Semantic analysis and atlas reasoning fused medical question and answer retrieval method and system

Abstract

The invention relates to the field of medical information processing, in particular to a medical question and answer retrieval method and system fusing semantic analysis and map reasoning. The method comprises the following steps: acquiring a context sequence feature based on a natural question and an FGM to execute a decoding operation, and obtaining an entity recognition result; meanwhile, on the basis of a natural question, obtaining a global representation vector marked by the CLS to execute a context semantic feature extraction operation, and obtaining an intention recognition result; according to the entity recognition result and the intention recognition result, a relation path reasoning operation is executed in a medical knowledge graph to obtain a graph retrieval result, and when matched target information is not obtained through the relation path reasoning operation, a fusion retrieval enhancement generation mechanism is executed, and the graph retrieval result is obtained from an external medical corpus; and based on a preset natural language template and a semantic slot filling strategy, obtaining response content according to the atlas retrieval result. And the ability of the medical question and answer retrieval system in processing complicated scenes such as unknown questions, fuzzy questions and open questions is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information processing, and in particular to a medical question and answer retrieval method and system fusing semantic analysis and graph reasoning. BACKGROUND

[0002] A medical question and answer system based on natural language processing is an intelligent system that retrieves or generates accurate answers from a medical knowledge base or text by understanding natural language medical questions raised by users. Its core goal is to provide efficient and reliable medical information query services to assist patients, doctors or researchers in quickly obtaining professional medical knowledge. Traditional medical question and answer systems mainly achieve basic information retrieval through keyword matching, rule templates or simple classification models. However, when dealing with complex semantics, unstructured expressions or ambiguous queries, such medical question and answer systems have the problem of insufficient accuracy and robustness, making it difficult to meet the diverse needs of medical consultation.

[0003] For a medical question and answer system combining a pre-trained language model and a knowledge graph, compared with traditional medical question and answer systems, it has higher question understanding ability and structured representation of medical knowledge describing semantic associations between entities, achieving better question and answer retrieval and relationship reasoning. However, the coordination of entity recognition and graph reasoning is prone to analysis failure or incomplete results when facing non-standard questions, insufficient graph coverage or unclear query intent, etc.

[0004] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The main purpose of the present application is to provide a medical question and answer retrieval method and system fusing semantic analysis and graph reasoning, aiming to solve the problem of easy analysis failure or incomplete results of medical question and answer systems when facing non-standard questions, insufficient graph coverage or unclear query intent, etc.

[0006] To achieve the above purpose, the present application provides a medical question and answer retrieval method fusing semantic analysis and graph reasoning, which comprises the following steps:

[0007] Performing a named entity recognition operation and a semantic relationship extraction operation on a medical data set to construct a medical knowledge graph;

[0008] Upon receiving a natural question, obtaining context sequence features based on the natural question and FGM, and performing a decoding operation according to the context sequence features to obtain a globally optimal label sequence as an entity recognition result; at the same time,

[0009] Upon receiving the natural question, the global representation vector of the CLS tag is obtained based on the natural question, and the context semantic feature extraction operation is performed based on the global representation vector of the CLS tag to obtain the target query relationship type as the intent recognition result;

[0010] Based on the entity recognition results and intent recognition results, a relational path reasoning operation is performed in the medical knowledge graph to perform semantic query and obtain target information matching the natural question as the graph retrieval result. When the relational path reasoning operation fails to obtain the target information matching the natural question, a fusion retrieval enhancement generation mechanism is executed to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result.

[0011] Based on a preset natural language template and semantic slot filling strategy, the graph retrieval results are transformed to obtain the response content corresponding to the natural question.

[0012] Optionally, based on the medical dataset, cleaning operations, text parsing operations, and structured extraction operations are performed to clean, standardize, and semantically classify the unstructured information in the medical dataset to obtain medical entities and semantic relationships. The medical entities include diseases, symptoms, drugs, and departments, and the semantic relationships include disease-symptom, disease-drug, and disease-department.

[0013] The medical knowledge graph is constructed based on the medical entities and the semantic relationships.

[0014] Optionally, the natural question is passed to the Chinese-RoBERTa-wwm-ext model to obtain the first context embedding vector;

[0015] Based on the first context embedding vector, a gradient-based perturbation is introduced to obtain the second context embedding vector.

[0016] The second context embedding vector is passed to a BiLSTM network for bidirectional encoding to obtain the context sequence features;

[0017] Based on the multi-head attention mechanism and CRF, the context sequence features are decoded to obtain the globally optimal label sequence.

[0018] Optionally, the natural question is passed to the RoBERTa semantic representation layer for overall semantic encoding to extract the global representation vector of the CLS tag;

[0019] The global representation vector of the CLS tag is passed to the TextCNN network to perform contextual semantic feature extraction to obtain the intent classification probability distribution;

[0020] The target query relationship type is determined based on the intent classification probability distribution.

[0021] Optionally, a query path is constructed based on the entity-relationship-entity structure. In the medical knowledge graph, a relationship path reasoning operation is performed on the entity recognition result and the intent recognition result to obtain target information that matches the natural question.

[0022] Optionally, a keyword-based inverted index structure can be constructed for fuzzy entity matching; and,

[0023] An extended retrieval operation is performed based on a similarity algorithm that calculates a weighted average of edit distance and Jaccard coefficient.

[0024] Optionally, when the target information obtained by the relational path reasoning operation is empty, or when the confidence of the target information obtained by the relational path reasoning operation is less than a preset confidence threshold, a fusion retrieval enhancement generation mechanism is executed to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result.

[0025] Furthermore, to achieve the above objectives, the present invention also provides a medical question-answering retrieval system that integrates semantic parsing and graph reasoning, characterized in that the medical question-answering retrieval system integrating semantic parsing and graph reasoning includes:

[0026] The knowledge graph construction module is used to perform named entity recognition and semantic relation extraction operations on medical datasets to build a medical knowledge graph.

[0027] The entity recognition module is used to obtain context sequence features based on the natural question and FGM when a natural question is received, and to perform a decoding operation based on the context sequence features to obtain the globally optimal label sequence as the entity recognition result;

[0028] The intent recognition module is used to obtain the global representation vector of the CLS tag based on the natural question when the natural question is received, and to perform a context semantic feature extraction operation based on the global representation vector of the CLS tag to obtain the target query relationship type as the intent recognition result.

[0029] The graph reasoning module is used to perform semantic queries by performing relational path reasoning operations in the medical knowledge graph based on the entity recognition results and intent recognition results, and to obtain target information matching the natural question as the graph retrieval result. When the relational path reasoning operation fails to obtain the target information matching the natural question, a fusion retrieval enhancement generation mechanism is executed to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result.

[0030] The answer generation module is used to transform the graph retrieval results based on a preset natural language template and semantic slot filling strategy to obtain the response content corresponding to the natural question.

[0031] Furthermore, to achieve the above objectives, the present invention also provides a medical question-answering retrieval device that integrates semantic parsing and graph reasoning. The medical question-answering retrieval device that integrates semantic parsing and graph reasoning includes a memory, a processor, and a medical question-answering retrieval program that integrates semantic parsing and graph reasoning, stored in the memory and executable on the processor. When the medical question-answering retrieval program that integrates semantic parsing and graph reasoning is executed by the processor, it implements the steps of the medical question-answering retrieval method that integrates semantic parsing and graph reasoning as described above.

[0032] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a medical question-answering retrieval program that integrates semantic parsing and graph reasoning. When the medical question-answering retrieval program that integrates semantic parsing and graph reasoning is executed by a processor, it implements the steps of the medical question-answering retrieval method that integrates semantic parsing and graph reasoning as described above.

[0033] This invention provides a medical question-answering retrieval method that integrates semantic parsing and graph reasoning. It constructs a standardized medical knowledge graph covering eight types of entities, including diseases, symptoms, drugs, and examinations, and their semantic relationships. In the semantic parsing stage, a named entity recognition model, CRoBERTa-BLA-FGM-CRF, is built using Chinese-RoBERTa-wwm-ext as a shared encoder, combined with a BiLSTM multi-head attention mechanism and a CRF decoder. FGM adversarial training is introduced to enhance robustness against misspellings and terminology variations. Simultaneously, a RoBERTa model combined with TextCNN is used to recognize user intent, extracting the query target and relationship type from natural language questions to form a graph retrieval path. A dynamic question-answering mechanism based on entity matching confidence is introduced. If the matching degree is higher than a set threshold, a structured answer is generated based on graph reasoning operations using relationship path reasoning; otherwise, a retrieval enhancement generation mechanism is invoked to integrate semantic retrieval and a language model to generate the answer. This system not only achieves higher accuracy in medical entity recognition and intent classification, and stronger interpretability and structured expression capabilities at the graph reasoning level, but also demonstrates significant advantages in handling complex scenarios such as unknown questions, fuzzy questions, and open-ended questions. The introduction of the RAG mechanism compensates for the limited coverage of medical knowledge graphs, forming a closed-loop complementary mechanism of structured reasoning and semantic generation, effectively improving the intelligence, applicability, and robustness of the question-answering system. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the description, serve to explain the principles of the invention. To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0035] Figure 1 This is a schematic diagram of the hardware operating environment of the medical question-answering retrieval device that integrates semantic parsing and graph reasoning, as described in an embodiment of the present invention.

[0036] Figure 2 This is a flowchart illustrating an embodiment of the medical question-answering retrieval method integrating semantic parsing and graph reasoning of the present invention.

[0037] Figure 3 This is a schematic diagram of the architecture of the medical question-answering retrieval system that integrates semantic parsing and graph reasoning according to the present invention.

[0038] Figure 4 This is a schematic diagram of the architecture of an embodiment of the medical question-answering retrieval system integrating semantic parsing and graph reasoning of the present invention.

[0039] Figure 5 This is a schematic diagram of the architecture of another embodiment of the medical question-answering retrieval system that integrates semantic parsing and graph reasoning according to the present invention.

[0040] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0041] This application presents a medical question-answering retrieval method integrating semantic parsing and graph reasoning. It constructs a medical knowledge graph by performing named entity recognition and semantic relation extraction operations on a medical dataset. Upon receiving a natural question, it obtains context sequence features based on the natural question and the FGM (Focused Generalized Knowledge Graph), and performs decoding operations based on these context sequence features to obtain the globally optimal label sequence as the entity recognition result. Simultaneously, upon receiving the natural question, it obtains the global representation vector of the CLS (Common Language Subject) tag based on the natural question, and performs context semantic feature extraction operations based on the global representation vector of the CLS tag to obtain the target query relation type. The system performs semantic queries based on the entity recognition and intent recognition results, performing relational path reasoning operations within the medical knowledge graph to obtain target information matching the natural question as the graph retrieval result. If the relational path reasoning operation fails to obtain the target information matching the natural question, a fusion retrieval enhancement generation mechanism is executed to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result. Based on a preset natural language template and semantic slot filling strategy, the graph retrieval result is transformed to obtain the response content corresponding to the natural question. This improves the medical question-answering retrieval system's ability to handle complex scenarios such as unknown questions, fuzzy questions, and open-ended questions.

[0042] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0043] As one implementation scheme, Figure 1 This is a schematic diagram of the hardware operating environment of the medical question-answering retrieval device that integrates semantic parsing and graph reasoning, which is involved in the embodiments of the present invention.

[0044] like Figure 1As shown, the medical question-answering retrieval device integrating semantic parsing and graph reasoning may include: a processor 101, such as a central processing unit (CPU), a memory 102, and a communication bus 103. The memory 102 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. Optionally, the memory 102 may also be a storage device independent of the aforementioned processor 101. The communication bus 103 is used to enable communication between these components.

[0045] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on a medical question-answering retrieval device that integrates semantic parsing and graph reasoning. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0046] like Figure 1 As shown, the memory 102, which is a computer-readable storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and a medical question-and-answer retrieval method program that integrates semantic parsing and graph reasoning.

[0047] exist Figure 1 In the medical question-answering retrieval device integrating semantic parsing and graph reasoning shown, the processor 101 and the memory 102 can be installed in the medical question-answering retrieval device integrating semantic parsing and graph reasoning. The medical question-answering retrieval device integrating semantic parsing and graph reasoning calls the medical question-answering retrieval program integrating semantic parsing and graph reasoning stored in the memory 102 through the processor 101 and performs the following operations:

[0048] Perform named entity recognition and semantic relation extraction operations on the medical dataset to construct a medical knowledge graph;

[0049] Upon receiving a natural language question, context sequence features are obtained based on the natural language question and the FGM (Form Generator Generalized System), and decoding is performed based on these context sequence features to obtain the globally optimal label sequence as the entity recognition result; simultaneously,

[0050] Upon receiving the natural question, the global representation vector of the CLS tag is obtained based on the natural question, and the context semantic feature extraction operation is performed based on the global representation vector of the CLS tag to obtain the target query relationship type as the intent recognition result;

[0051] Based on the entity recognition results and intent recognition results, a relational path reasoning operation is performed in the medical knowledge graph to perform semantic query and obtain target information matching the natural question as the graph retrieval result. When the relational path reasoning operation fails to obtain the target information matching the natural question, a fusion retrieval enhancement generation mechanism is executed to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result.

[0052] Based on a preset natural language template and semantic slot filling strategy, the graph retrieval results are transformed to obtain the response content corresponding to the natural question.

[0053] In one embodiment, the processor 101 can be used to invoke a medical question-answering retrieval program that integrates semantic parsing and graph reasoning, stored in the memory 102, and perform the following operations:

[0054] Based on the medical dataset, cleaning, text parsing, and structured extraction operations are performed to clean, standardize, and semantically classify the unstructured information in the medical dataset to obtain medical entities and semantic relationships. The medical entities include diseases, symptoms, drugs, and departments, and the semantic relationships include disease-symptom, disease-drug, and disease-department.

[0055] The medical knowledge graph is constructed based on the medical entities and the semantic relationships.

[0056] In one embodiment, the processor 101 can be used to invoke a medical question-answering retrieval program that integrates semantic parsing and graph reasoning, stored in the memory 102, and perform the following operations:

[0057] The natural language question is passed to the Chinese-RoBERTa-wwm-ext model to obtain the first context embedding vector;

[0058] Based on the first context embedding vector, a gradient-based perturbation is introduced to obtain the second context embedding vector.

[0059] The second context embedding vector is passed to a BiLSTM network for bidirectional encoding to obtain the context sequence features;

[0060] Based on the multi-head attention mechanism and CRF, the context sequence features are decoded to obtain the globally optimal label sequence.

[0061] In one embodiment, the processor 101 can be used to invoke a medical question-answering retrieval program that integrates semantic parsing and graph reasoning, stored in the memory 102, and perform the following operations:

[0062] The natural question is passed to the RoBERTa semantic representation layer for overall semantic encoding, and the global representation vector of the CLS tag is extracted.

[0063] The global representation vector of the CLS tag is passed to the TextCNN network to perform contextual semantic feature extraction to obtain the intent classification probability distribution;

[0064] The target query relationship type is determined based on the intent classification probability distribution.

[0065] In one embodiment, the processor 101 can be used to invoke a medical question-answering retrieval program that integrates semantic parsing and graph reasoning, stored in the memory 102, and perform the following operations:

[0066] Based on the entity-relationship-entity structure, a query path is constructed. In the medical knowledge graph, a relationship path reasoning operation is performed on the entity recognition result and the intent recognition result to obtain the target information that matches the natural question.

[0067] In one embodiment, the processor 101 can be used to invoke a medical question-answering retrieval program that integrates semantic parsing and graph reasoning, stored in the memory 102, and perform the following operations:

[0068] Construct a keyword-based inverted index structure for fuzzy entity matching; and,

[0069] An extended retrieval operation is performed based on a similarity algorithm that calculates a weighted average of edit distance and Jaccard coefficient.

[0070] In one embodiment, the processor 101 can be used to invoke a medical question-answering retrieval program that integrates semantic parsing and graph reasoning, stored in the memory 102, and perform the following operations:

[0071] When the target information obtained by the relational path reasoning operation is empty, or when the confidence of the target information obtained by the relational path reasoning operation is less than a preset confidence threshold, the fusion retrieval enhancement generation mechanism is executed to retrieve the target content matching the natural question from the external medical corpus as the graph retrieval result.

[0072] Based on the hardware architecture of the medical question-answering retrieval device that integrates semantic parsing and graph reasoning, an embodiment of the medical question-answering retrieval method integrating semantic parsing and graph reasoning of the present invention is proposed.

[0073] Reference Figure 2 In the first embodiment, the medical question-answering retrieval method integrating semantic parsing and graph reasoning includes the following steps:

[0074] Step S100: Perform named entity recognition and semantic relationship extraction operations on the medical dataset to construct a medical knowledge graph.

[0075] In this embodiment, the medical dataset used in constructing the medical knowledge graph can be publicly available medical data resources. That is, based on publicly available medical datasets, text parsing and structured extraction techniques are employed to clean and standardize the original medical entities and semantic relationships, constructing a standardized medical knowledge graph with multiple semantic relationships. The medical knowledge graph uses unified data modeling specifications to abstractly represent entity nodes and relationship edges, thereby improving the standardization, consistency, and scalability of the graph structure. The medical knowledge graph can cover various semantic relationships such as disease and symptom, disease and department, disease and drug, and disease and food.

[0076] As an optional implementation, step 100 above includes performing cleaning, text parsing, and structured extraction operations based on the medical dataset to clean, standardize, and semantically classify the unstructured information in the medical dataset to obtain medical entities and semantic relationships. The medical entities include diseases, symptoms, drugs, and departments, and the semantic relationships include disease-symptom, disease-drug, and disease-department relationships. Then, the medical knowledge graph is constructed based on the medical entities and the semantic relationships.

[0077] Specifically, based on publicly available medical data resources, a standardized knowledge graph covering six types of entities—diseases, symptoms, examinations, drugs, food, and treatments—and various semantic relationships is constructed through data cleaning, named entity recognition, and semantic relation extraction methods. A graph-structure optimization-oriented modeling approach is used to abstract attributes for each entity and normalize the representation of relation edges, constructing a graph triplet structure.<h, r, t> Simultaneously, graph databases, such as Neo4j, are introduced to provide efficient storage and query support, thereby improving the response speed of graph retrieval.

[0078] The purpose of this approach is to utilize publicly available medical datasets as a data foundation during the knowledge graph construction phase. By combining text parsing and structured extraction techniques, unstructured information in the original medical corpus is cleaned, standardized, and semantically categorized. Through identifying medical entities such as diseases, symptoms, examinations, medications, food, and treatments, and mining the relationships between them (e.g., "disease-symptom," "disease-medication," "disease-department"), a standardized medical knowledge graph with multiple semantic types, multiple node categories, and an extensible structure is constructed. This medical knowledge graph adopts unified modeling specifications, providing a high-quality knowledge foundation for subsequent question-answering parsing and relational reasoning.

[0079] Step S200: Upon receiving a natural question, obtain context sequence features based on the natural question and FGM (Fast Gradient Method), and perform decoding operation based on the context sequence features to obtain the globally optimal label sequence as the entity recognition result.

[0080] In this embodiment, an entity recognition module that integrates a pre-trained language model, a bidirectional long short-term memory network, and a multi-head attention mechanism is used to achieve semantic modeling and contextual feature extraction of natural language questions, accurately identifying medical entities and their categories as input for graph reasoning. The overall design of this entity recognition module can be collectively referred to as the CroBERTa-BLA-FGM-CRF model.

[0081] The aforementioned CRoBERTa-BLA-FGM-CRF model is built upon the pre-trained Chinese-RoBERTa-wwm-ext language model. It integrates a Bidirectional Long Short-Term Memory (BiLSTM) network, a multi-head attention mechanism, and a Conditional Random Field (CRF) decoding structure, and incorporates a fast gradient method for adversarial training to enhance the model's robustness and generalization ability in medical question recognition tasks. BiLSTM is an improved recurrent neural network specifically designed for processing sequential data, capable of capturing bidirectional dependencies in sequences by combining the outputs of two LSTM networks (forward and backward). The FGM strategy primarily operates on the embedding layer during model training to improve the stability of semantic modeling.

[0082] As an optional implementation, step 200 above includes: passing the natural question to the Chinese-RoBERTa-wwm-ext model to obtain a first context embedding vector; based on the first context embedding vector, introducing a gradient-based perturbation to obtain a second context embedding vector; passing the second context embedding vector to a BiLSTM network for bidirectional encoding to obtain the context sequence features; and then, based on a multi-head attention mechanism and CRF, performing a decoding operation on the context sequence features to obtain the globally optimal label sequence.

[0083] Specifically, the user-input Chinese natural language question is passed to the Chinese-RoBERTa-wwm-ext model to obtain the first context embedding vector. Understandably, the Chinese-RoBERTa-wwm-ext model described above is a pre-trained model. Assume the input sequence is x={x1,x2,...,x...} n}, then its word embedding output is:

[0084]

[0085] Subsequently, the first context embedding vector is passed to the BiLSTM network for bidirectional encoding to obtain context sequence features:

[0086]

[0087]

[0088] Next, the output of the BiLSTM is fed into a multi-head attention module to further model the semantic dependencies between different positions. For each head i∈{1,...,h}, let the query, key, and value matrix be... Then the output of the i-th head is:

[0089]

[0090] The multi-head attention output is:

[0091]

[0092] Let it be denoted as H (2) The input is then processed by a Conditional Random Field (CRF) layer to decode the globally optimal label sequence. The CRF layer's scoring function for the label sequence y is as follows:

[0093]

[0094] Its normalized prediction probability is:

[0095]

[0096] Understandably, this approach achieves highly robust recognition from input questions to BIO entity labels, making it particularly suitable for parsing easily confused words and highly similar entities in Chinese medical scenarios.

[0097] Furthermore, to improve the model's robustness in handling misspellings and synonym variants in natural language questions within medical scenarios, an FGM adversarial training method is introduced into the word embedding layer using a CroBERTa-BLA-FGM-CRF structure. A gradient-based micro-perturbation is added to the first context embedding vector to obtain a second context embedding vector. This second context embedding vector is then used for bidirectional encoding to obtain context sequence features, thereby enhancing generalization ability. The adversarial perturbation vector is calculated as follows:

[0098]

[0099] in Let L(e) be the perturbation amplitude, and L(e) be the training loss function.

[0100] This is done to significantly improve the model's tolerance to input perturbations, making it more robust and stable in natural language questions input by real users.

[0101] It should be noted that in model evaluation, the performance of entity recognition tasks can be measured by precision, recall, and F1 score, which are calculated as follows:

[0102]

[0103] Here, TP represents the number of correctly predicted entities, FP represents the number of false positives, and FN represents the number of false negatives. These metrics comprehensively measure the model's overall performance in entity detection and help to continuously optimize and iterate the model.

[0104] For example, the CRoBERTa-BLA-FGM-CRF model is evaluated using the CCKS 2019 Chinese Medical Named Entity Recognition dataset. This dataset covers multiple medical entity categories such as diseases, drugs, surgeries, and tests, and possesses good label quality and domain diversity, making it a recognized standard test set in the field of Chinese medical text processing. A comparative experiment is conducted by constructing two models: the baseline model Chinese-RoBERTa-wwm-ext + BiLSTM + CRF and the proposed model CRoBERTa-BLA-FGM-CRF, to compare their ability to discriminate semantically ambiguous boundaries.

[0105] In the experiment, precision, recall, and F1 score were used as evaluation metrics to compare the overall performance and entity recognition capabilities of the two models. The results show that the CroBERTa-BLA-FGM-CRF model achieves superior performance across multiple entity categories, particularly showing significant improvements in categories such as "laboratory testing," "imaging examination," and "drugs." This verifies that the CroBERTa-BLA-FGM-CRF model possesses stronger robustness and practicality in Chinese medical question scenarios.

[0106] Table 1: Performance Comparison of Model Structures Across Various Medical Entity Categories

[0107]

[0108] As shown in Table 1, the CRoBERTa-BLA-FGM-CRF model achieved varying degrees of performance improvement across all six categories of medical entities, with the most significant improvements observed in the categories of "drugs," "laboratory tests," and "imaging examinations." Specifically, the overall F1 score increased from 79.85% of the original model to 81.96%, indicating that the CRoBERTa-BLA-FGM-CRF model has more accurate recognition boundaries for named entities, stronger recall capabilities, and higher robustness and generalization ability in entity extraction. Therefore, the experimental results verify that the proposed CRoBERTa-BLA-FGM-CRF structure has superior entity recognition performance in Chinese medical question-answering scenarios, providing more accurate input support for subsequent knowledge graph queries.

[0109] Understandably, the named entity recognition model CRoBERTa-BLA-FGM-CRF, which integrates multiple deep learning modules, is employed in the medical entity recognition stage. CRoBERTa-BLA-FGM-CRF uses the Chinese-RoBERTa-wwm-ext pre-trained language model as its semantic representation foundation, further integrating a bidirectional long short-term memory network to enhance contextual understanding, and introducing a multi-head attention mechanism to strengthen the capture of key information within sentences. Finally, sequence labeling and tag decoding are achieved through a Conditional Random Field (CRF) layer. During training, to improve the model's generalization ability and robustness to anomalous samples, FGM is introduced for adversarial training, thereby significantly improving entity recognition accuracy, especially demonstrating good performance in handling complex situations such as variations of medical terminology and spelling errors.

[0110] Step S300: While performing the above step S200, when the natural question is received, the global representation vector of the CLS tag is obtained based on the natural question, and the context semantic feature extraction operation is performed according to the global representation vector of the CLS tag to obtain the target query relationship type as the intent recognition result.

[0111] In this embodiment, the target query relation type is determined by semantic parsing of natural language questions to guide the subsequent graph reasoning process. Specifically, an intent recognition model composed of a RoBERTa semantic representation layer and a multi-channel convolutional neural network (TextCNN) structure is used to extract contextual semantic features from natural language questions and perform multi-classification reasoning, thereby outputting relation type labels corresponding to the query intent of the natural language questions.

[0112] As an optional implementation, step S300 includes: passing the natural question to the RoBERTa semantic representation layer for overall semantic encoding and extracting the global representation vector of the CLS tag; passing the global representation vector of the CLS tag to the TextCNN network for contextual semantic feature extraction to obtain the intent classification probability distribution; and then determining the target query relationship type based on the intent classification probability distribution.

[0113] Specifically, firstly, the shared RoBERTa is used to perform overall semantic encoding on natural language questions, extracting the global representation vector of CLS tags; then, this vector is input into the TextCNN network, and local n-gram features are extracted through multiple convolutional kernel windows, such as 3, 4, and 5; the features at each scale are pooled and concatenated into a fixed-length vector, which is then input into a fully connected layer for intent classification. The calculation process is as follows:

[0114]

[0115] in Let p represent the concatenated feature representation, and p be the output intent classification probability distribution. The model achieves intent recognition by maximizing the probability corresponding to the true label, demonstrating good discriminative ability in multi-class tasks.

[0116] The purpose of this approach is to perform contextual modeling and semantic feature extraction on user-input natural language questions using an intent classification model composed of a RoBERTa semantic representation layer and a TextCNN multi-channel convolutional network. This accurately identifies the user's true query needs and maps natural language questions to predefined query relationship types in the knowledge graph, such as "querying symptoms of a disease" or "obtaining indications for a drug," providing clear and reliable query guidance signals for subsequent graph inference operations.

[0117] Step S400: Based on the entity recognition result and intent recognition result, perform a relational path reasoning operation in the medical knowledge graph to perform semantic query and obtain target information matching the natural question as the graph retrieval result. If the relational path reasoning operation fails to obtain the target information matching the natural question, execute a fusion retrieval enhancement generation mechanism to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result.

[0118] In this embodiment, entity recognition results can be understood as the identified entity nodes, and intent recognition results can be understood as the identified relationship types; the external medical corpus includes external medical text corpora, literature databases, etc.

[0119] Optionally, based on the identified entity nodes and relationship types, a relational path reasoning operation is performed in the medical knowledge graph to complete the semantic query according to the "entity-relationship-entity" structure and obtain the target information that matches the natural language statement. If the returned graph retrieval results are empty or insufficient to meet the needs of the natural language statement, the fusion retrieval-Augmented Generation (RAG) mechanism is further invoked to retrieve semantically related content of the natural language statement from external medical text corpora or literature databases as the target content. Then, the answer is generated by combining the generative language model, thereby improving the system's adaptability and robustness to complex question-answering scenarios.

[0120] In other words, step S400 may include a fault tolerance and compensation mechanism. When the natural language question input by the user cannot be parsed into a valid entity node or intent type, a prompt message is generated according to the preset fault tolerance mechanism to provide feedback on the reason for the question parsing failure and guide the user to re-input the natural language question. At the same time, during the execution of the graph relational path reasoning operation, if the entity node cannot achieve an exact match, the entity fuzzy matching and extended retrieval are completed by constructing a keyword-based inverted index structure and combining the similarity algorithm calculated by editing distance and Jaccard coefficient weighting.

[0121] Furthermore, if the above relational path reasoning and fuzzy matching still fail to return valid results, the fusion retrieval enhancement generation mechanism is automatically triggered. This mechanism retrieves relevant content from an external medical corpus by vectorizing natural questions and combines it with a generative language model to reason about its semantic context, thereby generating more comprehensive and semantically flexible question-and-answer content to improve responsiveness in complex, fuzzy, or unknown query scenarios.

[0122] As an optional implementation, step S400 includes constructing a query path based on the entity-relationship-entity structure, performing a relationship path reasoning operation on the entity recognition result and intent recognition result in the medical knowledge graph, and obtaining target information that matches the natural question.

[0123] As another optional implementation, the steps of performing relational path reasoning operations on the entity recognition results and intent recognition results further include: constructing a keyword-based inverted index structure for fuzzy matching of entities; and performing extended retrieval operations based on a similarity algorithm calculated by weighting edit distance and Jaccard coefficient.

[0124] As another optional implementation, the step of executing a fusion retrieval enhancement generation mechanism to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result when the relational path reasoning operation fails to obtain the target information matching the natural question includes: when the target information obtained by the relational path reasoning operation is empty, or when the confidence of the target information obtained by the relational path reasoning operation is less than a preset confidence threshold, executing a fusion retrieval enhancement generation mechanism to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result.

[0125] Specifically, based on the extracted entity nodes h and intent categories r, relational path reasoning is performed on the graph.<h,r, ?> The query path is constructed using triples, and entity expansion is performed by combining inverted indexes and fuzzy matching mechanisms. Specifically, edit distance and Jaccard similarity are used to comprehensively score the query.

[0126]

[0127] in For entity keyword set, represents the weighting coefficient. This fusion similarity metric effectively alleviates the path loss problem caused by inaccurate entity matching in knowledge graphs.

[0128] If the returned graph retrieval results are empty or insufficient to support the question-answering task, the RAG mechanism is automatically triggered. Based on the semantic vector of the natural question, relevant corpus fragments are retrieved from an external medical corpus and concatenated with the natural question as input to the generative model to generate supplementary answers. The combination of document retrieval and answer generation improves the system's ability to handle unknown or open-ended questions.

[0129] In this embodiment, step S400 further includes a fusion control mechanism to determine whether the confidence score of the graph inference return result is lower than a set threshold, and to dynamically schedule the RAG mechanism when the confidence score does not meet the condition, thereby avoiding invalid calls to the generation module and improving system efficiency. This mechanism integrates rules and model collaboration strategies, achieving good engineering practicality.

[0130] Understandably, the purpose of this embodiment is to perform relational path reasoning in the medical knowledge graph during the graph reasoning and question-answering generation stages, using "identified entity nodes + relation types determined by intent classification" as input, and completing structured semantic queries through multi-hop logic of "entity-relationship-entity". In the event of graph retrieval failure or incomplete results, a fault-tolerance mechanism is implemented. First, fuzzy matching and expanded retrieval of entities are performed using a keyword inverted index structure, combined with edit distance and Jaccard similarity calculations. If the semantic coverage requirements of natural language questions are still insufficient, a fusion retrieval enhancement generation mechanism is further invoked. By vectorizing natural language questions and retrieving relevant content from an external medical corpus, and utilizing generative language models such as T5, BART, or BioGPT, semantic fusion and response generation are performed on the retrieval results, ultimately outputting a more flexible natural language answer with semantic generalization capabilities.

[0131] Step S500: Based on the preset natural language template and semantic slot filling strategy, the graph retrieval results are converted to obtain the response content corresponding to the natural question.

[0132] In this embodiment, the graph retrieval results are generated into natural language text according to a preset natural language template, improving user comprehension and information acquisition efficiency. Specifically, the structured query results obtained from the graph retrieval are automatically converted based on a preset natural language template and semantic slot filling strategy to generate response text that conforms to user reading habits, thereby achieving user-oriented interactive information presentation.

[0133] Specifically, by combining structured graph retrieval results with the output of the generative model, template filling, ranking fusion, and other techniques are used to generate natural language answers for end users, achieving accurate responses to user-inputted medical natural language questions. The response content supports multi-turn expansion and enhanced interpretability.

[0134] The purpose of this approach is to standardize the language organization and style of the structured results returned by relational path reasoning operations or the unstructured results returned by RAG during the natural language response generation stage. By using a semantic slot-based template filling method, the response content is mapped to user-friendly natural language text, and the response format is beautified, completed, and context-sensitively enhanced, thereby improving the naturalness of human-computer interaction and the user experience.

[0135] In the technical solution provided in this embodiment, a standardized medical knowledge graph is constructed, covering eight types of entities and their semantic relationships, including diseases, symptoms, drugs, and examinations. In the semantic parsing stage, Chinese-RoBERTa-wwm-ext is used as a shared encoder, combined with a BiLSTM multi-head attention mechanism and a CRF decoder to construct a named entity recognition model, CroBERTa-BLA-FGM-CRF. FGM adversarial training is introduced to enhance robustness against misspellings and terminology variations. Simultaneously, RoBERTa combined with a TextCNN model is used to recognize user intent, extracting the query target and relationship type of natural language questions to form the graph retrieval path. A dynamic question-answering mechanism based on entity matching confidence is introduced. If the matching degree is higher than a set threshold, a structured answer is generated based on graph reasoning operations using relationship path reasoning; otherwise, a retrieval enhancement generation mechanism is invoked to integrate semantic retrieval and a language model to complete the answer generation. This achieves not only higher accuracy in medical entity recognition and intent classification, and stronger interpretability and structured expression capabilities at the graph reasoning level, but also demonstrates significant advantages in handling complex scenarios such as unknown questions, fuzzy questions, and open-ended questions. The introduction of the RAG mechanism compensates for the limited coverage of medical knowledge graphs, forming a closed-loop complementary mechanism of structured reasoning and semantic generation, which effectively improves the intelligence, applicability and robustness of the question answering system.

[0136] In addition, refer to Figure 3 This embodiment also proposes a medical question-answering retrieval system that integrates semantic parsing and graph reasoning. The system comprises:

[0137] Knowledge graph construction module 10 is used to perform named entity recognition and semantic relation extraction operations on medical datasets to construct a medical knowledge graph;

[0138] The entity recognition module 20 is used to obtain context sequence features based on the natural question and FGM when a natural question is received, and to perform a decoding operation based on the context sequence features to obtain the globally optimal label sequence as the entity recognition result;

[0139] The intent recognition module 30 is used to obtain the global representation vector of the CLS tag based on the natural question when the natural question is received, and to perform a context semantic feature extraction operation based on the global representation vector of the CLS tag to obtain the target query relationship type as the intent recognition result.

[0140] The graph reasoning module 40 is used to perform a relational path reasoning operation in the medical knowledge graph based on the entity recognition result and the intent recognition result to perform semantic query and obtain target information matching the natural question as the graph retrieval result. When the relational path reasoning operation fails to obtain the target information matching the natural question, a fusion retrieval enhancement generation mechanism is executed to retrieve the target content matching the natural question from an external medical corpus as the graph retrieval result.

[0141] The answer generation module 50 is used to convert the graph retrieval results based on a preset natural language template and semantic slot filling strategy to obtain the response content corresponding to the natural question.

[0142] Please refer to Figure 4 As an optional implementation, the medical question-answering retrieval system integrating semantic parsing and graph reasoning consists of six main functional modules. Among them, the knowledge graph construction module 10 serves as a preprocessing step before system deployment, used to generate a standardized medical semantic knowledge base. The preprocessing part cleans the raw data, then extracts entity relationships, and finally models the data using the Neo4j graph database.

[0143] The remaining modules form the real-time semantic parsing and question-answering reasoning path when users ask questions online, including entity recognition module 20, intent recognition module 30, graph reasoning module 40, RAG module 60, and answer generation module 50. The entire system adopts a modular design to ensure efficient collaboration and decoupling between offline knowledge accumulation and online semantic interaction.

[0144] Specifically, after the user inputs a natural question, the entity recognition module 20 and the intent recognition module 30 process in parallel: extracting key entities from the natural question, constructing graph head nodes and parsing the question intent, and determining the query relationship type.

[0145] Then, the graph reasoning module 40 constructs a ternary query path based on the graph retrieval intent formed by entity h and relation type r.<h,r,?> The system quickly locates candidate entities in the medical knowledge graph through inverted index retrieval and matching, improving retrieval efficiency. Then, a fuzzy matching mechanism is applied, combining edit distance and Jaccard similarity to enhance recall. Finally, a confidence level is determined: if the confidence level is greater than a threshold, the structured graph nodes and attributes are returned; if it is less than the threshold, the RAG module 60 is called to perform corpus retrieval and generate results.

[0146] Finally, the answer generation module 50 integrates the structured and generative results to generate natural language as the result and returns it.

[0147] Please refer to Figure 5Furthermore, in terms of medical question parsing, the aforementioned medical question-answering retrieval system integrating semantic parsing and graph reasoning also includes a joint modeling framework 70 based on a shared semantic encoder. This framework first uniformly encodes the Chinese questions input by the user, and then sends the encoding results to the entity recognition module 20 and the intent recognition module 30 for parallel processing. This joint structure enables the model to capture the collaborative information between entities and intents in natural questions, improving the overall semantic parsing performance.

[0148] Specifically, after the input of a natural sentence, Chinese-RoBERTa-wwm-ext obtains the contextual semantic representation of each character / word in the input natural question; at the same time, FGM adversarial training perturbation is introduced in the embedding layer.

[0149] Then, the entity recognition module 20 and the intent recognition module 30 operate in parallel. In the entity recognition module 20, the BILSTM layer first extracts contextual order dependency features; then, a multi-head attention mechanism layer uses multiple attention heads to focus on semantics at different locations in parallel; then, the CRF decoder predicts the BIO sequence label path; finally, the BIO label sequence is obtained. In the intent recognition module 30, the TextCNN layer first extracts n-gram features of local context fragments using convolutional kernels of different sizes; then, a fully connected layer concatenates the convolutional features from multiple channels and maps them to a unified dimensional space; then, a Softmax layer outputs the intent classification probability distribution; finally, the intent classification label is obtained. Understandably, the BIO label sequence represents the entities output by NER, and the intent classification label represents the relation type output by the intent.

[0150] Then, the relation types of the entities and intents output by the NER are passed to the graph inference module 40, which then uses these relation types to form the graph query input.<h,r,?> .

[0151] It should be noted that the system embodiments described above are merely illustrative. The devices for medical question-answering retrieval that integrate semantic parsing and graph reasoning may or may not be physically separate. The medical question-answering retrieval system integrating semantic parsing and graph reasoning may or may not be a physical unit; that is, it may be located in one place or mapped to a terminal backend via a network. Some or all of the devices can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between devices indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0152] Therefore, the present invention also provides a computer-readable storage medium storing a medical question-answering retrieval program that integrates semantic parsing and graph reasoning. When the medical question-answering retrieval program that integrates semantic parsing and graph reasoning is executed by a processor, it implements the various steps of the medical question-answering retrieval method that integrates semantic parsing and graph reasoning as described in the above embodiments.

[0153] The computer-readable storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0154] It should be noted that, since the storage medium provided in the embodiments of this application is the storage medium used to implement the methods of the embodiments of this application, those skilled in the art can understand the specific structure and variations of the storage medium based on the methods described in the embodiments of this application, and therefore will not be repeated here. All storage media used in the methods of the embodiments of this application fall within the scope of protection of this application.

[0155] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0156] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0157] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxesFigure 1 The function specified in one or more boxes.

[0158] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0159] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, third, etc., does not indicate any order. These words can be interpreted as names.

[0160] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0161] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A medical question and answer retrieval method fusing semantic parsing and graph reasoning, characterized in that, The medical question and answer retrieval method fusing semantic analysis and graph reasoning comprises the following steps: Performing named entity recognition and semantic relation extraction on a medical data set to construct a medical knowledge graph; Upon receiving a natural question, obtaining context sequence features based on the natural question and FGM, and performing decoding operation according to the context sequence features to obtain a globally optimal label sequence as an entity recognition result; meanwhile, Upon receiving the natural question, obtaining a CLS marked global representation vector based on the natural question, and performing context semantic feature extraction operation according to the CLS marked global representation vector to obtain a target query relationship type as an intent recognition result; Performing relationship path reasoning operation in the medical knowledge graph according to the entity recognition result and the intent recognition result to perform semantic query, and obtaining target information matching the natural question as a graph retrieval result, wherein when the relationship path reasoning operation fails to obtain the target information matching the natural question, a fusion retrieval enhancement generation mechanism is performed to retrieve target content matching the natural question in an external medical corpus as the graph retrieval result; Converting the graph retrieval result based on a preset natural language template and a semantic slot filling strategy to obtain a response content corresponding to the natural question. 2.The medical question answering retrieval method of fusing semantic parsing and graph reasoning according to claim 1, wherein, The step of performing named entity recognition and semantic relation extraction on a medical data set to construct a medical knowledge graph comprises: Based on the medical data set, performing cleaning operation, text analysis operation and structured extraction operation, cleaning, standardizing and semantically classifying the unstructured information in the medical data set to obtain medical entities and semantic relations, wherein the medical entities include diseases, symptoms, drugs and departments, and the semantic relations include disease-symptom, disease-drug and disease-department; According to the medical entities and the semantic relations, the medical knowledge graph is constructed. 3.The medical question search method of fusing semantic analysis and graph reasoning of claim 1, wherein, The step of, upon receiving a natural question, obtaining context sequence features based on the natural question and FGM, and performing decoding operation according to the context sequence features to obtain a globally optimal label sequence as an entity recognition result comprises: Passing the natural question to a Chinese-RoBERTa-wwm-ext model to obtain a first context embedding vector; Based on the first context embedding vector, introducing gradient-based perturbation to obtain a second context embedding vector; Passing the second context embedding vector to a BiLSTM network for bidirectional encoding to obtain the context sequence features; Based on multi-head attention mechanism and CRF, decoding the context sequence features to obtain the globally optimal label sequence. 4.The medical question answering retrieval method of fusing semantic analysis and graph reasoning as claimed in claim 1, wherein, The step of, upon receiving the natural question, obtaining a CLS marked global representation vector based on the natural question, and performing context semantic feature extraction operation according to the CLS marked global representation vector to obtain a target query relationship type as an intent recognition result comprises: The natural question is transmitted to a RoBERTa semantic representation layer for overall semantic coding, and a global representation vector of the CLS label is extracted; The global representation vector of the CLS label is transmitted to a TextCNN network for context semantic feature extraction operation to obtain an intent classification probability distribution; According to the intent classification probability distribution, the target query relationship type is determined.

5. The medical question retrieval method of fusing semantic parsing and graph reasoning of claim 1, wherein, The step of performing relationship path reasoning operation in the medical knowledge graph based on the entity recognition result and the intent recognition result to perform semantic query and obtaining target information matched with the natural question as a graph retrieval result comprises: Based on the entity-relation-entity structure, a query path is constructed, and the relationship path reasoning operation is performed on the entity recognition result and the intent recognition result in the medical knowledge graph to obtain the target information matched with the natural question.

6. The medical question retrieval method of fusing semantic parsing and graph reasoning of claim 5, wherein, The step of performing the relationship path reasoning operation on the entity recognition result and the intent recognition result further comprises: An inverted index structure based on keywords is constructed for fuzzy matching of entities; and An extended retrieval operation is performed according to a similarity algorithm weighted by an edit distance and a Jaccard coefficient.

Images