071] like Figure 5 As shown, the computing device provided in the embodiment, at the hardware level, includes not only a processor and a memory, but also hardware required for other services such as an internal bus, a network interface, and a memory. The memory is a non-volatile memory, and the processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above-mentioned S1-S4 methods for extracting and standardizing discharge medical order information. Of course, in addition to software implementations, the present invention does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.
[0072] The embodiment further provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the above-mentioned method for extracting and standardizing discharge medical order information is implemented, which specifically includes the following steps:
[0073] S1, matching and annotating initial entity terms from discharge instructions text based on seed terms as extractive samples;
[0074] S2, build a training framework that includes an extractive model and a generative model, and perform alternating reinforcement training on the extraction task based on the training framework;
[0075] S3, based on the multi-round dialogue mechanism, conducts multi-round dialogue enhancement training on standardized tasks for the generative model after alternating reinforcement training;
[0076] S4, using the generative model trained with multi-round dialogue enhancement, extracts and standardizes discharge order information using a multi-round dialogue approach.
[0077] In the embodiment, computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data.
[0078] In summary, the generative model used in the present invention has the following advantages: (1) There is no need to build additional time parsers and medical terminology standardization modules. By standardizing the training data, the model can directly generate standardized extraction results; (2) The powerful language understanding and generation capabilities of the large language model enable it to reason based on the context and extract implicit information, such as converting relative time expressions into absolute dates; (3) The pre-trained large language model has excellent generalization capabilities and can process texts up to 4096 characters in length, making it suitable for information extraction tasks in most discharge medical order documents.
[0079] The scheme for extracting and standardizing discharge medical order information provided by the present invention makes full use of the latest progress of large language models in the field of natural language processing, simplifies the post-processing links in traditional methods, and improves the accuracy and efficiency of information extraction. The method can be seamlessly integrated with the information system of medical institutions, supports intelligent patient follow-up management and post-hospital services, and has broad application prospects.
[0080] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for extracting and standardizing discharge medical advice information, characterized in that: The following steps are involved: Based on the seed terms, the initial entity terms are matched and annotated from the discharge instructions text as extractive samples; A training framework including extractive and generative models was constructed, and the following extraction tasks were iteratively performed in an alternating reinforcement training process: the extractive model was trained with extractive samples, and entity terms were extracted from the medical discharge order text using the trained extractive model as generative samples; the generative model was trained with generative samples, and entity terms were extracted from the medical discharge order text using the trained generative model as the next round of extractive samples; Based on a multi-round dialogue mechanism, the generative model that has been trained through alternating reinforcement is trained through multi-round dialogue for standardization tasks, including: using the generative model to extract entity terms to be standardized from the discharge order text, and matching candidate standard entity terms for the entity terms to be standardized, inputting the entity terms to be standardized, the candidate standard entity terms, and the discharge order text into the generative model for standardization to obtain standardized entity terms, and constructing a loss function to train the generative model; Among them, when multi-round dialogue enhancement training is performed based on a multi-round dialogue mechanism, unique token characters for identification are added to the entity terms to be standardized, the candidate standard entity terms, the discharge order text, and the standardized entity terms, so as to accurately select the entity terms to be standardized and the standardized entity terms when constructing the loss function, and the cross entropy loss is constructed based on the extracted entity terms to be standardized and the true values of the entity terms to be standardized. At the same time, the cross entropy loss is constructed based on the generated standardized entity terms and the true values of the standardized entity terms, and the generative model is trained based on the two cross entropy losses; A generative model trained with multi-round dialogue enhancement is used to extract and standardize discharge order information using a multi-round dialogue approach.
2. The method for extracting and standardizing discharge medical order information according to claim 1, characterized in that: The method of matching and labeling initial entity terms from the discharge medical order text based on seed terms as extractive samples includes: Based on the existing seed terms in the medical term dictionary, entities matching the seed terms are found in the discharge instructions text through keyword matching and marked as initial entity terms, and the initial entity terms are used as extraction samples.
3. The method for extracting and standardizing discharge medical order information according to claim 1, characterized in that: The extractive model adopts the Med-BERT model, which uses the BERT model as a text encoder and adds a token classification layer on this basis to perform entity classification on each token to achieve noun entity recognition and extraction.
4. The method for extracting and standardizing discharge medical order information according to claim 1, characterized in that: The generative model includes a Tranformer Decoder structure.
5. The method for extracting and standardizing discharge medical order information according to claim 1, characterized in that: The candidate standard entity terms are matched to the entity terms to be standardized, including: The similarity between the entity terms to be standardized and the terms in the standard term library is calculated, and the top K terms with high similarity are selected as matching candidate standard entity terms. When calculating the similarity, the BERT model is used to convert the term encoding into a semantic vector, and the similarity between the two semantic vectors is calculated based on the semantic vector.
6. A device for extracting and standardizing discharge medical advice information, characterized in that: include: An initial entity term extraction module, which is used to match and annotate initial entity terms from discharge medical order texts as extraction samples based on seed terms; An alternating reinforcement training module is used to construct a training framework including an extractive model and a generative model, and iteratively perform the alternating reinforcement training process of the following extraction tasks: using the extractive samples to train the extractive model, and using the trained extractive model to extract entity terms from the discharge medical order text as generative samples, using the generative samples to train the generative model, and using the trained generative model to extract entity terms from the discharge medical order text as the next round of extractive samples; Among them, when multi-round dialogue enhancement training is performed based on a multi-round dialogue mechanism, unique token characters for identification are added to the entity terms to be standardized, the candidate standard entity terms, the discharge order text, and the standardized entity terms, so as to accurately select the entity terms to be standardized and the standardized entity terms when constructing the loss function, and the cross entropy loss is constructed based on the extracted entity terms to be standardized and the true values of the entity terms to be standardized. At the same time, the cross entropy loss is constructed based on the generated standardized entity terms and the true values of the standardized entity terms, and the generative model is trained based on the two cross entropy losses; A multi-round dialogue enhancement training module is used to perform multi-round dialogue enhancement training of a generative model that has undergone alternating enhancement training on a standardization task based on a multi-round dialogue mechanism, including: using a generative model to extract entity terms to be standardized from a discharge order text, and matching candidate standard entity terms for the entity terms to be standardized, inputting the entity terms to be standardized, the candidate standard entity terms, and the discharge order text into the generative model for standardization processing to obtain standardized entity terms, and constructing a loss function to train the generative model; An application reasoning module is used to extract and standardize discharge order information using a multi-round dialogue method using a generative model that has been trained through multi-round dialogue enhancement.
7. A computing device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the one or more processors execute the executable code, they are used to implement the method for extracting and standardizing discharge medical order information according to any one of claims 1-5.
8. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the method for extracting and standardizing discharge medical order information described in any one of claims 1-5 is implemented.
