Data enhancement method, device and equipment, computer readable storage medium and computer program product
By training a target large language model and combining it with an interactive model, the problem of the lack of professionalism and relevance of first aid guidance knowledge in existing technologies has been solved, generating more professional and diversified first aid guidance information and improving citizens' first aid capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA MOBILE CHENGDU INFORMATION & TELECOMM TECH CO LTD
- Filing Date
- 2024-10-22
- Publication Date
- 2026-04-24
AI Technical Summary
The first aid guidance knowledge generated by existing technologies lacks professionalism and relevance, and cannot effectively improve citizens' first aid capabilities.
The initial large language model is trained based on sample data from the target domain to generate a target large language model. The similarity between candidate guidance information and target guidance information is determined by combining interactive models, thereby selecting professional and diverse augmented data.
The generated first aid guidance information is more professional and relevant, improving citizens' first aid knowledge and self-rescue capabilities.
Smart Images
Figure CN121922397A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, and in particular to a data augmentation method, apparatus, device, computer-readable storage medium, and computer program product. Background Technology
[0002] Currently, pre-hospital emergency care in China is relatively underdeveloped. Citizens have a weak willingness to provide first aid, low awareness of first aid, and a serious lack of first aid knowledge and skills. In response to this situation, it is essential to strengthen the dissemination of first aid knowledge, improve the first aid capabilities of the public, and ensure that citizens can rationally apply first aid measures for self-rescue before emergency personnel arrive, thus buying more time for subsequent pre-hospital emergency care.
[0003] Existing technologies typically generate data directly based on first aid guidance knowledge and large language models. This involves constructing prompts from first aid guidance knowledge and inputting these prompts into a large language model to generate new first aid guidance knowledge. However, the new first aid guidance knowledge generated directly from large language models in existing technologies cannot guarantee the professionalism and relevance of the data. Summary of the Invention
[0004] To address the aforementioned technical problems, this application aims to provide a data enhancement method, apparatus, device, computer-readable storage medium, and computer program product that can solve the problem of low professionalism and relevance of generated data in related technologies.
[0005] The technical solution of this application is implemented as follows:
[0006] A data augmentation method, the method comprising:
[0007] Based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model, multiple candidate guidance information are determined; wherein, the target large language model is obtained by training an initial large language model using sample data from the target domain;
[0008] Based on the candidate guidance information, the target guidance information, and the interactive model, the similarity between the candidate guidance information and the target guidance information is determined;
[0009] Based on the similarity, target candidate guidance information is determined from multiple candidate guidance information, and the target candidate guidance information is determined to be augmented data of the target guidance information.
[0010] In the above scheme, before determining multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patient, and the target large language model, the following steps are included:
[0011] Acquire sample data in the medical field that has a multimodal form, and process the sample data to obtain target sample data; wherein, the target field includes the medical field;
[0012] The target large language model is obtained by training the initial large language model using the target sample data.
[0013] In the above scheme, determining multiple candidate guidance information based on the current description information of the target patient, the target guidance information, and the target large language model includes:
[0014] Based on the current description information of the target patient and the target guidance information, prompt words are generated;
[0015] Based on the prompt words, the target patient's historical description information, and the target large language model, multiple candidate guidance messages are generated.
[0016] In the above scheme, before determining multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patient, and the target large language model, the following steps are also included:
[0017] Based on the sample description information of the sample patients in different scenarios and the initial guidance information for the sample patients, multiple target guidance information is generated;
[0018] Multiple sets of target guidance information are stored in a guidance library, and the target guidance information in the guidance library is corrected to update the guidance library.
[0019] In the above scheme, determining the similarity between the candidate guidance information and the target guidance information based on the candidate guidance information, the target guidance information, and the interactive model includes:
[0020] Multiple candidate guidance information and target guidance information are concatenated and serialized to obtain multiple processed guidance information;
[0021] Based on the processed guidance information and the interactive model, the similarity between each candidate guidance information and the target guidance information is determined.
[0022] In the above scheme, determining the similarity between each candidate guidance information and the target guidance information based on multiple processed guidance information and the interactive model includes:
[0023] A coding layer is used to encode multiple processed guidance information to obtain first feature information corresponding to each processed guidance information;
[0024] A pooling layer is used to perform pooling processing on each of the first feature information to obtain the second feature information;
[0025] Based on the output layer and each of the second feature information, the similarity between each candidate guidance information and the target guidance information is obtained; wherein, the interactive model includes the encoding layer, the pooling layer and the output layer.
[0026] A data augmentation apparatus, the apparatus comprising:
[0027] The first processing unit is used to determine multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model; wherein, the target large language model is obtained by training an initial large language model using sample data from the target domain;
[0028] The second processing unit is used to determine the similarity between the candidate guidance information and the target guidance information based on the candidate guidance information, the target guidance information, and the interactive model;
[0029] The determining unit is configured to determine target candidate guidance information from multiple candidate guidance information based on the similarity, and determine the target candidate guidance information as enhanced data of the target guidance information.
[0030] A data enhancement device, the device comprising: a processor, a memory, and a communication bus;
[0031] The communication bus is used to realize the communication connection between the processor and the memory;
[0032] The processor is used to execute the data augmentation program stored in the memory to implement the steps of the above-described data augmentation method.
[0033] A computer-readable storage medium storing one or more programs that can be executed by one or more processors to perform the steps of the data augmentation method described above.
[0034] A computer program product includes a computer program that, when executed by a processor, implements the steps of the data augmentation method described above.
[0035] The data augmentation method, apparatus, device, computer-readable storage medium, and computer program product provided in this application firstly determine multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patient, and the target large language model. The target large language model is obtained by training an initial large language model using sample data from the target domain. Then, based on the candidate guidance information, the target guidance information, and the interactive model, the similarity between the candidate guidance information and the target guidance information is determined. Subsequently, based on the similarity, the target candidate guidance information is determined from the multiple candidate guidance information, and the target candidate guidance information is determined as the augmented data for the target guidance information. In this way, by using sample data from the target domain to learn professional domain knowledge from the initial large language model, the target large language model gains more professional domain knowledge capabilities, thereby generating more professional and diverse candidate guidance information. Furthermore, based on the professional and diverse candidate guidance information and the target guidance information, deep matching is performed in conjunction with the interactive model, thereby ensuring the relevance and professionalism of the final generated augmented knowledge. Attached Figure Description
[0036] Figure 1 A flowchart illustrating a data augmentation method provided in an embodiment of this application;
[0037] Figure 2 A flowchart illustrating yet another data augmentation method provided in an embodiment of this application;
[0038] Figure 3 This is a schematic diagram illustrating the determination of augmented data in a data augmentation method provided in an embodiment of this application;
[0039] Figure 4 This is a schematic diagram illustrating the acquisition of target sample data in a data augmentation method provided in an embodiment of this application;
[0040] Figure 5 A schematic diagram illustrating the training of an initial large language model in a data augmentation method provided in an embodiment of this application;
[0041] Figure 6 A schematic diagram of the interaction model in a data augmentation method provided in an embodiment of this application;
[0042] Figure 7 This is a schematic diagram of the structure of a data enhancement device provided in an embodiment of this application;
[0043] Figure 8 This is a schematic diagram of the structure of a data enhancement device provided in an embodiment of this application. Detailed Implementation
[0044] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.
[0045] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.
[0046] It's important to note that existing first aid knowledge generation schemes generally fall into two main categories: one is a self-instructive approach directly based on a Large Language Model (LLM). This involves constructing task prompts and then having the LLM generate new first aid guidance knowledge. The LLM then evaluates and filters the generated knowledge, relying entirely on its capabilities. The other approach directly generates new first aid guidance knowledge based on existing first aid knowledge and the LLM. This method combines first aid knowledge with a general LLM, constructs prompts, and allows the LLM to generate new first aid guidance knowledge.
[0047] However, the self-instructive approach heavily relies on the capabilities of third-party large language models. Both generation and discrimination depend entirely on these models, leading to data bias due to this reliance. Furthermore, general-domain models may lack specialized knowledge in specific domains, potentially resulting in issues with the quality and professionalism of the generated first-aid guidance. While methods based on large language models and direct generation of first-aid guidance knowledge can alleviate the problems of the self-instructive approach by incorporating first-aid guidance knowledge, they cannot guarantee professionalism and diversity.
[0048] Based on this, embodiments of this application provide a data augmentation method, which can be applied to a data augmentation device, as described above. Figure 1 As shown, the method includes the following steps:
[0049] Step 101: Based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model, determine multiple candidate guidance information.
[0050] The target large language model is obtained by training the initial large language model using sample data from the target domain.
[0051] In this embodiment, the target patient can refer to a patient for whom guidance information needs to be determined; the current description information of the target patient can include the patient's current chief complaint, consciousness, and respiration; the target guidance information can refer to expert emergency care guidance knowledge; the candidate guidance information can refer to the guidance information output by the target large language model, and the candidate guidance information can specifically refer to candidate emergency care guidance knowledge; the large language model can be a generative pre-trained Transformer (GPT) model; the description information of the target patient and the target guidance information can be processed first, and then multiple candidate guidance information can be obtained based on the processing results and the target large language model.
[0052] Step 102: Based on the candidate guidance information, the target guidance information, and the interactive model, determine the similarity between the candidate guidance information and the target guidance information.
[0053] In this embodiment, the interactive model (InterAction-Based Model) can be a Sim-Bert model; the similarity between the candidate guidance information and the target guidance information can be represented by the matching score (sccore) between the candidate guidance information and the target guidance information; the candidate guidance information and the target guidance information can be processed first, and then the processing result can be input into the interactive model to obtain the similarity between the candidate guidance information and the target guidance information.
[0054] Step 103: Based on similarity, determine the target candidate guidance information from multiple candidate guidance information, and determine the target candidate guidance information as the augmented data of the target guidance information.
[0055] In this embodiment of the application, the target candidate guidance information is the guidance information selected from multiple candidate guidance information, and specifically it can be the guidance information with the highest similarity among multiple candidate guidance information; the similarity of multiple candidate guidance information can be compared and filtered, and the guidance information with the highest similarity can be selected as the target candidate guidance information, and it can be used as the enhanced first aid guidance knowledge data of the target guidance information.
[0056] The data augmentation method provided in this application uses sample data from the target domain to learn professional domain knowledge from an initial large language model, enabling the target large language model to have more professional domain knowledge capabilities. This results in more professional and diverse candidate guidance information. Furthermore, based on the professional and diverse candidate guidance information and the target guidance information, deep matching is performed using an interactive model, thereby ensuring the relevance and professionalism of the final generated augmented knowledge.
[0057] Based on the foregoing embodiments, this application provides yet another data augmentation method, referring to... Figure 2 and Figure 3 As shown, the method includes the following steps:
[0058] Step 201: The data augmentation device acquires sample data in the medical field with multimodal forms and processes the sample data to obtain target sample data.
[0059] The target areas include the medical field.
[0060] Step 202: The data augmentation device uses the target sample data to train the initial large language model to obtain the target large language model.
[0061] It's important to note that when dealing with relatively limited first-aid guidance knowledge data, training and fine-tuning various models require data augmentation to improve their generalization and robustness. Currently, generating data using large language models is a good approach. However, ensuring the reliability and professionalism of data generated by large language models remains a challenge, as they are prone to introducing errors or noisy data. Furthermore, the general capabilities of large language models are insufficient for specific industry or scenario applications. In practical industry applications, it's often necessary to augment and infuse general language models with industry-specific knowledge to make them more professional and better able to solve industry-specific or scenario-specific problems.
[0062] Therefore, in this embodiment, the initial large language model is trained by acquiring sample data from the medical field that possesses multimodal forms (including text, speech, images, and videos, etc.). This involves subjecting the large language model to domain-specific knowledge learning, enabling the trained large language model (i.e., the target large language model) to possess more specialized domain knowledge capabilities, thereby ensuring the professionalism of the enhanced data. Specifically, for example... Figure 4 As shown, sample data can include subject-specific books, professional literature, public health data, emergency medical records, medical examination data, and publicly available medical datasets. Data processing can specifically include data cleaning, redundancy removal, and quality screening. Specifically, after obtaining the sample data, it can be cleaned (e.g., using heuristic rules or other rules), and the cleaned data can be redundant removed (e.g., using the minimum hash algorithm (minHash) or Local Sensitive Hashing (LSH)). Then, the data after redundancy removal can be quality screened (e.g., using Perplexity (PPL) or N-gram) to obtain the target sample data.
[0063] It should be noted that the subject-specific books cover 18 departments, from undergraduate to doctoral levels; professional literature covers monographs and experimental research; public health covers diseases, symptoms, and procedures; emergency medical cases cover well-known case databases and images; medical examination data covers qualification examinations such as practicing physicians, head nurses, Chinese and Western pharmacists, and medical postgraduate entrance examinations; and public medical datasets cover knowledge Q&A and consultation records.
[0064] In this embodiment of the application, target sample data can be input into an initial large language model, and the initial large language model can be trained to obtain the target large language model, such as... Figure 5 As shown, model training can include continuous pre-training (CPT) and data fine-tuning. Specifically, the initial large language model can be incrementally pre-trained based on target sample data using strategies such as distributed training and zero-redundancy optimizer (ZeRo) optimization. Then, the model can be fine-tuned using instruction fine-tuning methods and reinforcement learning to obtain the target large language model. In this way, by collecting data from fields such as emergency care, public health, and medicine, the initial large language model can be incrementally pre-trained and fine-tuned to improve the professional domain capabilities of the target large language model.
[0065] It should be noted that, as Figure 5 As shown, distributed training can include data parallelism (DP), tensor parallelism (TP), and pipelined parallelism (PP); ZeRo optimization includes ZeRo Stage 3, Optimizer Offload, and Param Offload; other strategies can include Automatic Mixed Precision (AMP), Gradient Accumulation (GA), and Gradient Check Pointing (GC); instruction fine-tuning methods can include Full Fine-tuning (FT) and Loop Refinement (LORA); reinforcement learning includes Direct Preference Optimization (DPO) and Proximal Policy Optimization (PPO).
[0066] Step 203: The data augmentation device generates multiple target guidance information based on the sample description information of the sample patients in different scenarios and the initial guidance information for the sample patients.
[0067] It should be noted that after step 203, step 204 can be executed first, step 205 can be executed first, or steps 204 and 205 can be executed simultaneously. In this embodiment, step 204 is executed after step 203, and then step 205 is executed.
[0068] Step 204: The data augmentation device stores multiple target guidance information into the guidance library and corrects the target guidance information in the guidance library to update the guidance library.
[0069] In this embodiment, sample description information can refer to the description information of the sample patient, and specifically may include questions consulted by the sample patient, requests made by the sample patient, etc.; initial guidance information can refer to relevant guidance information based on the sample description information of the sample patient, i.e., first aid-related guidance knowledge; initial guidance information and sample description information of the sample patient can be collected, and multiple target guidance information can be generated through professional doctors and professional first aid guidance manuals; it should be noted that the sample description information of the sample patient should be collected and organized to cover scenarios of multiple patients as much as possible; professional knowledge correspondence and sorting of the current situation of the sample patient should be carried out in combination with professional doctors and professional guidance books to ensure the scientific feasibility of the sample description information and first aid guidance knowledge; then, multiple target guidance information can be stored in the guidance library, and the target guidance information in the guidance library can be corrected to update the guidance library, so as to ensure the timeliness and advancement of the guidance library.
[0070] Step 205: The data augmentation device generates prompts based on the current description information and target guidance information of the target patient.
[0071] In this embodiment, the prompt can specifically refer to a new guidance message generated by combining the current description information of the target patient and the target guidance information, and the regenerated guidance information should maintain the same meaning as the target guidance information as much as possible. The prompt can be generated based on the current description information of the target patient (i.e., the patient's consultation questions, including chief complaint, consciousness, and breathing) and expert emergency care guidance knowledge. In this way, the prompt constructed based on the current description information and the target guidance information generates a variety of candidate answers, thereby ensuring the diversity of enhanced data.
[0072] Step 206: The data augmentation device generates multiple candidate guidance messages based on prompt words, the target patient's historical descriptive information, and the target large language model.
[0073] In this embodiment of the application, the guidance information may refer to first aid guidance knowledge; the historical description information may refer to the description of the target patient's past illness and symptoms; the prompt words and the target patient's historical illness information can be input into the target large language model, and the target large language model can output multiple candidate guidance information, specifically N candidate guidance information.
[0074] Step 207: The data augmentation device performs splicing and serialization processing on multiple candidate guidance information and target guidance information to obtain multiple processed guidance information.
[0075] In this embodiment, each candidate guidance information from multiple candidate guidance information can be concatenated with the target guidance information first, and then the concatenated information can be serialized. For example, N candidate guidance information can be represented as C = {C1, C2, ..., C...} N The target guidance information can be represented as S = (S1, S2, ..., S...). n The processed guidance information obtained after splicing and serialization can be represented as {C1-S, C2-S, ..., C...}. N -S}.
[0076] Step 208: The data augmentation device determines the similarity between each candidate guidance information and the target guidance information based on multiple processed guidance information and interactive models.
[0077] In this embodiment, each processed guidance information corresponds to a similarity score. Specifically, multiple processed guidance information can be input into the interactive model to obtain multiple similarities. In this way, based on diverse candidate answers, deep matching is performed in combination with the interactive model to ensure the relevance of the final enhanced knowledge.
[0078] In the embodiments of this application, such as Figure 6 As shown, the interactive model includes an encoder, a pooling layer, and an output layer (i.e., MLP). Therefore, the encoder, pooling, and output layers can be used sequentially to process multiple processed guidance information to obtain similarity. It should be noted that before processing the multiple processed guidance information sequentially using the encoder, pooling, and output layers, the multiple processed guidance information can be embedded first.
[0079] It should be noted that step 208 can be achieved in the following way:
[0080] Step 208A1: The data augmentation device uses an encoding layer to encode multiple processed guidance information to obtain the first feature information corresponding to each processed guidance information.
[0081] In this embodiment of the application, the first feature vector may refer to the feature information obtained by processing multiple processed guidance information through the coding layer. Specifically, it may be a feature vector. That is, the first feature vector may refer to the feature vector obtained by using the coding layer to perform high-dimensional representation of the processed guidance information.
[0082] Step 208A2: The data augmentation device uses a pooling layer to perform pooling processing on each first feature information to obtain the second feature information.
[0083] In this embodiment of the application, the second feature information may refer to the feature vector obtained by pooling the first feature information using a pooling layer.
[0084] Step 208A3: The data augmentation device obtains the similarity between each candidate guidance information and the target guidance information based on the output layer and each second feature information.
[0085] The interactive model includes an encoding layer, a pooling layer, and an output layer.
[0086] In this embodiment, the obtained second feature information can be input to the output layer to obtain a matching score between the candidate guidance information and the target guidance information. This matching score is used to characterize the similarity between the candidate guidance information and the target guidance information; that is, the higher the matching score, the higher the similarity between the candidate guidance information and the target guidance information. Figure 5 As shown, the multiple matching scores output can be represented as {C1-S: score1, C2-S: score2, ..., C...} N -S:sccoreN}.
[0087] Step 209: The data augmentation device determines the target candidate guidance information from multiple candidate guidance information based on similarity, and determines the target candidate guidance information as the augmented data of the initial guidance information.
[0088] It should be noted that the descriptions of the same steps and contents as in other embodiments in this embodiment can be found in the descriptions in other embodiments, and will not be repeated here.
[0089] The data augmentation method provided in this application uses sample data from the target domain to learn professional domain knowledge from an initial large language model, enabling the target large language model to have more professional domain knowledge capabilities. This results in more professional and diverse candidate guidance information. Furthermore, based on the professional and diverse candidate guidance information and the target guidance information, deep matching is performed using an interactive model, thereby ensuring the relevance and professionalism of the final generated augmented knowledge.
[0090] Based on the foregoing embodiments, this application provides a data enhancement device that can be applied to... Figure 1 and Figure 2 In the data augmentation method provided in the corresponding embodiment, refer to Figure 7 As shown, the data enhancement device 3 may include: a first processing unit 31, a second processing unit 32, and a determination unit 33, wherein:
[0091] The first processing unit 31 is used to determine multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model; wherein, the target large language model is obtained by training an initial large language model using sample data from the target domain;
[0092] The second processing unit 32 is used to determine the similarity between the candidate guidance information and the target guidance information based on the candidate guidance information, the target guidance information, and the interactive model;
[0093] The determination unit 33 is used to determine the target candidate guidance information from multiple candidate guidance information based on similarity, and to determine the target candidate guidance information as the augmented data of the target guidance information.
[0094] In other embodiments of this application, the first processing unit 31 is further configured to perform the following steps:
[0095] Acquire sample data in the medical field that has multiple modalities, and process the sample data to obtain target sample data; wherein, the target field includes the medical field;
[0096] The initial large language model is trained using target sample data to obtain the target large language model.
[0097] In other embodiments of this application, the first processing unit 31 is further configured to perform the following steps:
[0098] Based on the current description information and target guidance information of the target patient, prompt words are generated;
[0099] Based on prompt words, historical descriptive information of the target patient, and the target large language model, multiple candidate guidance messages are generated.
[0100] In other embodiments of this application, the second processing unit 32 is further configured to perform the following steps:
[0101] Based on the sample description information of sample patients in different scenarios and the initial guidance information for sample patients, multiple target guidance information is generated;
[0102] Store multiple target guidance information in the guidance library, and correct the target guidance information in the guidance library to update the guidance library.
[0103] In other embodiments of this application, the second processing unit 32 is further configured to perform the following steps:
[0104] Multiple candidate guidance information and target guidance information are concatenated and serialized to obtain multiple processed guidance information;
[0105] Based on multiple processed guidance information and interactive models, the similarity between each candidate guidance information and the target guidance information is determined.
[0106] In other embodiments of this application, the second processing unit 32 is further configured to perform the following steps:
[0107] A coding layer is used to encode multiple processed guidance information to obtain the first feature information corresponding to each processed guidance information;
[0108] A pooling layer is used to perform pooling processing on each first feature information to obtain the second feature information;
[0109] Based on the output layer and each second feature information, the similarity between each candidate guidance information and the target guidance information is obtained.
[0110] It should be noted that the specific implementation process of the steps performed by each module in the embodiments of this application can be referred to Figure 1 and Figure 2 The implementation process of the data augmentation method provided in the corresponding embodiments will not be described in detail here.
[0111] The data augmentation apparatus provided in the embodiments of this application learns professional domain knowledge from the initial large language model by using sample data from the target domain, thereby enabling the target large language model to have more professional domain knowledge capabilities. As a result, the generated candidate guidance information is more professional and diverse. Furthermore, based on the professional and diverse candidate guidance information and the target guidance information, deep matching is performed in combination with an interactive model, thereby ensuring the relevance and professionalism of the final generated augmented knowledge.
[0112] Based on the foregoing embodiments, embodiments of this application provide a data enhancement device that can be applied to... Figure 1 and Figure 2 In the data augmentation method provided in the corresponding embodiment, refer to Figure 8 As shown, the data enhancement device 4 may include: a processor 41, a memory 42, and a communication bus 43, wherein:
[0113] Communication bus 43 is used to realize the communication connection between processor 41 and memory 42;
[0114] Processor 41 is used to execute the data enhancement program in memory 42 to perform the following steps:
[0115] Based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model, multiple candidate guidance information are identified; among them, the target large language model is obtained by training the initial large language model using sample data from the target domain;
[0116] Based on candidate guidance information, target guidance information, and interactive model, the similarity between candidate guidance information and target guidance information is determined;
[0117] Based on similarity, target candidate guidance information is identified from multiple candidate guidance information, and the target candidate guidance information is identified as augmented data of the target guidance information.
[0118] In other embodiments of this application, processor 41 is used to execute a data augmentation program in memory 42 to determine multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patient, and the target large language model, in order to perform the following steps:
[0119] Acquire sample data in the medical field that has multiple modalities, and process the sample data to obtain target sample data; wherein, the target field includes the medical field;
[0120] The initial large language model is trained using target sample data to obtain the target large language model.
[0121] In other embodiments of this application, processor 41 is used to execute a data augmentation program in memory 42 to determine multiple candidate guidance information based on the current description information of the target patient, target guidance information, and target large language model, in order to achieve the following steps:
[0122] Based on the current description information and target guidance information of the target patient, prompt words are generated;
[0123] Based on prompt words, historical descriptive information of the target patient, and the target large language model, multiple candidate guidance messages are generated.
[0124] In other embodiments of this application, processor 41 is used to execute a data enhancement program in memory 42 to perform the following steps:
[0125] Based on the sample description information of sample patients in different scenarios and the initial guidance information for sample patients, multiple target guidance information is generated;
[0126] Store multiple target guidance information in the guidance library, and correct the target guidance information in the guidance library to update the guidance library.
[0127] In other embodiments of this application, processor 41 is configured to execute a data augmentation program in memory 42 to determine the similarity between candidate guidance information and target guidance information based on candidate guidance information, target guidance information, and an interactive model, in order to perform the following steps:
[0128] Multiple candidate guidance information and target guidance information are concatenated and serialized to obtain multiple processed guidance information;
[0129] Based on multiple processed guidance information and interactive models, the similarity between each candidate guidance information and the target guidance information is determined.
[0130] In other embodiments of this application, processor 41 is configured to execute a data augmentation program in memory 42 based on multiple processed guidance information and an interactive model to determine the similarity between each candidate guidance information and the target guidance information, in order to perform the following steps:
[0131] A coding layer is used to encode multiple processed guidance information to obtain the first feature information corresponding to each processed guidance information;
[0132] A pooling layer is used to perform pooling processing on each first feature information to obtain the second feature information;
[0133] Based on the output layer and each second feature information, the similarity between each candidate guidance information and the target guidance information is obtained; wherein, the interactive model includes an encoding layer, a pooling layer and an output layer.
[0134] It should be noted that a detailed description of the steps performed by the processor can be found in [reference needed]. Figure 1 and Figure 2 The implementation process of the data augmentation method provided in the corresponding embodiments will not be described in detail here.
[0135] The data augmentation device provided in this application uses sample data from the target domain to learn professional domain knowledge from an initial large language model, enabling the target large language model to have more professional domain knowledge capabilities. This results in more professional and diverse candidate guidance information. Furthermore, based on the professional and diverse candidate guidance information and the target guidance information, deep matching is performed using an interactive model, thereby ensuring the relevance and professionalism of the final generated augmented knowledge.
[0136] Based on the foregoing embodiments, this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to achieve... Figure 1 and Figure 2 The steps in the data augmentation method provided in the corresponding embodiment.
[0137] Based on the foregoing embodiments, this application also provides a computer program product, including a computer program that can be executed by the processor 41 of the data enhancement device 4 to achieve... Figure 1 and Figure 2 The steps in the data augmentation method provided in the corresponding embodiment.
[0138] It should be noted that the aforementioned computer-readable storage media can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.
[0139] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0140] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0141] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.
[0142] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0143] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0144] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0145] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A data augmentation method, characterized in that, The method includes: Based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model, multiple candidate guidance information are determined; wherein, the target large language model is obtained by training an initial large language model using sample data from the target domain; Based on the candidate guidance information, the target guidance information, and the interactive model, the similarity between the candidate guidance information and the target guidance information is determined; Based on the similarity, target candidate guidance information is determined from multiple candidate guidance information, and the target candidate guidance information is determined to be augmented data of the target guidance information.
2. The method according to claim 1, characterized in that, Before determining multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model, the following steps are also included: Acquire sample data in the medical field that has a multimodal form, and process the sample data to obtain target sample data; wherein, the target field includes the medical field; The target large language model is obtained by training the initial large language model using the target sample data.
3. The method according to claim 2, characterized in that, Based on the current description information of the target patient, the target guidance information, and the target large language model, multiple candidate guidance information are determined, including: Based on the current description information of the target patient and the target guidance information, prompt words are generated; Based on the prompt words, the target patient's historical description information, and the target large language model, multiple candidate guidance messages are generated.
4. The method according to claim 1, characterized in that, Before determining multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model, the following steps are also included: Based on the sample description information of the sample patients in different scenarios and the initial guidance information for the sample patients, multiple target guidance information is generated; Multiple sets of target guidance information are stored in a guidance library, and the target guidance information in the guidance library is corrected to update the guidance library.
5. The method according to claim 1, characterized in that, The step of determining the similarity between the candidate guidance information and the target guidance information based on the candidate guidance information, the target guidance information, and the interactive model includes: Multiple candidate guidance information and target guidance information are concatenated and serialized to obtain multiple processed guidance information; Based on the processed guidance information and the interactive model, the similarity between each candidate guidance information and the target guidance information is determined.
6. The method according to claim 5, characterized in that, The step of determining the similarity between each candidate guidance information and the target guidance information based on multiple processed guidance information and the interactive model includes: A coding layer is used to encode multiple processed guidance information to obtain first feature information corresponding to each processed guidance information; A pooling layer is used to perform pooling processing on each of the first feature information to obtain the second feature information; Based on the output layer and each of the second feature information, the similarity between each candidate guidance information and the target guidance information is obtained; wherein, the interactive model includes the encoding layer, the pooling layer and the output layer.
7. A data augmentation device, characterized in that, The device includes: The first processing unit is used to determine multiple candidate guidance information based on the current description information of the target patient, the target guidance information for the sample patients, and the target large language model; wherein, the target large language model is obtained by training an initial large language model using sample data from the target domain; The second processing unit is used to determine the similarity between the candidate guidance information and the target guidance information based on the candidate guidance information, the target guidance information, and the interactive model; The determining unit is configured to determine target candidate guidance information from multiple candidate guidance information based on the similarity, and determine the target candidate guidance information as enhanced data of the target guidance information.
8. A data enhancement device, characterized in that, The device includes: a processor, a memory, and a communication bus; The communication bus is used to realize the communication connection between the processor and the memory; The processor is used to execute a data augmentation program in memory to implement the steps of the data augmentation method as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs that can be executed by one or more processors to implement the steps of the data augmentation method as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the data augmentation method as described in any one of claims 1 to 6.