Question answering prediction method for large oral medical model based on knowledge retrieval enhancement

Abstract

This invention discloses a question-answering prediction method for a large-scale oral healthcare model based on knowledge retrieval enhancement. The method includes: S1 receiving a user-inputted question and converting it into a semantic vector representation using a pre-trained model; S2 obtaining the semantic vector representation of each trigger keyword in the knowledge base and calculating the similarity between the semantic vector representation in step S1 and the semantic vector representation of each trigger keyword; S3 determining whether there exists a similarity greater than or equal to a preset threshold among all similarities; if so, proceeding to step S4; otherwise, using the user-inputted question as the model input question; S4 concatenating the trigger keywords corresponding to the similarity that meets the condition with their structured prompts and the user-inputted question as the model input question; and S5 inputting the model input question into the large-scale oral healthcare model for prediction to obtain the question response.

Description

Technical Field

[0001] This invention belongs to the field of medical decision support, specifically involving a question-answering prediction method for a large-scale oral medical model based on knowledge retrieval enhancement. Background Technology

[0002] Large-scale pre-trained language models have achieved significant results in tasks such as medical question answering and intelligent triage. However, in medical question answering, there are still some questions that cannot be accurately understood or answered in practical applications, especially long-tail questions, marginal questions, or cases where users express themselves vaguely. This can cause the model's output to deviate from the user's actual needs. For example, in a large-scale oral healthcare model, when a user inputs their oral health-related questions, hoping to know the exact name of the department to register for and the operating instructions, the prediction process often fails because the user is unfamiliar with some professional terms, leading to vague expressions and causing the large model to provide inaccurate department recommendations, thus affecting the user's medical treatment.

[0003] To improve the overall system experience, common optimization methods include: continuing to fine-tune the large model, building FAQ templates, supplementing rules, or using manual prompts. However, these methods have drawbacks such as poor versatility, high maintenance costs, or incomplete coverage of new issues, making them difficult to effectively handle dynamically updated user feedback.

[0004] Therefore, there is an urgent need for a general, efficient, and scalable lightweight enhancement mechanism to address occasional "unanswerable questions" in models, while also having good engineering adaptability and practical effectiveness. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, the oral medical large-scale model question-answering prediction method based on knowledge retrieval enhancement provided by this invention can introduce information most relevant to the user's question as contextual prompts before generation, effectively compensating for knowledge blind spots within the generated model.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0007] A question-answering prediction method for a large-scale oral healthcare model based on knowledge retrieval enhancement is provided, which includes the following steps:

[0008] S1. Receive user input questions and use a pre-trained model to convert them into semantic vector representations;

[0009] S2. Obtain the semantic vector representation of each trigger keyword in the knowledge base, and calculate the similarity between the semantic vector representation in step S1 and the semantic vector representation of each trigger keyword.

[0010] S3. Determine whether there is a similarity greater than or equal to a preset threshold among all similarities. If there is, proceed to step S4; otherwise, use the user's input question as the model input question.

[0011] S4. Concatenate the trigger keywords and their structured prompts corresponding to the similarity that meet the conditions with the user's input question to form the model input question;

[0012] S5. Input the problem into the oral healthcare big data model for prediction and obtain the problem response.

[0013] Furthermore, the preset threshold is obtained using an adaptive threshold adjustment mechanism:

[0014] S31. Based on the user's input question, predict the complexity of the user's input question using a logistic regression classifier:

[0015]

[0016] ,

[0017]

[0018] in, The complexity of the questions input by the user; For questions entered by the user; This is a logistic regression classifier; n is the total number of features; b is the bias obtained during training; Semantic density; It is an L2 norm; For the number of entities; From The set of keywords identified in the text; For syntactic complexity; For analysis The grammatical relationships between words in the text generate a tree-like structure that reflects the logical structure of the sentence; , and These are the functions corresponding to the 1st, 2nd, and 3rd features, respectively;

[0019] S32. Calculate a preset threshold based on the complexity of the user-input question:

[0020]

[0021] in, The preset threshold; and These are the lower threshold and the upper threshold, respectively; The complexity of the question for optimal user input.

[0022] Furthermore, the expression for the knowledge base is:

[0023] ,

[0024] Where K represents the knowledge base; For natural language sentence space; This is the i-th trigger keyword; for The corresponding structured prompt information; N is the total number of triggering keywords.

[0025] Furthermore, the knowledge retrieval-enhanced question-answering prediction method for large-scale oral healthcare models also includes updating the knowledge base at preset intervals:

[0026] S21. Within the current preset time period, based on errors predicted by the model or user feedback, the oral healthcare big data model is found to have deviations in the actual question-and-answer process;

[0027] S22. Based on the user-inputted questions that show deviations, compare the question responses generated by the corresponding large-scale oral healthcare model with the expert-annotated answers to obtain the answer deviation degree:

[0028]

[0029] in, is the confidence level of the model in its generated question response 'a'; is the question input by the user. The semantic vector distribution of the response 'a' to the question; To verify the distribution of answers for experts; and All are weighting coefficients; Kullback-Leibler divergence measures the difference between two probability distributions and is used to quantify the semantic deviation between the generated answer and the standard answer.

[0030] S23. Update the knowledge base based on the answer deviation degree and deviation threshold:

[0031]

[0032] in, and These are the knowledge bases before and after the update, respectively. Annotate the answers for experts; This is the deviation threshold; For existing query-answer pairs in the knowledge base; This is the expert-corrected query-answer pair;

[0033] S24. When updating the knowledge base, dynamically calculate the structured hint information added to the knowledge base. The edge:

[0034]

[0035] in, This is the newly added set of edges; The existing structured prompts in the knowledge base; The weight of the edge; To and Existing entities with the top-k similarity; This is an edge weight prediction model based on graph neural networks; The edges, or relationships, between structured prompts.

[0036] Furthermore, when cache space is insufficient, it also includes updating the knowledge base:

[0037] Calculate the weighted score for each triggering keyword in the knowledge base:

[0038]

[0039] in, For the knowledge base Number of visits within the time window; for Hours since the last visit;

[0040] Retain trigger keywords in the database that have a retention time less than or equal to the minimum retention time, as well as trigger keywords whose retention time is greater than the minimum retention time and whose weighted score is greater than the preset score.

[0041] Furthermore, the pre-trained model construction method includes: fine-tuning the SentenceTransformer embedding model in the open-source library sentence-transformers using a dataset composed of oral medical-related literature and data; during the fine-tuning process, the maximum mean difference is used to measure the general text domain. With the field of oral healthcare Distribution differences:

[0042]

[0043] in, For general text fields With the field of oral healthcare The distribution differences; It is a feature mapping function; and They are respectively and The number of samples; and Each represents a single sample in the dataset; This is used to calculate the difference between the means of the embedding distributions of the two domains.

[0044] Furthermore, the expression for calculating similarity is:

[0045]

[0046] in, The transpose of the semantic vector representation of the question input by the user; T is the transpose symbol; For the i-th trigger keyword in the knowledge base semantic vector representation, ; To determine the absolute value sign; For taking the mold;

[0047] Model input problem in step S4 The expression is:

[0048] + User-input question x

[0049]

[0050] in, The set consisting of triggering keywords and their structured prompts corresponding to the similarity of the conditions; The preset threshold; for The corresponding structured prompt information.

[0051] Furthermore, the pre-trained model will incorporate the user-input question. and each trigger keyword in the knowledge base Mapped to a normalized semantic vector representation:

[0052] x), ( ), =1

[0053] in, For vectorized embedding encoders; for The corresponding semantic vector representation; for The corresponding semantic vector representation.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] 1. This solution proposes a question enhancement strategy, namely, the model input question obtained by concatenation. This allows for the intelligent selection and introduction of an appropriate amount of structured knowledge based on the type, difficulty, and contextual information of the user's question. For simple questions, it can provide concise answers, while for complex questions, it will inject more professional knowledge and background information to ensure the accuracy and comprehensiveness of the answer.

[0056] 2. This solution innovatively introduces a semantic knowledge base, specifically designed to supplement and strengthen the knowledge gaps of language models when dealing with specific domains (such as medical issues). Unlike traditional knowledge bases, this solution's semantic knowledge base is not merely a static collection of entries, but a dynamically updated and context-aware adaptive library; it can respond in real time to specific needs in user input and is continuously expanded and optimized through expert feedback.

[0057] 3. This solution not only enhances cross-contextual knowledge matching capabilities but also proposes a similarity retrieval method, making semantic retrieval more efficient and accurate. Compared to traditional keyword-based retrieval methods, this invention, through deep learning technology, can understand the semantic similarity between different expressions, contexts, and surrounding texts.

[0058] 4. Compared with traditional methods based on FAQ templates or rule systems, this invention uses a knowledge base to search for relevant trigger keywords, i.e., it adopts a deep semantic matching method to solve the problem of missing coverage caused by expression variations. It has stronger generalization ability and can adapt to different user expression styles. It can automatically identify synonyms, context changes and question phrasing differences, thereby covering a wider range of question scenarios and avoiding the problem of insufficient coverage in traditional methods.

[0059] 5. Regarding the fine-tuning of pre-trained models, compared to methods that directly fine-tune large models, this invention does not require changes to the main model structure and parameters. It provides enhancements to the original large model through a domain knowledge base and adaptive prompting mechanism, without the need for complex fine-tuning and retraining. This approach reduces manual intervention and computational resource requirements, thereby significantly reducing development costs and time, and improving the system's adaptability and maintenance efficiency.

[0060] 6. Validation in a real-world medical question-and-answer system shows that the accuracy and user satisfaction of this solution are significantly higher than traditional methods in medical question-and-answer tasks. Especially when dealing with complex medical problems, the system can provide more accurate diagnostic and treatment suggestions. This achievement demonstrates that the solution of this invention not only possesses high technical advancement but also significant commercial application potential. Attached Figure Description

[0061] Figure 1 This is a flowchart of a knowledge retrieval-enhanced question-answering prediction method for a large-scale oral healthcare model. Detailed Implementation

[0062] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0063] refer to Figure 1 , Figure 1 The flowchart illustrates a question-answering prediction method for a large-scale oral healthcare model based on knowledge retrieval enhancement; for example... Figure 1 As shown, the method S includes steps S1 to S5.

[0064] In step S1, the user input question is received and converted into a semantic vector representation using a pre-trained model; in implementation, the preferred method for constructing the pre-trained model in this scheme includes:

[0065] The SentenceTransformer embedding model in the open-source library sentence-transformers was fine-tuned using a dataset consisting of oral medical literature and data; during the fine-tuning process, the maximum mean difference was used to measure the general text domain. With the field of oral healthcare Distribution differences:

[0066]

[0067] in, For general text fields With the field of oral healthcare The distribution differences; It is a feature mapping function; and They are respectively and The number of samples; and Each represents a single sample in the dataset; This is used to calculate the difference between the means of the embedding distributions of the two domains.

[0068] SentenceTransformer is a text vector encoding framework trained on BERT and its derivative models (such as RoBERTa, DistilBERT, MiniLM, etc.). It is specifically designed to map sentences, phrases, or paragraphs in natural language into high-dimensional vector representations in semantic space. Compared to the original BERT model, it can efficiently handle sentence-level semantic comparisons and is suitable for tasks such as question-answering matching, semantic search, and clustering.

[0069] To improve the adaptability of the model, this invention employs domain adaptation technology. By fine-tuning on oral medical literature and data, the semantic embedding model can focus on specific knowledge in that domain, such as oral diseases, treatment methods, and dental drugs.

[0070] This approach uses a vectorized embedding encoder based on a pre-trained model (such as Sentence-BERT). , will user input and each in the knowledge base Mapped to a normalized semantic vector representation:

[0071] x), ( ), =1

[0072] The encoder obtains a unified embedding space through training, where and The vector lengths are all 1, and text pairs with similar semantics are closer in distance. After normalization, the similarity can be simplified to the vector dot product.

[0073] In step S2, the semantic vector representation of each triggering keyword in the knowledge base is obtained, and the similarity between the semantic vector representation in step S1 and the semantic vector representation of each triggering keyword is calculated.

[0074] The preferred expression for the knowledge base in this solution is:

[0075] ,

[0076] Where K represents the knowledge base; For natural language sentence space; This is the i-th trigger keyword; for The corresponding structured prompt information; N is the total number of triggering keywords.

[0077] The knowledge base provided in this solution is used to cover typical problem areas where the oral healthcare big model performs poorly. The knowledge base is organized in key-value pair format: K is the extracted high-frequency keywords in the domain or error-prone keywords in the oral healthcare big model, and v is the standard answer or guidance suggestion that has been manually confirmed or verified by experts.

[0078] In this scheme, the expression for calculating similarity is:

[0079]

[0080] in, The transpose of the semantic vector representation of the question input by the user; T is the transpose symbol; For the i-th trigger keyword in the knowledge base semantic vector representation, ; To determine the absolute value sign; To calculate the difference between the means of the embedding distributions of the two domains, the optimization objective is to fine-tune to make... , Reduce by more than 50%.

[0081] By reducing distributional differences across domains, this approach improves the performance of pre-trained models in the field of oral medicine, ensuring that the models can be customized and fine-tuned to meet the specific needs of oral healthcare. This is particularly evident when dealing with medical terminology and complex questions, where the solution demonstrates greater accuracy in understanding and matching. This technology adapts to the semantic features of different medical fields, rapidly enhancing the performance of pre-trained models in oral healthcare.

[0082] In step S3, it is determined whether there is a similarity greater than or equal to a preset threshold among all similarities. If there is, proceed to step S4; otherwise, the user-input question is used as the model input question.

[0083] In one embodiment of the present invention, the preset threshold is obtained using an adaptive threshold adjustment mechanism:

[0084] S31. Based on the user's input question, predict the complexity of the user's input question using a logistic regression classifier:

[0085]

[0086] ,

[0087]

[0088] in, The complexity of the questions input by the user; For questions entered by the user; This is a logistic regression classifier; n is the total number of features; b is the bias obtained during training; Semantic density; It is an L2 norm; For the number of entities; From The set of keywords identified in the text; For syntactic complexity; For analysis The grammatical relationships between words in the text generate a tree-like structure that reflects the logical structure of the sentence; , and These are the functions corresponding to the 1st, 2nd, and 3rd features, respectively;

[0089] S32. Calculate a preset threshold based on the complexity of the user-input question:

[0090]

[0091] in, The preset threshold; and These are the lower threshold and the upper threshold, respectively; The complexity of the question for optimal user input.

[0092] This solution introduces an adaptive threshold adjustment mechanism, which dynamically adjusts the preset threshold for similarity matching, thereby flexibly balancing matching accuracy and efficiency according to actual scenarios or user needs. Specifically, this solution uses a deep learning model to progressively optimize the preset threshold setting based on the semantic complexity or background information of the user input, enabling the solution to automatically adjust its matching strategy in different environments. For example, for some simple queries, the system may choose a lower similarity threshold to increase the number of matching candidates; while for complex medical questions, the threshold can be increased to ensure more accurate matching. This not only improves matching accuracy but also enhances efficiency in large-scale knowledge bases.

[0093] In step S4, the trigger keywords corresponding to the similarity scores that meet the conditions, along with their structured prompts, are concatenated with the user-inputted question to form the model input question; the model input question... The expression is:

[0094] + User-input question x

[0095]

[0096] in, The set consisting of triggering keywords and their structured prompts corresponding to the similarity of the conditions; The preset threshold; for The corresponding structured prompt information.

[0097] The final input question in this solution consists of several knowledge items, which are concatenated into the original question and then input into a large-scale oral healthcare model for response generation. This approach is equivalent to performing a "refined retrieval + information injection" before generation, significantly improving the model's response quality to the target question. The final concatenated complete input is then directly sent as prompts to the large-scale oral healthcare model (such as a medical question-answering model) for answer generation. By adding highly relevant structured knowledge prompts before the original question, the large-scale model can refer to this information during the inference phase, thereby improving performance.

[0098] In step S5, the model input problem is input into the oral healthcare big data model for prediction, and the problem response is obtained.

[0099] In implementation, this solution preferably utilizes a knowledge retrieval-enhanced oral healthcare large-scale model question-answering prediction method, which also includes updating the knowledge base at preset intervals.

[0100] S21. Within the current preset time period, based on errors predicted by the model or user feedback, the oral healthcare big data model is found to have deviations in the actual question-and-answer process;

[0101] S22. Based on the user-inputted questions that show deviations, compare the question responses generated by the corresponding large-scale oral healthcare model with the expert-annotated answers to obtain the answer deviation degree:

[0102]

[0103] in, is the confidence level of the model in its generated question response 'a'; is the question input by the user. The semantic vector distribution of the response 'a' to the question; To verify the distribution of answers for experts; and All are weighting coefficients; Kullback-Leibler divergence measures the difference between two probability distributions and is used to quantify the semantic deviation between the generated answer and the standard answer.

[0104] S23. Update the knowledge base based on the answer deviation degree and deviation threshold:

[0105]

[0106] in, and These are the knowledge bases before and after the update, respectively. Annotate the answers for experts; This is the deviation threshold; For existing query-answer pairs in the knowledge base; This is the expert-corrected query-answer pair;

[0107] S24. When updating the knowledge base, dynamically calculate the structured hint information added to the knowledge base. The edge:

[0108]

[0109] in, This is the newly added set of edges; The existing structured prompts in the knowledge base; The weight of the edge; To and Existing entities with the top-k similarity; This is an edge weight prediction model based on graph neural networks; The edges, or relationships, between structured prompts.

[0110] When a large-scale oral healthcare model encounters a problem that it cannot solve or has low accuracy, experts can directly input verified answers or modification suggestions. These suggestions are then added to the knowledge base and used for the next round of model training or updates. This dynamic review and feedback mechanism ensures that the knowledge base can be continuously optimized and updated during application. Especially in professional fields such as medicine, the knowledge base must remain up-to-date and accurate to enhance the system's adaptability and long-term applicability.

[0111] Meanwhile, this solution also introduces semantic cache hot-reloading technology for incremental updates, meaning that the knowledge base is updated even when cache space is insufficient—an improved LFU cache update strategy. Traditional LFU is a classic cache eviction algorithm used to determine which data should be removed when cache space is insufficient. Its core idea is to prioritize eviction of entries with the lowest access frequency. However, relying solely on access frequency (Freq) may lead to: long-term hot topics occupying the cache for extended periods (such as "how to register for an appointment"), squeezing out space for new entries; and sudden high-frequency queries due to sudden traffic surges (such as trending searches for a specific disease) failing to respond quickly.

[0112] To address this, this solution also includes updating the knowledge base when cache space is insufficient:

[0113] Calculate the weighted score for each triggering keyword in the knowledge base:

[0114]

[0115] in, For the knowledge base Number of visits within the time window; for Hours since the last visit;

[0116] Retain trigger keywords in the database that have a retention time less than or equal to the minimum retention time, as well as trigger keywords whose retention time is greater than the minimum retention time and whose weighted score is greater than the preset score.

[0117] Since this invention is intended for medical applications, the parameters are designed as follows: Time window (Matching hospital consultation cycle), minimum retention time (A temporary cache for emergency-related queries).

[0118] Assume the cache contains the following entries:

[0119]

[0120] Updating the knowledge base based on weighted scoring has the following effects:

[0121] 1) The ranking of emergency room queries (low Freq but new) will improve, and traditional LFU will eliminate them.

[0122] 2) Long-term, low-frequency but important medical entries (such as "diabetic tooth extraction") will not be permanently retained.

[0123] This solution updates the knowledge base using the methods described above, achieving targeted enhancement of question-and-answer performance without requiring fine-tuning of the main model parameters. It exhibits good scalability and versatility.

[0124] In summary, this solution calculates similarity between the user-input question and the semantic vector representation of triggering keywords in the knowledge base, achieving high-precision matching, automatically concatenating prompt words, and guiding the large model to generate more accurate answers. Compared to traditional keyword matching or hard-coded rule methods, this solution has stronger fault tolerance and generalization capabilities, achieving intelligent enhancement without modifying the large model itself.

Claims

1. A question-answering prediction method for a large-scale oral healthcare model based on knowledge retrieval enhancement, characterized in that, Including the following steps: S1. Receive user input questions and use a pre-trained model to convert them into semantic vector representations; S2. Obtain the semantic vector representation of each trigger keyword in the knowledge base, and calculate the similarity between the semantic vector representation in step S1 and the semantic vector representation of each trigger keyword. S3. Determine whether there is a similarity greater than or equal to a preset threshold among all similarities. If there is, proceed to step S4; otherwise, use the user's input question as the model input question. S4. Concatenate the trigger keywords and their structured prompts corresponding to the similarity that meet the conditions with the user's input question to form the model input question; S5. Input the model problem into the large-scale oral healthcare model for prediction and obtain the problem response; The preset threshold is obtained using an adaptive threshold adjustment mechanism: S31. Based on the user's input question, predict the complexity of the user's input question using a logistic regression classifier: ， in, The complexity of the questions input by the user; For questions entered by the user; This is a logistic regression classifier; n is the total number of features; for The weight matrix; for The h-th feature; b is the bias obtained during training; Semantic density; It is an L2 norm; For the number of entities; From The set of keywords identified in the text; For syntactic complexity; For analysis The grammatical relationships between words in the text generate a tree-like structure that reflects the logical structure of the sentence; , and These are the functions corresponding to the 1st, 2nd, and 3rd features, respectively; S32. Calculate a preset threshold based on the complexity of the user-input question: in, The preset threshold; and These are the lower threshold and the upper threshold, respectively; The complexity of the question for optimal user input.

2. The knowledge retrieval-enhanced question-answering prediction method for a large-scale oral healthcare model according to claim 1, characterized in that, The expression for the knowledge base is: , Where K represents the knowledge base; For natural language sentence space; This is the i-th trigger keyword; for The corresponding structured prompt information; N is the total number of triggering keywords.

3. The knowledge retrieval-enhanced question-answering prediction method for a large-scale oral healthcare model according to claim 2, characterized in that, This also includes updating the knowledge base at preset intervals: S21. Within the current preset time period, based on errors predicted by the model or user feedback, the oral healthcare big data model is found to have deviations in the actual question-and-answer process; S22. Based on the user-inputted questions that show deviations, compare the question responses generated by the corresponding large-scale oral healthcare model with the expert-annotated answers to obtain the answer deviation degree: in, is the confidence level of the model in its generated question response 'a'; is the question input by the user. The semantic vector distribution of the response 'a' to the question; The distribution of answers was verified by experts; and All are weighting coefficients; Kullback-Leibler divergence measures the difference between two probability distributions and is used to quantify the semantic deviation between the generated answer and the standard answer. S23. Update the knowledge base based on the answer deviation degree and deviation threshold: in, and These are the knowledge bases before and after the update, respectively. Annotate the answers for experts; This is the deviation threshold; For existing query-answer pairs in the knowledge base; The query-answer pairs have been corrected by experts. S24. When updating the knowledge base, dynamically calculate the structured hint information added to the knowledge base. The edge: in, This is the newly added set of edges; The existing structured prompts in the knowledge base; The weight of the edge; To and Existing entities with the top-k similarity; This is an edge weight prediction model based on graph neural networks; The edges, or relationships, between structured prompts.

4. The knowledge retrieval-enhanced question-answering prediction method for a large-scale oral healthcare model according to claim 2, characterized in that, When cache space is insufficient, it also includes updating the knowledge base: Calculate the weighted score for each triggering keyword in the knowledge base: in, For the knowledge base Number of visits within the time window; for Hours since the last visit; Retain trigger keywords in the database that have a retention time less than or equal to the minimum retention time, as well as trigger keywords whose retention time is greater than the minimum retention time and whose weighted score is greater than the preset score.

5. The knowledge retrieval-enhanced question-answering prediction method for a large-scale oral healthcare model according to claim 1, characterized in that, The pre-trained model construction method includes: fine-tuning the SentenceTransformer embedding model in the open-source library sentence-transformers using a dataset consisting of oral medical literature and data; during the fine-tuning process, the maximum mean difference is used to measure the general text domain. With the field of oral healthcare Distribution differences: in, For general text fields With the field of oral healthcare The distribution differences; It is a feature mapping function; and They are respectively and The number of samples; and Each represents a single sample in the dataset; This is used to calculate the difference between the means of the embedding distributions of the two domains.

6. The knowledge retrieval-enhanced question-answering prediction method for a large-scale oral healthcare model according to any one of claims 1-5, characterized in that, The expression for calculating similarity is: in, The transpose of the semantic vector representation of the question input by the user; T is the transpose symbol; For the i-th trigger keyword in the knowledge base semantic vector representation, ; To determine the absolute value sign; This is used to calculate the difference between the means of the embedding distributions of the two domains; Model input problem in step S4 The expression is: + User-input question x in, The set consisting of triggering keywords and their structured prompts corresponding to the similarity of the conditions; The preset threshold; for The corresponding structured prompt information.

7. The knowledge retrieval-enhanced question-answering prediction method for a large-scale oral healthcare model according to any one of claims 1-5, characterized in that, The pre-trained model will take the user-input question and each trigger keyword in the knowledge base Mapped to a normalized semantic vector representation: x), ( ), =1 in, For vectorized embedding encoders; for The corresponding semantic vector representation; for The corresponding semantic vector representation.

Images