Medical teaching exercise answer retrieval method based on feature extraction
Through feature extraction and approximate nearest neighbor search algorithms, the answers are efficiently retrieved in the medical teaching learning question answer database, and the large language model is used to generate answers, which solves the problem of low efficiency and inaccurate retrieval of traditional medical textbook exercises, and achieves fast and accurate search results and systematic optimization.
Patent Information
- Application Number
- CN202411931973.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-26
- Publication Date
- 2025-04-29
AI Technical Summary
The search methods for traditional medical textbook exercises are inefficient and inaccurate enough, making it difficult to meet the integration and optimization of massive exercise resources and knowledge.
The medical teaching learning question answer search method based on feature extraction is adopted, and the pictures are converted into processable text data through the image information extraction model. The feature extraction model is used to generate the text feature vector of the medical teaching learning question, and the answer is efficiently retrieved in the medical teaching learning question answer library through the approximate nearest neighbor search algorithm. At the same time, the answer is generated using a large language model and the answer library is optimized.
It realizes fast and accurate medical teaching study questions search results, improves search efficiency and accuracy, and continuously optimizes the accuracy and practicality of the system through feedback mechanisms.
Smart Images
Figure CN119903147A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical teaching auxiliary technology, and in particular to a method for retrieving answers to medical teaching questions based on feature extraction. Background Art
[0002] Medical textbooks are extensive and in-depth, covering everything from basic medical theory to clinical practice. With the continuous deepening of medical research and the continuous innovation of medical technology, the content of textbooks is also constantly updated and expanded. This makes traditional exercise retrieval methods, such as manually flipping through books or relying on limited online resources, increasingly inefficient and difficult to meet needs. Students and educators often need to spend a lot of time and energy to find relevant exercises and answers, which affects learning efficiency and educational effectiveness.
[0003] When inputting medical exercises into the system to retrieve answers, traditionally, text input is mainly used, which is inefficient and may also result in input errors. There are also some methods and systems in the prior art that perform answer retrieval through image recognition. However, due to the extensiveness and complexity of the content of medical textbooks, the prior art is not always accurate and fast enough for retrieving answers to medical exercises. Summary of the invention
[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for retrieving answers to medical teaching exercises based on feature extraction, so as to provide users with fast and accurate medical exercise retrieval results.
[0005] The purpose of the present invention can be achieved by the following technical solutions:
[0006] A method for retrieving answers to medical teaching questions based on feature extraction, the method comprising:
[0007] Obtain the image information of medical teaching questions that need to be queried;
[0008] Use the image information extraction model to extract image data and convert the image into processable text data;
[0009] Use the feature extraction model to extract features from text data and generate feature vectors of medical teaching questions text;
[0010] By using an approximate nearest neighbor search algorithm, a medical teaching text feature vector with the highest cosine similarity to a medical teaching text feature vector is retrieved from a medical teaching question answer database containing multiple medical domains, and the corresponding page in the textbook is located according to the medical teaching text feature vector;
[0011] The retrieved medical teaching text feature vectors and the corresponding pages of the textbooks are displayed to the user in a visual form;
[0012] Receive the user's first feedback on the displayed results. If the first feedback is positive, end the search. Otherwise, use the large language model to generate answers based on the text data and return them to the user again.
[0013] Receive second feedback from users on the answers generated by the large language model, and optimize the medical teaching question answer library and the large language model based on the second feedback.
[0014] Furthermore, the medical teaching question answer library is constructed before searching, and the process of constructing the medical teaching question answer library includes:
[0015] Obtain medical teaching content, including medical teaching questions, medical textbooks, and user-uploaded exercises, and extract teaching data from them using an image information extraction model;
[0016] Use the feature extraction model to extract features from teaching data and generate feature vectors of medical teaching text;
[0017] Preprocess the feature vector of medical teaching text, including denoising, normalization and dimension reduction;
[0018] According to the medical teaching content, determine the number of subject categories to be divided, and randomly select the same number of medical teaching text feature vectors as the initial clustering centers;
[0019] Calculate the distance between each medical teaching text feature vector and each cluster center, and classify each medical teaching text feature vector and the cluster center with the closest distance into one category; after all medical teaching text feature vectors are classified, recalculate the cluster center of each category according to the medical teaching text feature vectors in each category, take the mean of all feature vectors in each category as the new cluster center, and repeat the classification and calculation until the position of the cluster center no longer changes or reaches a predetermined number of iterations;
[0020] According to the final cluster centers, the medical teaching text feature vectors are divided into different subject categories;
[0021] The feature vectors of medical teaching texts of different subject categories are stored in different medical domains, and corresponding medical domain keywords are assigned according to the subject categories to establish a medical domain index;
[0022] According to the keywords and indexes of each medical domain, a corresponding vector library is created, and the medical teaching text feature vectors stored in the medical domain are imported into the corresponding vector library to form a structured medical teaching question answer library;
[0023] Regularly update the vector library based on user feedback and network searches;
[0024] The medical teaching text feature vector from the medical textbook includes the textbook page number and the textbook page map corresponding to the medical teaching text feature vector;
[0025] The medical domain includes: basic medicine domain, clinical medicine domain, public health domain, medical engineering domain and medical humanities domain. Among them, the basic medicine domain includes the feature vectors of medical teaching texts such as human anatomy, physiology, biochemistry and pathology; the clinical medicine domain includes the feature vectors of medical teaching texts such as internal medicine, surgery, obstetrics and gynecology, pediatrics, ophthalmology and dentistry; the public health domain includes the feature vectors of medical teaching texts such as disease prevention, disease control and preventive medicine; the medical engineering domain includes the feature vectors of medical teaching texts such as medical devices and medical imaging; the medical humanities domain includes the feature vectors of medical teaching texts such as medical ethics, medical history, patient communication and interpersonal relationships and cross-cultural medicine.
[0026] Furthermore, the image information extraction model includes a text detection module, a text recognition module and a post-processing module, wherein the text detection module uses the pruned, distilled and quantized DB network as the basic network, and the text recognition module uses the transformer as the basic network.
[0027] Furthermore, the process of converting the image into processable text data by the image information extraction model includes:
[0028] Use the text detection module to process the image information and obtain the text area bounding box;
[0029] Use the text recognition module to perform text recognition on the image in the text area boundary box to obtain the text content and confidence level in the boundary box;
[0030] The post-processing module is used to take the remainder of the recognized text, and the recognition results after taking the remainder are spliced and corrected to obtain processable text data.
[0031] Furthermore, the text data for feature extraction includes not only text converted by using the image information extraction model, but also text manually input by the user.
[0032] Furthermore, the feature extraction model includes an encoding layer and a transformer layer. The encoding layer converts the text data of the input model into word vectors through word embedding, position encoding and segment encoding. The transformer layer extracts deeper semantic information from the word vectors through a multi-head attention mechanism and a feedforward neural network, and finally outputs a medical teaching question text feature vector containing deep semantic information.
[0033] Furthermore, the approximate nearest neighbor search algorithm includes hierarchical clustering, product quantization and inverted files.
[0034] Furthermore, the visualization form includes directly displaying the answer content, displaying the page containing the answer in a preview manner, and being able to continue to display the remaining pages of the entire teaching material in a preview manner after displaying the page containing the answer.
[0035] Furthermore, the large language model is obtained through mechanical algorithm training using the contents of a medical teaching question answer library as a data set.
[0036] Furthermore, the optimization includes:
[0037] Receive the second feedback from the user on the answer generated by the large language model. If the second feedback is positive, use the feature extraction model to vectorize the answer generated by the large language model and store it in the medical teaching question answer library. If the second feedback is negative, send the medical teaching question image information queried this time to the maintenance personnel, receive the answer information sent back by the maintenance personnel, use the feature extraction model to vectorize the answer information, store it in the medical teaching question answer library, and re-train the large language model according to the updated medical teaching question answer library.
[0038] Compared with the prior art, the beneficial effects of the present invention include:
[0039] 1. Aiming at the difficulties in retrieving answers to medical textbooks, the present invention discloses a method for retrieving answers to medical teaching questions based on feature extraction, integrates and optimizes massive exercise resources and knowledge, and provides users with fast and accurate medical teaching question retrieval results; the medical teaching question answer library used for retrieval contains multiple medical domains, and the knowledge content is stored in categories, which can improve the speed of establishing the early data set and the speed of subsequent retrieval; the approximate nearest neighbor search algorithm is used to efficiently retrieve and query the text feature vector with the highest cosine similarity of the feature vector in the medical teaching question answer library, so as to quickly locate the answer, not only retrieve the answer content, but also locate the corresponding page in the textbook according to the medical teaching text feature vector, which is convenient for users to view and understand; by receiving user feedback on the displayed results, the answer is generated using a large language model, and the medical teaching question answer library and the large language model are optimized according to the user's feedback on the answer. This feedback mechanism helps to continuously improve the accuracy and practicality of the system and ensure the timeliness and accuracy of the answer library;
[0040] 2. The present invention has established a medical teaching question answer library containing multiple medical domains. By covering multiple medical domains such as basic medicine, clinical medicine, public health, medical engineering and medical humanities, the answer library can provide more comprehensive medical knowledge coverage. The structured answer library makes the retrieval process more efficient and also makes it easier to update and maintain. With the continuous updating and development of medical knowledge, new exercises and answers can be added to the answer library regularly to maintain its timeliness and accuracy;
[0041] 3. In the present invention, the text detection module of the image information extraction model uses an improved DB network, and the model is lightweight through pruning, distillation and quantization, which improves the speed of model image processing. The text recognition module uses transformer as the basic network, which has better recognition ability for long texts. Compared with traditional OCR image recognition, the image information extraction model takes into account both the performance and efficiency of the model, has a higher precision text recognition ability, can accurately read text even in complex backgrounds, and has strong robustness;
[0042] 4. The medical teaching question answer library containing multiple medical domains established by the present invention can not only be used to retrieve answers, but also to train a large language model for generating answers, which can be reused, saving the time for training a large language model to establish a data set, and making the return of answers more complete, thereby improving retrieval and learning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0045] Example 1
[0046] This embodiment intends to disclose a method for retrieving answers to medical teaching questions based on feature extraction, the method is as follows Figure 1 As shown, it includes the following steps:
[0047] Step S1, obtaining the medical teaching question image information to be queried.
[0048] The content of the picture information is course questions and exercises related to medical education.
[0049] Step S2, extracting image data using an image information extraction model, and converting the image into processable text data.
[0050] The image information extraction model includes a text detection module, a text recognition module and a post-processing module. The text detection module uses the pruned, distilled and quantized DB network as the basic network, and the text recognition module uses the transformer as the basic network.
[0051] The image information extraction model used in this embodiment is PP-OCRv4. Pruning, distillation and quantization can lighten the model and speed up the processing speed of the image information extraction model. Transformer is selected as the basic network because the attention mechanism has excellent performance in the recognition ability of long texts.
[0052] The process of converting images into processable text data by the image information extraction model includes:
[0053] Use the text detection module to process the image information and obtain the text area bounding box;
[0054] Use the text recognition module to perform text recognition on the image in the text area boundary box to obtain the text content and confidence level in the boundary box;
[0055] The post-processing module is used to take the remainder of the recognized text, and the recognition results after taking the remainder are spliced and corrected to obtain processable text data.
[0056] Step S3, using a feature extraction model to extract features from the text data, and generating a feature vector of the medical teaching question text.
[0057] The text data for feature extraction here includes not only the text converted by the image information extraction model, but also the text manually input by the user.
[0058] This embodiment also provides a verification function of the OCR recognition result on the basis of automatic recognition, allowing the user to manually correct erroneous recognition results and improve the accuracy of the final answer.
[0059] The feature extraction model includes an encoding layer and a transformer layer. The encoding layer generates a word vector for each word by combining position encoding and segment encoding, and converts the text data of the input model into a word vector. The transformer layer is used to capture the deep semantic relationship between sentences, and extracts deeper semantic information from the word vector through a multi-head attention mechanism and a feedforward neural network, processes cross-sentence semantics, and finally outputs a text feature vector of medical teaching questions containing deep semantic information.
[0060] The feature extraction model used in this embodiment is the SimBERT model. While maintaining the powerful semantic representation capability of BERT, this model further improves the accuracy and robustness of sentence-level semantic similarity.
[0061] Step S4, using an approximate nearest neighbor search algorithm, retrieve the medical teaching text feature vector with the highest cosine similarity to the medical teaching question text feature vector in the medical teaching question answer database, and locate the corresponding page in the textbook according to the medical teaching text feature vector.
[0062] The approximate nearest neighbor search algorithm includes hierarchical clustering (Hierarchical Clustering), product quantization (Product Quantization, PQ) and inverted file indexing (Inverted File Indexing, IVF).
[0063] These algorithms can significantly reduce computational complexity while ensuring retrieval accuracy, thereby achieving efficient query of large-scale data.
[0064] The medical teaching question answer library mentioned in step S4 is constructed before the search. The specific construction process includes:
[0065] Obtain medical teaching content, including medical teaching questions, medical textbooks, and user-uploaded exercises, and extract teaching data from them using an image information extraction model;
[0066] Use the feature extraction model to extract features from teaching data and generate feature vectors of medical teaching text;
[0067] Preprocess the feature vector of medical teaching text, including denoising, normalization and dimension reduction;
[0068] According to the medical teaching content, determine the number of subject categories to be divided, and randomly select the same number of medical teaching text feature vectors as the initial clustering centers;
[0069] Calculate the distance between each medical teaching text feature vector and each cluster center, and classify each medical teaching text feature vector and the cluster center with the closest distance into one category; after all medical teaching text feature vectors are classified, recalculate the cluster center of each category according to the medical teaching text feature vectors in each category, take the mean of all feature vectors in each category as the new cluster center, and repeat the classification and calculation until the position of the cluster center no longer changes or reaches a predetermined number of iterations;
[0070] According to the final cluster centers, the medical teaching text feature vectors are divided into different subject categories;
[0071] The feature vectors of medical teaching texts of different subject categories are stored in different medical domains, and corresponding medical domain keywords are assigned according to the subject categories to establish a medical domain index;
[0072] According to the keywords and indexes of each medical domain, a corresponding vector library is created, and the medical teaching text feature vectors stored in the medical domain are imported into the corresponding vector library to form a structured medical teaching question answer library;
[0073] Regularly update the vector library based on user feedback and network searches;
[0074] The medical teaching text feature vector from the medical textbook includes the textbook page number and the textbook page map corresponding to the medical teaching text feature vector;
[0075] The medical domain includes: basic medicine domain, clinical medicine domain, public health domain, medical engineering domain and medical humanities domain. Among them, the basic medicine domain includes the feature vectors of medical teaching texts such as human anatomy, physiology, biochemistry and pathology; the clinical medicine domain includes the feature vectors of medical teaching texts such as internal medicine, surgery, obstetrics and gynecology, pediatrics, ophthalmology and dentistry; the public health domain includes the feature vectors of medical teaching texts such as disease prevention, disease control and preventive medicine; the medical engineering domain includes the feature vectors of medical teaching texts such as medical devices and medical imaging; the medical humanities domain includes the feature vectors of medical teaching texts such as medical ethics, medical history, patient communication and interpersonal relationships and cross-cultural medicine.
[0076] In this embodiment, the database used is the Faiss database, which is an efficient similarity search and dense vector clustering library that supports GPU acceleration. While supporting storage of databases in different medical domains, it can also use the parallel computing capabilities of the GPU for efficient data retrieval.
[0077] Step S5, displaying the retrieved medical teaching text feature vectors and the corresponding pages of the teaching materials to the user in a visual form.
[0078] The visualization forms include directly displaying the answer content, displaying the page containing the answer in preview mode, and being able to continue to display the remaining pages of the entire teaching material in preview mode after displaying the page containing the answer.
[0079] Step S6, receiving the user's first feedback on the displayed results. If the first feedback is positive, the search is terminated. Otherwise, the large language model is used to generate an answer based on the text data and the answer is returned to the user again.
[0080] Step S7, receiving a second feedback from the user regarding the answer generated by the large language model, and optimizing the medical teaching question answer library and the large language model according to the second feedback content.
[0081] In this embodiment, the user's feedback on the search results includes "very relevant", "not very relevant" and "completely irrelevant", among which "very relevant" is positive feedback, and "not very relevant" and "completely irrelevant" are negative feedback.
[0082] The large language model uses the content in the medical teaching question answer library as the data set and is trained through a mechanical algorithm.
[0083] When receiving the second feedback from the user on the answer generated by the large language model, if the second feedback is positive, the answer generated by the large language model is vectorized using the feature extraction model and stored in the medical teaching question answer library;
[0084] If the second feedback is negative, the medical teaching question image information queried this time will be sent to the maintenance personnel, and the answer information sent back by the maintenance personnel will be received. The answer information will be vectorized using the feature extraction model and stored in the medical teaching question answer library. At the same time, the large language model will be re-trained based on the updated medical teaching question answer library.
[0085] In another embodiment of the present invention, the optimization also includes performing a network search based on user feedback and recognized text content to obtain answers on the Internet, and after confirmation by maintenance personnel, using a feature extraction model to vectorize the retrieved answers and store them in a medical teaching question answer library.
[0086] Next, an example of actually applying the method for retrieving answers to medical teaching questions disclosed in this embodiment is given.
[0087] Obtain photos of medical teaching questions that users do not understand;
[0088] The image data was extracted using the image information extraction model PP-OCRv4, and the medical question in the image was extracted as "characteristics of the DNA double helix structure";
[0089] The feature extraction model SimBERT is used to transform the "DNA double helix structure characteristics" into feature vectors and generate feature vectors for medical teaching question texts;
[0090] Through the approximate nearest neighbor search algorithm, the medical teaching text feature vector with the highest cosine similarity to the vector of "DNA double helix structure characteristics" is retrieved from the pre-built medical teaching question answer database containing multiple medical domains, the most similar vector is found, and the specific page number in the textbook is located accordingly;
[0091] The query results are presented to the user in the form of visual images, showing the page where the relevant answers are located in the textbook;
[0092] The user provides feedback on the accuracy of the query results, including but not limited to marking the relevance of the query results. If the first feedback is "very relevant", the search is terminated. Otherwise, the large language model is used to generate the answer "characteristics of the DNA double helix structure" based on the text data and return it to the user again;
[0093] The user provides feedback on the accuracy of the query result again. If the second feedback is positive, the feature extraction model is used to vectorize the answer "Characteristics of DNA double helix structure" generated by the large language model and store it in the medical teaching question answer database;
[0094] If the second feedback is negative, the image information of "Characteristics of DNA Double Helix Structure" will be transmitted to the maintenance personnel, and the answer information sent back by the maintenance personnel will be received. The answer information of "Characteristics of DNA Double Helix Structure" will be vectorized using the feature extraction model and stored in the medical teaching question answer database. At the same time, the large language model will be re-trained according to the updated medical teaching question answer database.
[0095] Example 2
[0096] Based on Example 1, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the aforementioned feature extraction-based medical teaching question answer retrieval method.
[0097] At the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above-mentioned feature extraction-based medical teaching question answer retrieval method. Of course, in addition to the software implementation, 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.
[0098] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.
[0099] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.
[0100] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.
Claims
1. A method for retrieving answers to medical teaching questions based on feature extraction, characterized in that: The method comprises: Obtain the image information of medical teaching questions that need to be queried; Use the image information extraction model to extract image data and convert the image into processable text data; Use the feature extraction model to extract features from text data and generate feature vectors of medical teaching questions text; By using an approximate nearest neighbor search algorithm, a medical teaching text feature vector with the highest cosine similarity to a medical teaching text feature vector is retrieved from a medical teaching question answer database containing multiple medical domains, and the corresponding page in the textbook is located according to the medical teaching text feature vector; The retrieved medical teaching text feature vectors and the corresponding pages of the textbooks are displayed to the user in a visual form; Receive the user's first feedback on the displayed results. If the first feedback is positive, end the search. Otherwise, use the large language model to generate answers based on the text data and return them to the user again. Receive second feedback from users on the answers generated by the large language model, and optimize the medical teaching question answer library and the large language model based on the second feedback.
2. A method for retrieving answers to medical teaching questions based on feature extraction according to claim 1, characterized in that: The medical teaching question answer library is constructed before searching, and the process of constructing the medical teaching question answer library includes: Obtain medical teaching content, including medical teaching questions, medical textbooks, and user-uploaded exercises, and extract teaching data from them using an image information extraction model; Use the feature extraction model to extract features from teaching data and generate feature vectors of medical teaching text; Preprocess the feature vector of medical teaching text, including denoising, normalization and dimension reduction; According to the medical teaching content, determine the number of subject categories to be divided, and randomly select the same number of medical teaching text feature vectors as the initial clustering centers; Calculate the distance between each medical teaching text feature vector and each cluster center, and classify each medical teaching text feature vector and the cluster center with the closest distance into one category; after all medical teaching text feature vectors are classified, recalculate the cluster center of each category according to the medical teaching text feature vectors in each category, take the mean of all feature vectors in each category as the new cluster center, and repeat the classification and calculation until the position of the cluster center no longer changes or reaches a predetermined number of iterations; According to the final cluster centers, the medical teaching text feature vectors are divided into different subject categories; The feature vectors of medical teaching texts of different subject categories are stored in different medical domains, and corresponding medical domain keywords are assigned according to the subject categories to establish a medical domain index; According to the keywords and indexes of each medical domain, a corresponding vector library is created, and the medical teaching text feature vectors stored in the medical domain are imported into the corresponding vector library to form a structured medical teaching question answer library; Regularly update the vector library based on user feedback and network searches; The medical teaching text feature vector from the medical textbook includes the textbook page number and the textbook page map corresponding to the medical teaching text feature vector; The medical domain includes: basic medicine domain, clinical medicine domain, public health domain, medical engineering domain and medical humanities domain. Among them, the basic medicine domain includes the feature vectors of medical teaching texts such as human anatomy, physiology, biochemistry and pathology; the clinical medicine domain includes the feature vectors of medical teaching texts such as internal medicine, surgery, obstetrics and gynecology, pediatrics, ophthalmology and dentistry; the public health domain includes the feature vectors of medical teaching texts such as disease prevention, disease control and preventive medicine; the medical engineering domain includes the feature vectors of medical teaching texts such as medical devices and medical imaging; the medical humanities domain includes the feature vectors of medical teaching texts such as medical ethics, medical history, patient communication and interpersonal relationships and cross-cultural medicine.
3. A method for retrieving answers to medical teaching questions based on feature extraction according to claim 1, characterized in that: The image information extraction model includes a text detection module, a text recognition module and a post-processing module, wherein the text detection module uses the pruned, distilled and quantized DB network as the basic network, and the text recognition module uses the transformer as the basic network.
4. A method for retrieving answers to medical teaching questions based on feature extraction according to claim 3, characterized in that: The process of converting an image into processable text data by the image information extraction model includes: Use the text detection module to process the image information and obtain the text area bounding box; Use the text recognition module to perform text recognition on the image in the text area boundary box to obtain the text content and confidence level in the boundary box; The post-processing module is used to take the remainder of the recognized text, and the recognition results after taking the remainder are spliced and corrected to obtain processable text data.
5. A method for retrieving answers to medical teaching questions based on feature extraction according to claim 1, characterized in that: The text data for feature extraction includes not only text converted by using the image information extraction model, but also text manually input by the user.
6. A method for retrieving answers to medical teaching questions based on feature extraction according to claim 1, characterized in that: The feature extraction model includes an encoding layer and a transformer layer. The encoding layer converts the text data of the input model into word vectors through word embedding, position encoding and segment encoding. The transformer layer extracts deeper semantic information from the word vectors through a multi-head attention mechanism and a feedforward neural network, and finally outputs a text feature vector of medical teaching questions containing deep semantic information.
7. A method for retrieving answers to medical teaching questions based on feature extraction according to claim 1, characterized in that: The approximate nearest neighbor search algorithm includes hierarchical clustering, product quantization and inverted files.
8. The method for retrieving answers to medical teaching questions based on feature extraction according to claim 1, characterized in that: The visualization form includes directly displaying the answer content, displaying the page containing the answer in a preview manner, and being able to continue to display the remaining pages of the entire teaching material in a preview manner after displaying the page containing the answer.
9. The method for retrieving answers to medical teaching questions based on feature extraction according to claim 1, characterized in that: The large language model is obtained by training with a mechanical algorithm using the contents of a medical teaching question answer library as a data set.
10. The method for retrieving answers to medical teaching questions based on feature extraction according to claim 1, characterized in that: The optimization includes: Receive the second feedback from the user on the answer generated by the large language model. If the second feedback is positive, use the feature extraction model to vectorize the answer generated by the large language model and store it in the medical teaching question answer library. If the second feedback is negative, send the medical teaching question image information queried this time to the maintenance personnel, receive the answer information sent back by the maintenance personnel, use the feature extraction model to vectorize the answer information, store it in the medical teaching question answer library, and re-train the large language model according to the updated medical teaching question answer library.