Respiratory anesthesia knowledge retrieval type question answering method and system

Abstract

The invention belongs to the technical field of artificial intelligence and natural language processing, and relates to a respiratory anesthesia knowledge retrieval type question answering method and system, and the method comprises the steps: receiving a respiratory anesthesia related question text; performing segmentation, vectorization and text embedding processing on the problem text to obtain corresponding vectors; inputting a pre-established and trained respiratory anesthesia professional knowledge model, generating a related text by the respiratory anesthesia professional knowledge model through vector search of a retrieval type question and answer engine, generating question-related answer cues through context integration, and calculating and outputting answers to the questions according to similarity; the respiratory anesthesia professional knowledge model is obtained by fusing respiratory anesthesia professional knowledge to carry out fine adjustment on a basic big language model, and the respiratory anesthesia professional knowledge model is based on an enhanced Transform network structure, adopts a multi-head dynamic attention mechanism, and adds domain semantic adaptation, bidirectional interaction and retrieval enhancement.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and natural language processing, and in particular relates to a respiratory anesthesia knowledge retrieval type question-answering method and system. Background Art

[0002] In the current field of respiratory anesthesia medicine, there are at least four ways to obtain relevant knowledge, namely: 1. Doctors, research institutes, and medical workers have professional knowledge of respiratory anesthesia medicine. They can acquire a mature knowledge system through long-term practice, but this knowledge is carried by individuals, has strong independence, and requires time to update new knowledge; 2. Obtain professional knowledge about respiratory anesthesia by consulting relevant materials, books, expert consensus, and treatment manuals, but the knowledge carrier of this method is paper or electronic book media, and the retrieval of knowledge is very cumbersome and inefficient, and may be affected by knowledge updates and may cause untimely knowledge updates; 3. Inquiry and search through general large model dialogue tools such as ChatGPT, but due to the universality of the ChatGPT model, this method lacks understanding of the professional field of respiratory anesthesia. When dealing with complex problems, the system often gives inaccurate or irrelevant answers, which increases the cost of users acquiring knowledge. And affected by the update speed of the ChatGPT large model, it may not be possible to obtain the correct respiratory anesthesia expertise; 4. Through Baidu, Bing and other search engines, this type of search engine is a real-time crawler and a technology that integrates search information. It may not be able to obtain the correct search results in this regard, or it may be subject to the inconsistency of multiple knowledge and obtain incorrect respiratory anesthesia expertise. For the implementation of Baidu, Bing, and ChatGPT, it is necessary to generate, update, and fine-tune the model and online knowledge connections before the knowledge base can be updated, which requires a lot of manpower, material resources, and time costs. In summary, there is an urgent need for a knowledge retrieval question-and-answer method for respiratory anesthesia that can answer quickly and accurately, and the knowledge update is accurate, convenient, and timely. Summary of the invention

[0003] In view of the problems of slow knowledge update and low answer accuracy in the prior art, the purpose of the present invention is to overcome the defects of the prior art and propose a method for constructing a large language model for respiratory anesthesia knowledge retrieval question and answer.

[0004] In view of this, the present invention proposes a respiratory anesthesia knowledge retrieval question-answering method, comprising:

[0005] Step 1) receiving a text of a respiratory anesthesia-related question;

[0006] Step 2) segmenting, vectorizing and text embedding the question text to obtain the corresponding vector;

[0007] Step 3) inputting a pre-established and trained respiratory anesthesia expertise model, the respiratory anesthesia expertise model generates related texts through vector search of a retrieval-type question-answering engine, generates question-related answer prompt words through context integration, and outputs the answer to the question based on similarity calculation;

[0008] The respiratory anesthesia expertise model is obtained by fine-tuning a basic large language model by integrating respiratory anesthesia expertise. The respiratory anesthesia expertise model is based on the network structure of an enhanced Transformer, adopts a multi-head dynamic attention mechanism, and increases domain semantic adaptation, two-way interaction, and retrieval enhancement.

[0009] Preferably, the respiratory anesthesia-related issues in step 1) include: the principles of respiratory anesthesia, equipment operation methods, risk assessment and cutting-edge knowledge of respiratory anesthesia.

[0010] Preferably, the step 2) includes: using an adaptive text segmentation method to segment the question text into multiple text blocks, and converting them into question vector representations, generating question vectors through a text embedding model, and using a dimensionality reduction algorithm to map the question vectors to a low-dimensional semantic space.

[0011] Preferably, the basic large language model is a large language model model, and the respiratory anesthesia expertise model includes a retrieval and generation layer;

[0012] The retrieval and generation layer includes a knowledge base retrieval module, a context integration module and an answer generation module; wherein,

[0013] The knowledge base retrieval module is used to retrieve a pre-built respiratory anesthesia knowledge base through a vector search engine based on the vectorized problem representation to obtain knowledge text related to the problem semantics;

[0014] The context integration module is used to integrate context information using a multi-head dynamic attention mechanism to generate the most relevant knowledge fragments;

[0015] The answer generation module is used to generate answers to questions using a retrieval-enhanced generation method, and further improve the accuracy and relevance of the answers through a two-way interactive mechanism.

[0016] Preferably, the respiratory anesthesia expertise model includes an output processing layer, and the output processing layer includes: an answer optimization module and an answer output module, wherein:

[0017] The answer optimization module is used to adjust or reconstruct the answers according to the similarity of the generated answers;

[0018] The answer output module is used to convert the adjusted or reconstructed answers into readable text.

[0019] Preferably, the method further includes real-time dynamic updating of the respiratory anesthesia knowledge base, and further optimizing the vector search engine of the knowledge base retrieval module based on the updated respiratory anesthesia knowledge base by embedding an updating mechanism.

[0020] Preferably, the method further includes providing a customized function for the respiratory anesthesia knowledge base, so as to further expand the update content of the respiratory anesthesia knowledge base according to user permissions.

[0021] Preferably, the fine-tuning process of the respiratory anesthesia expertise model includes:

[0022] Collect public knowledge on respiratory anesthesia and annotate it to obtain a professional knowledge dataset;

[0023] Collect local documents of professional knowledge and extract text content;

[0024] Performing vectorization processing to obtain a set of multi-dimensional digital vectors, wherein each dimension represents a feature of the text, and the features include: vocabulary frequency and semantic relationship;

[0025] Perform text embedding to map digital vectors into a low-dimensional semantic space;

[0026] The mapped digital vector is input into the basic large language model, and the parameters are modulated until the training requirements are met to obtain a trained respiratory anesthesia expertise model.

[0027] Preferably, the public knowledge includes papers, expert consensus, annual reports and physician experience.

[0028] On the other hand, the present invention also provides a respiratory anesthesia knowledge retrieval type question-answering system, comprising:

[0029] A user interaction layer, used to receive texts of questions related to respiratory anesthesia;

[0030] Intelligent question-answering processing layer, used to segment, vectorize and embed question text to obtain corresponding vectors;

[0031] The intelligent model execution layer is used to input a pre-established and trained respiratory anesthesia professional knowledge model, the respiratory anesthesia professional knowledge model generates related texts through vector search of the retrieval-type question-answering engine, generates question-related answer prompt words through context integration, and outputs the answer to the question based on similarity calculation; and

[0032] The dynamic knowledge support layer is used to integrate respiratory anesthesia expertise to fine-tune the basic large language model to obtain a respiratory anesthesia expertise model, and is also used to dynamically update the respiratory anesthesia knowledge base in real time;

[0033] The respiratory anesthesia expertise model is based on the network structure of the enhanced Transformer, adopts a multi-head dynamic attention mechanism, and increases domain semantic adaptation, bidirectional interaction, and retrieval enhancement.

[0034] Compared with the prior art, the advantages of the present invention are:

[0035] 1. The present invention adopts fine-tuning technology to improve the model's ability to handle professional knowledge issues in respiratory anesthesia by fine-tuning the basic large language model, laying the foundation for subsequent steps in the present invention.

[0036] 2. The present invention integrates respiratory anesthesia expertise and, through designated processing of literature, guidelines and documents in the professional field, can update the knowledge base of the respiratory anesthesia expertise large model in near real time to ensure the accuracy of the system's answers.

[0037] 3. The present invention constructs a respiratory anesthesia knowledge retrieval question-answering engine with excellent performance, which accelerates the question-answering response speed in the field of respiratory anesthesia by fine-tuning the large model and actively integrating respiratory anesthesia professional knowledge.

[0038] 4. The respiratory anesthesia professional knowledge model of the present invention is customizable, and the knowledge base of the respiratory anesthesia professional knowledge model is adjusted to ensure that some professional knowledge that is not convenient to disseminate externally is disseminated within a set range.

[0039] 5. The respiratory anesthesia knowledge retrieval question-answering method and system proposed in the present invention provide efficient and intelligent question-answering generation and optimization services for the medical field. It can not only significantly improve the efficiency of professional knowledge retrieval, but also improve the accuracy, convenience and intelligence of the question-answering system by expanding knowledge supplementation or privatizing the large language model of respiratory anesthesia. In the future, this technology can be widely used in multiple sub-fields such as medical knowledge base construction, intelligent question-answering system, clinical auxiliary decision support, etc., and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a system architecture diagram of the present invention;

[0041] Figure 2 It is a work flow chart of the present invention.

[0042] Figure 3 It is a model structure diagram of the present invention. DETAILED DESCRIPTION

[0043] With the development of artificial intelligence technologies such as large language model technology, intelligent and data-driven applications in the medical field have gradually become mainstream. Among them, professional knowledge questions and answers in professional fields can greatly increase the work efficiency of researchers, doctors, and medical workers, and reduce their work error rate. The technology described in the present invention is a retrieval question-answering model in the field of respiratory anesthesia. In medical practice, doctors and researchers need to acquire professional knowledge efficiently and accurately, especially in the field of respiratory anesthesia, which involves complex clinical problems and a large amount of professional knowledge reserves. Traditional knowledge retrieval methods are usually difficult to meet the needs of professionalism and timeliness. Therefore, the intelligent question-answering system based on large language models has become an important way to solve this problem. The technology described in the present invention is dedicated to solving the problem of knowledge retrieval in the field of respiratory anesthesia medicine. By using professional respiratory anesthesia medical knowledge, data, expert consensus and other knowledge, the general large language model is fine-tuned so that this large model not only has general question-answering capabilities, but also has the ability to answer professional knowledge in the field of respiratory anesthesia medicine. In addition, the technology described in the present invention also provides a function for users to automatically update professional knowledge in the field of respiratory anesthesia medicine, that is, users can create and update new knowledge databases for professional documents on respiratory anesthesia such as new books and new papers that they want to use, thereby obtaining a new way of professional knowledge in the field of respiratory anesthesia medicine. The present invention also provides a dynamic knowledge updating technology. When at least one knowledge content about a respiratory anesthesia knowledge is retrieved, it is sorted by time and the latest knowledge is selected to be fed back to the user.

[0044] The present invention discloses a method for constructing a large language model based on respiratory anesthesia expertise, which aims to construct a question-answering large language model in the field of respiratory anesthesia based on user autonomous retrieval and autonomous maintenance by fine-tuning papers, expert consensus, and theoretical knowledge in the field of respiratory anesthesia expertise, and then updating the knowledge base of professional knowledge such as new books in this professional field, so as to provide users with accurate and rich knowledge retrieval and question-answering generation services.

[0045] This system uses advanced large language model technology to combine professional knowledge in the field of respiratory anesthesia with public data, and provides an efficient and accurate intelligent question-answering platform through model fine-tuning and question-answering generation optimization. The system architecture includes three core parts: respiratory anesthesia professional knowledge processing part, respiratory anesthesia knowledge retrieval question-answering engine and respiratory anesthesia question generation and optimization part. The whole system provides a complete set of intelligent question-answering services through the quantification, embedding, retrieval and generation of knowledge texts.

[0046] 1. Criteria for selecting large language models:

[0047] Model scale: Ensure large-scale parameters and efficient cross-semantic understanding capabilities.

[0048] Domain adaptability: Models pre-trained on medical text corpora (such as PubMed datasets, open source datasets, which can be directly described without paying attention to the source) are preferred.

[0049] Efficiency and deployment: If lightweight requirements are considered, models such as DistilBERT can be selected. In one embodiment, QWen is used, but it should be noted that it is not limited to this. For example, Baidu's Qianfan model and Tencent's Hunyuan Turbo can be used. Compared with other large models, QWen is a lightweight large model base model. When the answer accuracy is comparable, QWen occupies the least hardware resources, which is in line with the lightweight usage scenario in the hospital scenario.

[0050] 2. Network structure improvements and innovations:

[0051] Based on the enhanced Transformer network structure, the specific improvements are as follows:

[0052] Multi-head dynamic attention mechanism: Improve the traditional multi-head attention mechanism and use a dynamic weight adjustment mechanism to adjust the attention weight distribution according to specific keywords in the field of respiratory anesthesia.

[0053] Domain semantic adaptation layer: Add a dedicated domain embedding layer to perform semantic preprocessing on respiratory anesthesia-specific vocabulary in the input text to enhance the ability to capture professional terminology.

[0054] Bidirectional interaction module: Adding a bidirectional interaction module allows the model to establish a dynamic bidirectional relationship between questions and knowledge base, improving the quality of answer generation for complex questions.

[0055] Retrieval-enhanced generation module: Integrate the retrieval-based question-answering mechanism into the generation module to achieve explainability and accuracy of answer generation.

[0056] 3. Functional division corresponds to specific structure:

[0057] Text preprocessing: operations corresponding to segmentation and embedding processing (such as input vectorization).

[0058] Retrieval: RAG and vector search based engines to correlate knowledge base content.

[0059] Answer generation: Integrate generation strategies and attention mechanisms to output answers.

[0060] Feedback optimization: Dynamically adjust parameters to optimize the performance of the fine-tuned model.

[0061] 4. Detailed description of the specific structure of the fine-tuned respiratory anesthesia expertise model:

[0062] The respiratory anesthesia expertise model includes a retrieval and generation layer and an output processing layer.

[0063] The retrieval and generation layer includes a knowledge base retrieval module, a context integration module, and an answer generation module;

[0064] The knowledge base retrieval module is used to retrieve the pre-built respiratory anesthesia knowledge base through a vector search engine based on the vectorized problem representation to obtain the knowledge text related to the problem semantics;

[0065] The context integration module is used to integrate context information using a multi-head dynamic attention mechanism to generate the most relevant knowledge fragments;

[0066] The answer generation module is used to generate answers to questions using a retrieval-enhanced generation method and further improve the accuracy and relevance of the answers through a two-way interaction mechanism.

[0067] The respiratory anesthesia expertise model includes an output processing layer, which includes: an answer optimization module and an answer output module, wherein:

[0068] The answer optimization module is used to adjust or reconstruct the answers according to the similarity of the generated answers;

[0069] The answer output module is used to convert the adjusted or reconstructed answers into readable text.

[0070] The core technologies include the following aspects:

[0071] 1) The respiratory anesthesia expertise model is finally generated by annotating the public knowledge or local documents of respiratory anesthesia and constructing a dataset through fine-tuning the basic large language model;

[0072] 2) Perform text vectorization on respiratory anesthesia expertise and integrate it into the respiratory anesthesia expertise model through text embedding technology;

[0073] 3) Through the above two steps, a respiratory anesthesia knowledge retrieval question-answering engine is constructed to facilitate subsequent knowledge retrieval and question-answering;

[0074] 4) For questions related to respiratory anesthesia, we first convert the questions input by users into vector form through question text vectorization, and input the text vector into the respiratory anesthesia knowledge retrieval question-answering engine through text embedding technology to generate related texts, so as to perform corresponding semantic analysis and generate prompt words related to the questions to better understand and match the needs of users;

[0075] 5) Based on the prompt words and related texts generated above, the respiratory anesthesia knowledge retrieval question-answering engine will further generate prompt words related to the question. Finally, the respiratory anesthesia professional knowledge model after fine-tuning will construct a complete answer to the respiratory anesthesia question, thereby ensuring that the generated answer not only conforms to the semantic logic, but also can meet the high accuracy requirements in the field of respiratory anesthesia.

[0076] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0077] Example:

[0078] The embodiment of the present invention proposes a respiratory anesthesia knowledge retrieval question-answering method, which focuses on the field of respiratory anesthesia and aims to provide users with high-quality, professional answers to respiratory anesthesia-related questions. It integrates text processing technology, a powerful knowledge retrieval engine, and a professional large language model to ensure that user questions can be accurately understood, relevant knowledge can be efficiently retrieved, and professional answers with depth and accuracy can be generated.

[0079] The steps include:

[0080] Step 1) receiving a text of a respiratory anesthesia-related question;

[0081] Step 2) segmenting, vectorizing and text embedding the question text to obtain the corresponding vector;

[0082] Step 3) inputting a pre-established and trained respiratory anesthesia expertise model, the respiratory anesthesia expertise model generates related texts through vector search of a retrieval-type question-answering engine, generates question-related answer prompt words through context integration, and outputs the answer to the question based on similarity calculation;

[0083] The respiratory anesthesia expertise model is obtained by fine-tuning a basic large language model by integrating respiratory anesthesia expertise. The respiratory anesthesia expertise model is based on the network structure of an enhanced Transformer, adopts a multi-head dynamic attention mechanism, and increases domain semantic adaptation, two-way interaction, and retrieval enhancement.

[0084] This is explained in detail below.

[0085] 1. Overview of Workflow Architecture

[0086] like Figure 1 As shown in the figure, starting from the user's question about respiratory anesthesia, the text segmentation operation is first performed to split the question into smaller units that are easier to handle, which helps to grasp the core points and key information of the question more accurately. By converting the segmented text into a vector representation, it lays the foundation for subsequent knowledge retrieval and processing.

[0087] The respiratory anesthesia knowledge retrieval question-answering engine plays a key role in the entire process, searching deeply in the respiratory anesthesia knowledge base. Using advanced algorithms and technologies, it can quickly and accurately find knowledge fragments related to user questions, providing an important basis for generating preliminary answers. In this process, the vector search function ensures the comprehensiveness and depth of the retrieval, leaving out any information that may be relevant to the question.

[0088] The context integration link further improves the quality and completeness of the answer. By loading the respiratory anesthesia knowledge base, the system can obtain a wider range of relevant knowledge, combine it with the preliminary answer, and consider the problem from multiple angles. At the same time, the generated question-related answer prompts provide users with additional thinking directions and clues, helping users better understand the relationship between questions and answers.

[0089] As the core of the entire process, the respiratory anesthesia expertise model gives full play to its powerful language understanding and generation capabilities. It receives information that has been processed through contextual integration, and combines its own in-depth understanding and extensive knowledge in the field of respiratory anesthesia to conduct in-depth analysis and reasoning. Through the comprehensive use of various information, the respiratory anesthesia expertise model can generate more accurate, comprehensive and professional answers, providing users with high-quality knowledge services.

[0090] Finally, the system outputs the professional answers generated by the large language model to the user, meeting the user's knowledge needs for respiratory anesthesia-related issues. The various links in this workflow diagram work closely together to form an efficient and accurate knowledge processing and answer generation system, providing strong support for users to solve problems in the field of respiratory anesthesia.

[0091] 2. Description of process steps

[0092] like Figure 2 As shown,

[0093] 1. Input of questions related to respiratory anesthesia: Users raise questions related to respiratory anesthesia on the application side. The questions here can be various types of questions about various aspects such as the principles, operating methods, risk assessment, etc. of respiratory anesthesia, including general questions or some customized questions.

[0094] 2. Injection of vectors of segmented text to be processed: After the system receives the question, it first segments the question text. The purpose of text segmentation is to break down the question into smaller units in order to better understand the structure and key information of the question. For example, for the question "What are the applications and precautions of respiratory anesthesia in pediatric surgery?", it can be divided into several key parts such as "respiratory anesthesia", "pediatric surgery", "application", and "precautions". The segmented question text is converted into a vector representation and injected into the system. Vector representation allows computers to process and analyze text more efficiently. Through specific algorithms and techniques, the features of the text are extracted and converted into question vectors to prepare for subsequent retrieval and processing steps.

[0095] 3. Respiratory anesthesia knowledge retrieval question-answering engine: The injected vector will trigger the respiratory anesthesia knowledge retrieval question-answering engine to start working. This question-answering engine is the core part of the entire respiratory anesthesia professional knowledge model, and it is responsible for searching for information related to the question in the respiratory anesthesia knowledge base.

[0096] 4. Generate preliminary answers to questions: The Q&A engine will conduct in-depth searches in the professional knowledge base, using various algorithms and techniques to quickly find knowledge fragments related to the question. For example, for the above segmented questions, the Q&A engine may search the knowledge base for relevant content such as "Specific application cases of respiratory anesthesia in children's surgery" and "Precautions for respiratory anesthesia in children's surgery". Based on the retrieved knowledge fragments, the Q&A engine will try to generate a preliminary answer to the question. This preliminary answer may be based on a combination of multiple knowledge fragments, or the extraction and organization of a specific knowledge fragment.

[0097] 5. Generate prompt words for answers related to the question: In order to make the generated preliminary answer more accurate and comprehensive, the system will perform context integration. The process of context integration includes considering the problem from multiple angles, combining factors such as the background of the problem, related knowledge fields, and the specific needs of the user. At the same time, the system will also generate prompt words for answers related to the question. These prompt words can help users further understand the key to the problem, or provide some clues to guide users to think more deeply about the problem. For example, for the question "What are the applications and precautions of respiratory anesthesia in pediatric surgery?", the prompt words that may be generated include "children's physiological characteristics", "type of surgery", "anesthesia risk assessment", etc.

[0098] 6. Loading the respiratory anesthesia professional knowledge base: During the context integration process, the system will load the respiratory anesthesia knowledge base again to obtain more relevant knowledge and information. For example, if a certain respiratory anesthetic drug is mentioned in the preliminary answer, the system will obtain detailed information about the drug from the knowledge base, including its mechanism of action, side effects, scope of application, etc., in order to better answer the user's question.

[0099] 7. Respiratory anesthesia professional knowledge model: After context integration and prompt word generation, the processed information will be input into the respiratory anesthesia professional knowledge model. The model has strong language understanding and generation capabilities. It can conduct in-depth analysis and reasoning on the input information, combine its extensive knowledge in the field of respiratory anesthesia, and generate more accurate and comprehensive professional answers based on similarity calculations.

[0100] 8. Output of professional answers to questions related to respiratory anesthesia: Finally, the system outputs the professional answers generated by the respiratory anesthesia professional knowledge model to the user. The user can solve his or her own problem based on this answer, or further ask more in-depth questions, thus forming a continuous cycle of knowledge acquisition and problem solving.

[0101] 3. Overview of the overall working principle

[0102] The present invention presents an advanced knowledge retrieval and question-answering system specifically for the field of respiratory anesthesia. Its design purpose is to efficiently integrate various knowledge resources related to respiratory anesthesia and provide users with accurate, professional and timely answers to questions and valuable tips.

[0103] The starting point of the system is a rich data source, including public knowledge of respiratory anesthesia and local documents. These resources cover a wide range of respiratory anesthesia expertise, from basic theory to clinical practice, from the latest research results to traditional experience summary. Although the unstructured text data contains a lot of valuable information, it needs to be carefully processed before it can be effectively used by the system.

[0104] The "Unstructured Data Loader" is a relatively new technical means that is dedicated to the preliminary organization and structuring of the originally loose and disordered unstructured text data. By extracting key information, identifying important concepts and relationships, etc., it lays the foundation for subsequent in-depth analysis.

[0105] 4. Specific technical analysis

[0106] 1. Data loading and collection

[0107] The unstructured data loader uses a multi-channel and multi-mode data collection mechanism to efficiently load a variety of unstructured data sources in the field of respiratory anesthesia:

[0108] (1) Data sources: public anesthesia guidelines and field consensus documents; electronic medical records and anesthesia operation records within the hospital; scientific research papers and meeting minutes, etc.

[0109] (2) Collection method:

[0110] ① For different data source types (such as text documents, scanned images), dynamically select adapter loading to support incremental loading and historical data synchronization.

[0111] ② Use OCR technology to solve the recognition problem of scanned or image-based anesthesia records.

[0112] 2. Data preprocessing and cleaning

[0113] (1) Multi-level cleaning mechanism:

[0114] ① Noise filtering: remove meaningless characters (such as redundant punctuation marks) and irrelevant paragraphs (such as copyright information).

[0115] ②Semantic segmentation: Based on NLP analysis, complex text is broken down into logically clear paragraphs or sentences.

[0116] (2) Dynamic data standardization:

[0117] ① Standardize the time, units, terminology and other contents in the data to ensure uniformity (such as normalizing "mg" and "milligram" in the data into a unified format).

[0118] ② Design an adaptive stop word list to filter out domain-irrelevant words.

[0119] 3. Extraction of key information

[0120] The unstructured data loader uses a combination of deep learning and domain knowledge graphs to accurately extract and structure respiratory anesthesia expertise from text data:

[0121] (1) Named Entity Recognition:

[0122] ① Use a pre-trained model based on domain fine-tuning (Transformer is used in this invention, which is a general technology. This technology can be directly described without paying attention to the implementation) to identify specific terms in the field of respiratory anesthesia, such as "drug name", "diagnosis result", "operation steps", etc.

[0123] ②Dynamically expand entity categories to support the rapid introduction of emerging medical terms.

[0124] (2) Relationship extraction:

[0125] ① Extract the logical and semantic relationships between entities, such as "a certain drug is suitable for a certain disease" and "a certain operation corresponds to a risk warning".

[0126] ② Combine bidirectional LSTM (a general technology that can be directly described without paying attention to its implementation) and self-attention mechanism to improve the accuracy of semantic relationship recognition.

[0127] (3) High-value paragraph identification: Using a multi-task learning model, the text content is scored according to domain importance and high-value paragraphs are extracted first.

[0128] 4. Knowledge structuring and storage

[0129] (1) Data structuring: The extracted knowledge is stored in JSON format, including entity types, semantic relationships, key paragraphs, etc.

[0130] (2) Knowledge fusion: The extracted structured data are deduplicated and fused to generate a unified respiratory anesthesia knowledge base, which serves as the core data input for model training and fine-tuning.

[0131] (3) Real-time update mechanism: Combined with the knowledge traceability function, it automatically triggers fusion and update every time new data is loaded.

[0132] 5. Application in the respiratory anesthesia expertise model

[0133] The structured knowledge data constructed by the unstructured data loader can be directly used in the training, fine-tuning and question-answering engine construction of the respiratory anesthesia expertise model described in the present invention, mainly including the following applications:

[0134] (1) Model training: Provide high-quality, consistently labeled domain datasets to improve the professionalism and accuracy of respiratory anesthesia expertise models.

[0135] (2) Model fine-tuning: Quickly adapt to dynamic changes in the field (such as the introduction of new drugs or new technologies) by loading new knowledge.

[0136] (3) Question and answer engine support: Provide knowledge graph embedding for question and answer engines to enhance the engine’s semantic understanding and retrieval capabilities.

[0137] The text segmentation step breaks down the huge amount of unstructured text data into text blocks to be processed. This segmentation enables the system to process each small unit more carefully, improving the accuracy and efficiency of processing. At the same time, it also facilitates better matching and association with user questions in subsequent processes.

[0138] Text embedding converts the segmented text block data and the user's questions into vector representations. This vector representation can capture the semantic features of the text, so that similar texts have similar positions in the vector space. In this way, the system can quickly find the text content most relevant to the user's question by calculating the distance or similarity between vectors.

[0139] When a user asks a question related to respiratory anesthesia, the system first vectorizes the question, that is, generates a question vector. This vector is in the same vector space as the text block data vector stored in the vector memory, providing a basis for subsequent similarity calculations.

[0140] (1) Basic principles of text embedding technology

[0141] Text embedding technology converts text content into vector representations to capture its semantics, contextual associations, and domain characteristics. Through pre-training or fine-tuning of deep learning models, a semantically rich and structured vector space is constructed:

[0142] Text segmentation: Documents, guidelines, case records, etc. in the field of respiratory anesthesia are segmented based on paragraphs, sentences, and even terms to ensure the integrity of semantic units.

[0143] Embedding generation: This paper uses a Transformer-based deep learning model to convert segmented text blocks into semantic vectors to capture the semantic relationships between words, phrases, and paragraphs.

[0144] Vector space construction: Through model training, a multi-dimensional vector space is constructed, and semantically similar texts are closer in position in the space.

[0145] (2) Key steps of text embedding technology in the system

[0146] (a) Vectorization of respiratory anesthesia knowledge text

[0147] Semantic feature extraction:

[0148] An embedding model specifically for the respiratory anesthesia field is designed. In one embodiment, SciBERT, a pre-trained model based on the medical field, is fine-tuned.

[0149] Enhance the model's semantic understanding of medical terms (such as drug names, anesthesia procedures) to avoid misjudgment or semantic loss.

[0150] Multi-granularity embedding: Generate multi-granularity embedding vectors for data at different levels (such as word level, sentence level, and paragraph level) to improve the model's ability to express complex semantics.

[0151] Dynamic embedding adjustment: Dynamically adjust the embedding representation of the text according to the context of different questions to highlight key semantic features.

[0152] (b) Vectorization of User Questions

[0153] Question semantic parsing: Based on the questions input by the user, the deep learning model is used to decompose the questions into multi-layer semantic units (such as core goals and constraints).

[0154] Vector generation and enhancement: Generate vector representations for the multi-dimensional attributes of the problem (such as time, object, and operation), and enhance the understanding of the problem domain through a dedicated respiratory anesthesia knowledge model.

[0155] Semantic focus optimization: Use the attention mechanism to focus on the key terms and contextual relationships in the question and eliminate noise content.

[0156] (c) Similarity calculation and matching

[0157] Semantic similarity measurement: This invention uses the cosine similarity calculation method to calculate the similarity between the user question and the text vector in the knowledge base. According to different retrieval requirements, the matching algorithm is adaptively selected.

[0158] Multi-level matching strategy: After the initial matching, secondary screening is performed based on contextual information and domain rules to ensure the accuracy and professionalism of the returned results.

[0159] Retrieval optimization: In view of the dynamic changes in the embedding vector space, an online training and update mechanism is used to improve the real-time retrieval efficiency.

[0160] (3) Innovative Application of Text Embedding Technology

[0161] ①Knowledge retrieval engine

[0162] Knowledge fragment location: Based on the embedding vector of the user's question, quickly locate the most relevant text content in the knowledge base and generate answers based on high-confidence results.

[0163] Personalized recommendation: Based on the historical questions and behaviors input by users, the text embedding generation strategy is dynamically adjusted to provide personalized knowledge services.

[0164] ② Question and answer system optimization

[0165] Accurate answer generation: The high-dimensional representation of the embedding vector provides strong semantic support, ensuring that the generated answers conform to the semantic logic and domain rules of professional knowledge.

[0166] Multi-round interaction support: In response to user questions or ambiguous input, text embedding technology is used to dynamically adjust semantic understanding and provide more intelligent response capabilities.

[0167] ③ Update of knowledge on respiratory anesthesia

[0168] Real-time knowledge fusion: When new domain knowledge is added, the new knowledge is quickly integrated into the vector space through the embedding update mechanism to ensure the real-time nature of the knowledge base.

[0169] Semantic shift processing: Dynamically adjust the embedding vector in response to changes in the meaning of terms (such as new drugs, new operations), solving the problem that traditional systems are difficult to adapt to dynamic knowledge.

[0170] (4) The value and innovation of text embedding technology

[0171] Domain-customized model design: Combined with the characteristics of the respiratory anesthesia field, a domain-specific text embedding model is constructed.

[0172] Dynamic embedding generation and optimization: Improve the model's ability to understand complex problems through multi-granularity dynamic embedding mechanism and semantic focus optimization model.

[0173] Efficient semantic matching: Design multi-level matching strategies and fast nearest neighbor search technology to significantly improve retrieval efficiency and answer quality.

[0174] Real-time knowledge updating capability: Dynamic embedding adjustment and online training mechanisms are used to ensure the model’s adaptability to dynamic changes in domain knowledge.

[0175] The vector memory is the core storage component of the system. It stores the text block data vector and question vector after text embedding. After the user asks a question, the system will perform vector similarity calculation in the vector memory. By comparing the similarity between the question vector and each text block data vector, the system can accurately find the text block most relevant to the question.

[0176] The overall structure of the vector memory:

[0177] Vector memory is an efficient storage and query system designed specifically for semantic retrieval. It consists of the following main modules:

[0178] (1) Vector storage unit

[0179] ①Semantic vectors used to store text block data in the field of respiratory anesthesia.

[0180] ② Each semantic vector corresponds one-to-one with its original text block and associated metadata (such as source, timestamp, credibility, etc.), supporting the tracking of context and semantic background.

[0181] ③ The storage adopts a high-performance sparse vector format, and in one embodiment, ANN-ApproximateNearestNeighbors is used to improve computing efficiency.

[0182] (2) Vector Index Module

[0183] ① Based on advanced indexing algorithms, in one embodiment, HNSW-Hierarchical Navigable Small World is used to build a multi-level vector indexing system.

[0184] ② Index layering design:

[0185] Global index: used to quickly locate text blocks similar to the question vector.

[0186] Local indexing: refines the search scope to ensure accurate matching of complex semantic features.

[0187] (3) Similarity calculation module

[0188] ①Measure the semantic similarity between the user question vector and the stored vector by calculating the cosine similarity.

[0189] ②Support dynamic adjustment of weights to adapt to the specific semantic features of different questions.

[0190] (4) Vector dynamic update module

[0191] Realize real-time updates of stored content, including:

[0192] ① New vector: When new content is added to the knowledge base, a vector is automatically generated and written to the storage.

[0193] ② Vector correction: In response to changes in domain terminology or knowledge updates, the original vector is adjusted through model fine-tuning to ensure the semantic consistency of the system.

[0194] (5) Retrieval Optimization Module

[0195] ① For large-scale vector data storage, parallel retrieval and distributed architecture are adopted.

[0196] ② Provide batch query capabilities and support simultaneous access by multiple users.

[0197] Technical workflow of vector memory:

[0198] (1) Data preprocessing and vector generation

[0199] The system extracts raw data from the knowledge base in the field of respiratory anesthesia, including clinical guidelines, research literature, case data, etc.

[0200] Using the respiratory anesthesia-specific text embedding model, the text block data is converted into a high-dimensional semantic vector and stored in the vector storage unit.

[0201] (2) User Question Vectorization

[0202] The question input by the user generates a question vector through the text embedding model, which shares the same vector space with the semantic vector storing the text block to ensure the consistency of semantic representation.

[0203] (3) Similarity calculation and matching

[0204] The system performs the following steps in vector memory:

[0205] ① Retrieve the global neighbors of the user’s question vector and quickly locate potentially relevant text blocks.

[0206] ② Combine the local index to perform secondary screening of candidate vectors.

[0207] ③ Calculate the similarity and sort it, and output the text block data most relevant to the question.

[0208] (4) Dynamic feedback and optimization

[0209] It supports users to evaluate the results (such as confirmation, adjustment, and denial) and optimizes the similarity calculation model and indexing strategy through a feedback mechanism.

[0210] Innovation of vector memory:

[0211] As a part of data storage, vector memory integrates a number of innovative technologies such as intelligent semantic understanding, dynamic update, real-time retrieval, and multimodal fusion, which significantly improves the performance and intelligence level of the system.

[0212] High-performance semantic index design:

[0213] The introduction of a multi-layer index architecture combines global fast retrieval with local semantic fine matching, significantly improving retrieval speed and accuracy.

[0214] Supports dynamic expansion of domain terminology to enhance the model's adaptability to respiratory anesthesia expertise.

[0215] Dynamic vector update mechanism:

[0216] A knowledge base dynamic update and vector automatic adjustment strategy are proposed to enable the vector memory to adapt to the rapid changes of domain knowledge.

[0217] Supports incremental updates, avoids large-scale storage reconstruction, and saves computing resources.

[0218] Domain semantic optimization:

[0219] Introduce respiratory anesthesia-specific semantic feature extraction algorithms (such as vector correction methods combined with knowledge graphs) in storage and retrieval to solve the problem of semantic understanding of domain-specific terms.

[0220] Real-time retrieval and high concurrency support:

[0221] ① Adopt a distributed parallel architecture and combine it with ANN search technology to achieve millisecond-level retrieval of massive vector data.

[0222] ②Support concurrent access by multiple users to ensure real-time and stability in medical scenarios.

[0223] The application value of vector memory in the large model of respiratory anesthesia professional knowledge language:

[0224] Knowledge retrieval accuracy:

[0225] Through high-dimensional vector representation and multi-level indexing, the system can quickly extract text content that is highly relevant to user questions from the respiratory anesthesia knowledge base, significantly improving the accuracy of knowledge matching.

[0226] System scalability and adaptability:

[0227] The dynamic update mechanism ensures that the knowledge base is synchronized with the latest content in real time. At the same time, the adaptive adjustment capability of the semantic vector enables the system to cope with changes in domain terminology and the emergence of new knowledge.

[0228] Multi-scenario applicability:

[0229] The design of vector memory supports a variety of usage scenarios from single retrieval to complex question and answer, meeting the diverse needs of knowledge exploration, education, training, etc. in the field of respiratory anesthesia.

[0230] Semantic intelligence supports:

[0231] Vector memory not only realizes semantic retrieval, but also provides semantic feature support for subsequent question reasoning and answer generation. It is the core supporting technology of the large model of respiratory anesthesia professional knowledge language.

[0232] After finding the relevant vector-related text blocks, the prompt template is used to generate prompt words for respiratory anesthesia related questions. These prompt words can provide users with more thinking angles and clues, help users further refine the questions, or guide users to think about the questions from different aspects, thereby improving the accuracy and comprehensiveness of the answers to the questions.

[0233] In the stage of using the large language model, the relevant text blocks found are first processed with the help of the basic large language model. With its powerful language understanding and generation capabilities, the basic large language model can deeply analyze and reason about the text and generate preliminary answers. However, in order to better meet the needs of question answering in the field of respiratory anesthesia, the basic large language model is optimized through fine-tuning. The fine-tuning process adjusts the parameters of the model according to specific respiratory anesthesia tasks and data sets, making the model more accurate and efficient in dealing with respiratory anesthesia problems. After fine-tuning, a respiratory anesthesia expertise model specifically for the field of respiratory anesthesia was formed. This model combines respiratory anesthesia expertise with advanced language processing technology to generate more accurate and professional answers to questions.

[0234] Finally, the respiratory anesthesia expertise model is used to generate answers to user questions and return them to the user. Users can solve their own questions based on the answers, or ask more in-depth questions based on the answers, thus forming a continuous cycle and continuous improvement of knowledge retrieval and question-answering process. This system not only provides users with a convenient channel for knowledge acquisition, but also provides powerful tools for professionals and researchers in the field of respiratory anesthesia, promoting knowledge exchange and development in this field.

[0235] Example 2

[0236] Embodiment 2 of the present invention proposes a respiratory anesthesia knowledge retrieval question-answering system, which is implemented based on the method of embodiment 1. Figure 3 As shown, including:

[0237] A user interaction layer, used to receive texts of questions related to respiratory anesthesia;

[0238] Intelligent question-answering processing layer, used to segment, vectorize and embed question text to obtain corresponding vectors;

[0239] The intelligent model execution layer is used to input a pre-established and trained respiratory anesthesia professional knowledge model, the respiratory anesthesia professional knowledge model generates related texts through vector search of the retrieval-type question-answering engine, generates question-related answer prompt words through context integration, and outputs the answer to the question based on similarity calculation; and

[0240] The dynamic knowledge support layer is used to integrate respiratory anesthesia expertise to fine-tune the basic large language model to obtain a respiratory anesthesia expertise model, and is also used to dynamically update the respiratory anesthesia knowledge base in real time;

[0241] The respiratory anesthesia expertise model is based on the network structure of the enhanced Transformer, adopts a multi-head dynamic attention mechanism, and increases domain semantic adaptation, bidirectional interaction, and retrieval enhancement.

[0242] The respiratory anesthesia knowledge base is based on external knowledge retrieval and real-time knowledge update, realizing multi-dimensional knowledge fusion.

[0243] This system adopts a modular design and mainly includes three modules: respiratory anesthesia knowledge retrieval question-answering engine module, respiratory anesthesia professional knowledge processing module and respiratory anesthesia related question generation module. At the fine-tuning end represented by the respiratory anesthesia knowledge retrieval question-answering engine module, a high-quality respiratory anesthesia professional knowledge model is constructed by constructing a data set of existing respiratory anesthesia professional knowledge and fine-tuning the basic large language model. Through this engine, answers to respiratory anesthesia related questions can be generated quickly and conveniently; at the local knowledge base end represented by the respiratory anesthesia professional knowledge processing module, professional knowledge and public documents are processed and connected to the large language model to build and update the knowledge base; and at the application end represented by the respiratory anesthesia related question generation module, the respiratory anesthesia professional knowledge question-answering service is realized through question vectorization, retrieval and answer generation optimization to ensure that the answers to questions meet actual application needs. The entire system has a strict logical process and clear technical steps from professional knowledge processing, model fine-tuning to knowledge base construction and question and answer generation, ensuring the accuracy and real-time nature of question and answer generation.

[0244] Each module is described in detail below.

[0245] 1. Respiratory anesthesia knowledge retrieval question-answering engine module: This module is used to construct a data set for existing respiratory anesthesia expertise and fine-tune the basic large language model to build a high-quality respiratory anesthesia expertise model. Through this engine, answers to respiratory anesthesia-related questions can be generated quickly and easily.

[0246] ● Professional knowledge dataset: The system first uses technical processing to generate a respiratory anesthesia professional knowledge dataset that can be used to train a basic large language model.

[0247] Fine-tuning of the basic large language model: The system uses a basic large language model that contains extensive language knowledge and fine-tunes it through the professional data in the field of respiratory anesthesia generated above. The fine-tuning process includes the annotation and dataset construction of public and local documents to ensure that the generated language model can accurately reflect the knowledge system in the field of respiratory anesthesia, which can obtain a professional knowledge model of respiratory anesthesia.

[0248] 2. Respiratory anesthesia professional knowledge processing module: This module is the knowledge base of the system. By processing professional knowledge and public documents, it connects to the large language model to build and update the knowledge base.

[0249] Knowledge text vectorization: The system first performs text vectorization on literature, guidelines, and documents in the field of respiratory anesthesia. Through calculation, the knowledge structure in the text is converted into vector form, which is convenient for embedding in the model.

[0250] Text embedding: By using a pre-trained language model, the vectorized text is embedded so that it can be represented in a high-dimensional vector space to form an embedding vector of professional knowledge.

[0251] Large language model fusion: The embedded vectors will be fused into the respiratory anesthesia expertise model for subsequent use.

[0252] 3. Respiratory anesthesia question answer generation module: Based on the generation of the question and answer engine, through question vectorization, retrieval and answer generation optimization, respiratory anesthesia professional knowledge question and answer service is realized to ensure that the answers to the questions meet the actual application needs.

[0253] Question text vectorization: When a user asks a question related to respiratory anesthesia, the system first vectorizes the question. By embedding the question into the same vector space, it can be efficiently matched and retrieved in the model.

[0254] Knowledge retrieval engine: After the question text is embedded, it is input into the respiratory anesthesia knowledge retrieval question-answering engine. The engine retrieves the most relevant knowledge to the question through vector similarity calculation. The knowledge retrieval is based on the data in the respiratory anesthesia knowledge base to ensure that the extracted content has medical expertise.

[0255] Generate relevant text: After the knowledge retrieval is completed, the system will generate relevant text that is highly matched to the question, providing users with timely, accurate and professional answers.

[0256] · Question-related prompt word generation: During the question-answer generation process, the system generates a series of prompt words related to the question by analyzing the semantic structure of the user's question. These prompt words help the system better understand user needs and generate more accurate answers by optimizing the matching degree of the question.

[0257] Answer generation: Based on the generated prompt words and related knowledge text, the system uses the fine-tuned respiratory anesthesia expertise model to generate the final answer. This process ensures that the answer is not only reasonable in medical semantic logic, but also highly accurate and meets the needs of clinical practice.

[0258] From the perspective of knowledge sources, the system integrates respiratory anesthesia expertise, including public respiratory anesthesia knowledge and professional content in local documents, forming a rich professional knowledge data set. This knowledge is the basic support for the entire system and provides an authoritative reference for answering various complex respiratory anesthesia-related questions.

[0259] In terms of technical implementation, we first convert respiratory anesthesia knowledge texts from different sources and questions raised by users into computer-processable digital vector forms through knowledge text vectorization and question text vectorization steps. This process enables the original text content to be accurately quantified and analyzed, laying the foundation for subsequent processing steps.

[0260] Then, the text embedding operation is performed to map the vectorized knowledge text and question text into a low-dimensional space, which improves the computational efficiency and accuracy of subsequent processing while retaining its semantic information. This step makes similar texts have a close position relationship in the embedding space, which is convenient for subsequent retrieval and matching.

[0261] The basic large language model plays a core role in the system. It uses its powerful language understanding and generation capabilities to deeply analyze and process the embedded text. This model can understand the meaning of the text, make inferences, and generate preliminary answers.

[0262] The fine-tuning end makes targeted adjustments and optimizations to the basic large language model. By fine-tuning it according to specific respiratory anesthesia tasks and datasets, the model is made more adaptable to the problem-solving needs of this professional field, improving its performance and accuracy.

[0263] The fusion step organically combines the fine-tuned model with respiratory anesthesia expertise. This ensures that the generated answers have both the intelligence and flexibility of the model and the accuracy and authority of the expertise.

[0264] The respiratory anesthesia knowledge retrieval question-answering engine built on this basis is the core part of the entire system for users. It can receive respiratory anesthesia-related questions input by users, quickly find relevant knowledge texts through efficient retrieval and matching algorithms, and generate accurate answers to questions using the fused model. At the same time, it can also generate related texts and question-related prompts to provide users with more information and guidance, helping users better understand questions and answers.

[0265] As an important part of the system, the local knowledge base stores a wealth of respiratory anesthesia expertise and documents. During the system operation, it can provide the original knowledge source for steps such as knowledge text vectorization at any time, ensuring that the system can obtain the latest expertise in a timely manner and provide users with the most accurate answers.

[0266] The application end is the interface for users to interact with the system. It can be presented in various forms such as web pages and mobile applications, providing users with convenient channels for inputting questions and obtaining answers, allowing users to easily use this professional respiratory anesthesia knowledge retrieval question and answer engine.

[0267] In summary:

[0268] 1. Domain-customized respiratory anesthesia expertise model and precise semantic vector embedding enable the relevant vector representation to not only capture the semantic information of general medical knowledge, but also to carefully express the complex medical terms, treatment methods and latest clinical practices related to respiratory anesthesia;

[0269] 2. Multimodal data fusion and cross-modal retrieval, so that users can input text descriptions or upload images, charts, etc. The system can perform effective vector retrieval and answer generation, which enhances the flexibility and applicability of the system;

[0270] 3. Incremental update mechanism and real-time retrieval optimization allow the system to quickly incorporate the latest knowledge when faced with new clinical research, anesthesia equipment updates or new treatment plans, without affecting the query efficiency of existing data;

[0271] 4. Intelligent knowledge recommendation and automatic reasoning capabilities. With the help of vector storage, the system can accurately recommend relevant respiratory anesthesia knowledge based on the input query content and historical question and answer records. In addition, it can also deduce possible treatment methods, anesthetic use plans, emergency treatment plans, etc. through the similarity between vectors, thereby improving the intelligence level of the system.

[0272] These innovative technologies can enable the respiratory anesthesia knowledge retrieval question-answering system to respond more accurately and quickly to complex problems faced by clinicians, anesthesiologists and other professionals in their actual work, and provide real-time and reliable support.

[0273] Based on the specific deployment, it includes the fine-tuning end, the local knowledge base end, and the application end.

[0274] 1. Fine-tuning end

[0275] First, collect public knowledge about respiratory anesthesia. The public knowledge here includes a series of public knowledge about respiratory anesthesia that can be recognized by professionals, such as papers, expert consensus, annual reports, and doctor experience. After obtaining this knowledge, it is necessary to annotate this knowledge and then generate a professional knowledge data set. This data set is like a huge knowledge treasure house that stores various digitally processed professional knowledge. It provides a rich data source for subsequent steps, allowing the system to retrieve and use relevant knowledge to answer users' questions. The basic large language model used in the present invention can be the Alibaba open source large language model QWen as the basic model. This basic large language model has powerful language understanding and generation capabilities. It can analyze the semantic structure of the text, extract key words, and generate semantic information. Information, and reasoning and generating preliminary answers based on this information, for example, when the model receives the embedded question text, it will try to generate a preliminary answer framework based on its knowledge understanding of the field of respiratory anesthesia; the above-mentioned basic large language model is further adjusted and optimized through the fine-tuning end; the fine-tuning process is to make subtle adjustments to the parameters of the model based on the above-mentioned respiratory anesthesia professional knowledge dataset to improve the performance and accuracy of the model in answering questions in this professional field, for example, by training on a large number of respiratory anesthesia-related questions and answer pairs, adjusting the model's weight parameters, so that the model can better adapt to the needs of answering respiratory anesthesia questions; after the above series of steps, a professional knowledge large language model specifically for the field of respiratory anesthesia is formed. This model integrates respiratory anesthesia expertise, advanced language processing technology, and intelligent question-answering engine functions, and can provide users with high-quality respiratory anesthesia question answering services.

[0276] 2. Local knowledge base

[0277] In today's society, medical technology is developing rapidly, and a lot of new expertise in respiratory anesthesia is generated every day. This respiratory anesthesia expertise can also be carried out in local documents, which can provide a customized content function for administrators of the entire system. Because fine-tuning the respiratory anesthesia expertise prediction model is costly, the cost here includes the cost of annotating the data set and the time cost. These more difficult problems can be perfectly solved through the steps described below.

[0278] First, we extract text content from the source of respiratory anesthesia expertise, including public knowledge of respiratory anesthesia and local documents. Through vectorization technology, we convert these expertise texts into digital vector representations. For texts describing the "mechanism of action of respiratory anesthesia", after vectorization, they will become a set of multi-dimensional digital vectors, in which each dimension represents a specific feature of the text, such as vocabulary frequency, semantic relationship, etc.; secondly, we perform text embedding operations on the vectorized knowledge texts and question texts. The text embedding process is like mapping these digital vectors into a low-dimensional semantic space, in which similar texts will be close to each other. The advantages of this are, on the one hand, it can reduce the complexity of calculation and improve the efficiency of subsequent processing; on the other hand, it can better retain the semantic information of the text, making the system more accurate in retrieval and matching. For example, after text embedding, the positions of knowledge text and question text related to "risk of respiratory anesthesia" in the embedding space will be closer, which is convenient for the system to quickly find relevant knowledge to answer questions; finally, the fine-tuned respiratory anesthesia expertise model is integrated with the respiratory anesthesia expertise. The main function of this step is to inject expertise into the model, making the answers generated by the model more accurate and authoritative.

[0279] 3. Application

[0280] In a series of steps described on this side, similar to the local knowledge base side, the respiratory anesthesia-related questions input by the user are vectorized through calculation. This step is similar to the text vectorization on the local knowledge base side, but there is no process of extracting text from the text file; then, through the text embedding process, the text vectors generated in the above steps are input into the respiratory anesthesia knowledge retrieval-type question-and-answer engine constructed by the respiratory anesthesia professional knowledge language large model; while generating answers to questions, the question-and-answer engine will also generate related texts related to the questions based on the content of the questions and the related information of the answers. These related texts can provide users with more background information and reference materials to help users better understand questions and answers. For example, if the user's question is about the use of a certain respiratory anesthetic drug, The Q&A engine may generate some related text about the side effects, precautions, etc. of the drug, so that users can have a more comprehensive understanding of the drug; in addition, the Q&A engine will also generate prompt words related to the question, which can help users further refine the question or provide more clues to obtain more accurate answers. For example, if the user's question is vague, the Q&A engine may generate some prompt words, such as "specific symptoms" and "patient age", to guide the user to provide a more specific description of the problem; when the user asks a question on the application side, the system will generate answers to questions related to respiratory anesthesia through the above process and return them to the user. The user can solve his or her own problem based on the answer, or further ask more in-depth questions, thus forming a continuous cycle of knowledge acquisition and problem-solving process.

[0281] Application scenarios:

[0282] 1. Medical knowledge retrieval: In actual clinical applications, doctors can use this system to quickly retrieve professional knowledge in the field of respiratory anesthesia, obtain authoritative and detailed answers, and assist in clinical decision-making.

[0283] 2. Medical education: The system can not only provide rich learning resources for medical students and researchers, but also help users understand and master the core concepts and cutting-edge knowledge in the field of respiratory anesthesia through interactive question-and-answer methods.

[0284] 3. Scientific research support: In the research of respiratory anesthesia-related topics, researchers can use the system to efficiently obtain the latest knowledge and literature related to their topics, saving literature retrieval time and improving scientific research efficiency.

[0285] 4. Customized respiratory anesthesia knowledge base: In practical applications, the present invention can update the knowledge base of the respiratory anesthesia professional knowledge model through the processing method of the respiratory anesthesia professional knowledge processing module, and perform customized processing of a series of professional respiratory anesthesia knowledge. By expanding this knowledge base, some professional knowledge that is not convenient for external dissemination can be disseminated in a small range within a certain range, which can increase the degree of protection of intellectual property rights.

[0286] Invention point:

[0287] Fine-tuning technology is used to fine-tune the basic large language model to improve the model's ability to handle professional knowledge issues in respiratory anesthesia, paving the way for subsequent steps in the present invention.

[0288] The integration of respiratory anesthesia expertise can update the knowledge base of the respiratory anesthesia expertise large model in near real time through the designated processing of literature, guidelines and documents in the professional field, ensuring the accuracy of the system's answers.

[0289] Build a high-performance respiratory anesthesia knowledge retrieval question-answering engine, and speed up the question-answering response speed of the question-answering system in the field of respiratory anesthesia by fine-tuning the large model and actively integrating respiratory anesthesia professional knowledge.

[0290] The respiratory anesthesia professional knowledge model is customizable. For some expert experience and other knowledge that can only be disseminated within a small range, the knowledge base of the respiratory anesthesia professional knowledge model can be adjusted through the processing method of the above-mentioned respiratory anesthesia professional knowledge processing module to ensure that some professional knowledge that is not convenient for external dissemination is disseminated within a small range within a certain range, which can increase the degree of protection of intellectual property rights.

[0291] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A respiratory anesthesia knowledge retrieval question-answering method, comprising: Step 1) receiving a text of a respiratory anesthesia-related question; Step 2) segmenting, vectorizing and text embedding the question text to obtain the corresponding vector; Step 3) inputting a pre-established and trained respiratory anesthesia expertise model, the respiratory anesthesia expertise model generates related texts through vector search of a retrieval-type question-answering engine, generates question-related answer prompt words through context integration, and outputs the answer to the question based on similarity calculation; The respiratory anesthesia expertise model is obtained by fine-tuning a basic large language model by integrating respiratory anesthesia expertise. The respiratory anesthesia expertise model is based on the network structure of an enhanced Transformer, adopts a multi-head dynamic attention mechanism, and increases domain semantic adaptation, two-way interaction, and retrieval enhancement.

2. The respiratory anesthesia knowledge retrieval question-answering method according to claim 1, characterized in that: The issues related to respiratory anesthesia in step 1) include: the principles of respiratory anesthesia, equipment operation methods, risk assessment and cutting-edge knowledge of respiratory anesthesia.

3. The respiratory anesthesia knowledge retrieval question-answering method according to claim 1, characterized in that: The step 2) includes: using an adaptive text segmentation method to segment the question text into multiple text blocks, and converting them into question vector representations, generating question vectors through a text embedding model, and using a dimensionality reduction algorithm to map the question vectors to a low-dimensional semantic space.

4. The respiratory anesthesia knowledge retrieval question-answering method according to claim 1, characterized in that: The basic large language model is a large language model model, and the respiratory anesthesia professional knowledge model includes a retrieval and generation layer; The retrieval and generation layer includes a knowledge base retrieval module, a context integration module and an answer generation module; wherein, The knowledge base retrieval module is used to retrieve a pre-built respiratory anesthesia knowledge base through a vector search engine based on the vectorized problem representation to obtain knowledge text related to the problem semantics; The context integration module is used to integrate context information using a multi-head dynamic attention mechanism to generate the most relevant knowledge fragments; The answer generation module is used to generate answers to questions using a retrieval-enhanced generation method, and further improve the accuracy and relevance of the answers through a two-way interactive mechanism.

5. The respiratory anesthesia knowledge retrieval question-answering method according to claim 4, characterized in that: The respiratory anesthesia expertise model includes an output processing layer, which includes: an answer optimization module and an answer output module, wherein: The answer optimization module is used to adjust or reconstruct the answers according to the similarity of the generated answers; The answer output module is used to convert the adjusted or reconstructed answers into readable text.

6. The respiratory anesthesia knowledge retrieval question-answering method according to claim 4, characterized in that: The method also includes real-time dynamic updating of the respiratory anesthesia knowledge base, and further optimizing the vector search engine of the knowledge base retrieval module based on the updated respiratory anesthesia knowledge base by embedding an updating mechanism.

7. The respiratory anesthesia knowledge retrieval question-answering method according to claim 6, characterized in that: The method also includes providing a customized function for the respiratory anesthesia knowledge base, so as to further expand the updated content of the respiratory anesthesia knowledge base according to user authority.

8. The respiratory anesthesia knowledge retrieval question-answering method according to claim 1, characterized in that: The process of fine-tuning the respiratory anesthesia expertise model included: Collect public knowledge on respiratory anesthesia and annotate it to obtain a professional knowledge dataset; Collect local documents of professional knowledge and extract text content; Performing vectorization processing to obtain a set of multi-dimensional digital vectors, wherein each dimension represents a feature of the text, and the features include: vocabulary frequency and semantic relationship; Perform text embedding to map digital vectors into a low-dimensional semantic space; The mapped digital vector is input into the basic large language model, and the parameters are modulated until the training requirements are met to obtain a trained respiratory anesthesia expertise model.

9. The respiratory anesthesia knowledge retrieval question-answering method according to claim 8, characterized in that: The public knowledge includes papers, expert consensus, annual reports and physician experience.

10. A respiratory anesthesia knowledge retrieval question-answering system, characterized in that: include: A user interaction layer, used to receive texts of questions related to respiratory anesthesia; Intelligent question-answering processing layer, used to segment, vectorize and embed question text to obtain corresponding vectors; The intelligent model execution layer is used to input a pre-established and trained respiratory anesthesia professional knowledge model. The respiratory anesthesia professional knowledge model generates related texts through vector search of a retrieval-type question-answering engine, generates question-related answer prompt words through context integration, and outputs the answer to the question based on similarity calculation; and The dynamic knowledge support layer is used to integrate respiratory anesthesia expertise to fine-tune the basic large language model to obtain a respiratory anesthesia expertise model, and is also used to dynamically update the respiratory anesthesia knowledge base in real time; The respiratory anesthesia expertise model is based on the network structure of the enhanced Transformer, adopts a multi-head dynamic attention mechanism, and increases domain semantic adaptation, bidirectional interaction, and retrieval enhancement.

Images