Medical health service system based on multi-granularity semantic parsing and knowledge reasoning engine

Abstract

The application belongs to the technical field of artificial intelligence medical treatment, and discloses a medical health service system and method based on a multi-granularity semantic analysis and knowledge reasoning engine, which comprises the following steps: through a multi-granularity semantic analysis module, coarse-grained intention recognition and fine-grained medical entity state extraction are performed on the unstructured input of a user; a dynamic probability reasoning engine starts two reasoning paths in parallel: a generative reasoning path generates a group of candidate diagnosis hypotheses by using an LLM, and a symbolic reasoning path performs multi-hop probability reasoning on a unique probability knowledge graph to calculate another group of candidate diagnosis paths and their cumulative probabilities; a reasoning fusion and verification module cross-verify the results of the two paths, and only when the hypothesis of the LLM is verified by the high-probability path of the PKG, a final health service response containing a conclusion and an interpretable path is generated. The application significantly improves the accuracy, reliability and transparency of automated medical consultation services.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and computer vision technology, and in particular relates to a medical and health service system based on multi-granularity semantic parsing and knowledge reasoning engine. Background Technology

[0002] With the rapid development and deep application of artificial intelligence technology, automated medical and health service systems, especially public-facing medical question-and-answer and intelligent dialogue systems, have become a research and application hotspot of common concern in academia and industry. These systems are dedicated to providing convenient and timely health consultation services to a wide range of users, while also assisting doctors in conducting preliminary diagnostic screenings. This is of great significance for improving the overall efficiency of medical services and expanding the accessibility of medical resources.

[0003] However, in the journey to build a highly reliable and transparent intelligent healthcare service system, existing technological approaches still face significant and complex technical bottlenecks and challenges, which restrict the progress of related technologies from the laboratory to clinical application.

[0004] Currently, the mainstream technical implementation paths can be broadly categorized into two types: systems based on large-scale language models and systems based on traditional knowledge graphs. Each of these systems has its own distinct technical characteristics, but also its own inherent limitations that are difficult to overcome.

[0005] The first type of technical approach, namely systems based on large-scale language models, has made remarkable progress in recent years. Advanced models, represented by generative pre-trained transformation models, have demonstrated powerful natural language understanding and generation capabilities after being fine-tuned or specially pre-trained in specific fields such as the medical field. For example, the CMed-GPT model, which is specifically pre-trained for the Chinese medical field, can deeply understand complex health questions raised by users in a natural and conversational manner and generate fluent and contextually coherent responses.

[0006] These systems possess a natural advantage in natural language understanding, effectively addressing the diversity of word meanings and the prevalent irregularities in everyday spoken language. However, their fundamental flaw lies in their inherent "black box" nature; the knowledge of large language models is encoded in a highly diffuse and implicit manner within their massive model parameters. Their internal reasoning processes are opaque, making them difficult for humans to directly understand, examine, and verify.

[0007] This "black box" characteristic directly leads to the so-called "illusion" problem, where the model may generate erroneous information that sounds extremely reasonable and logically consistent, but is actually completely inconsistent with medical facts. In the high-risk, high-precision field of medicine, any factual error may lead to incalculable serious clinical consequences. Therefore, systems that rely solely on large language models face enormous and insurmountable challenges in ensuring the reliability and security of their output results, which constitutes a key "translation gap" between technology demonstration and actual clinical application.

[0008] The second type of technical approach is based on knowledge graph systems. To compensate for the lack of verifiable reasoning capabilities in large language models, researchers have developed medical service systems based on knowledge graphs. Knowledge graphs explicitly store massive amounts of medical entities, such as diseases, symptoms, drugs, and examination items, as well as the intricate relationships between them, in a structured graph form. This structure provides a solid and formalized knowledge foundation for the system to perform rigorous logical reasoning.

[0009] Such systems can reason along clearly defined logical paths in a knowledge graph, making their decision-making process transparent and traceable. In principle, every conclusion given by the system can be traced back to its exact basis in the knowledge base.

[0010] However, traditional knowledge graph systems also have their inherent limitations. The primary problem is that they are usually deterministic and struggle to handle the inherent uncertainty that is prevalent in medical knowledge.

[0011] For example, the strength of the association between a symptom and a disease is essentially a probabilistic problem in clinical practice, rather than a simple binary "yes" or "no" relationship. Most existing knowledge graphs lack effective modeling of this uncertainty, resulting in their reasoning results often being too rigid and failing to truly reflect the complexity and probabilistic nature of real-world clinical practice.

[0012] Secondly, these systems are relatively weak in natural language understanding. They usually require users to input structured query statements or rely on relatively simple keyword matching technology. Therefore, they are difficult to effectively handle colloquial, vague, or even ambiguous natural language dialogues input by users, which greatly limits their usability and user experience in real user interaction scenarios.

[0013] To combine the advantages of the two types of technologies mentioned above, some hybrid systems have recently emerged in academia and industry. For example, a model called DR.KNOWS attempts to use knowledge graphs to enhance the diagnostic reasoning capabilities of large language models. Its core idea is to retrieve pre-existing, deterministic knowledge paths related to patient input information from the knowledge graph, and then provide these paths as contextual information to the large language model in order to guide it in generating more accurate diagnoses.

[0014] However, such methods have a fundamental limitation: they are essentially a one-way “guidance” relationship, rather than a true two-way “fusion and verification”.

[0015] On the one hand, the knowledge graph it uses is still deterministic, failing to solve the modeling problem of uncertainty in medical knowledge; more importantly, the system lacks a closed-loop verification mechanism to check the authenticity of the final output of the large language model; after receiving the guidance information from the knowledge graph, the generation process of the large language model itself remains a "black box," and it is impossible to know whether its final output strictly follows the provided logical path or whether it generates new "illusions" outside of the guidance. This architecture merely uses the knowledge graph to assist the large language model, but does not fundamentally solve the core pain point of verifying the reliability of the output results of the large language model.

[0016] In summary, there is a pressing comprehensive technical challenge in the existing technology: the lack of an integrated solution that can simultaneously achieve the following four objectives: 1) To achieve deep and accurate understanding of complex, colloquial natural language; 2) Transparent and verifiable reasoning based on structured medical knowledge; 3) Explicitly model and effectively handle the inherent uncertainties in medical information; 4) And to deeply coordinate and integrate these capabilities to produce a final healthcare service output that is both highly accurate and fully interpretable. Summary of the Invention

[0017] The technical problem to be solved by this invention is to overcome the shortcomings of existing automated medical service systems. Specifically, single large language model (LLM) systems have low reliability of output results and are prone to producing "illusions" that do not conform to the facts due to their "black box" characteristics. Traditional knowledge graph (KG) systems, due to their deterministic models, cannot effectively handle the inherent uncertainty of medical knowledge and are rigid in understanding spoken language user input.

[0018] To address the aforementioned technical problems, this invention provides a healthcare service system and method based on multi-granularity semantic parsing and a knowledge reasoning engine. This system and method utilize an innovative hybrid dual-path parallel reasoning and verification architecture to synergistically leverage the advantages of generative artificial intelligence and symbolic artificial intelligence while avoiding their inherent defects, thereby significantly improving the accuracy, reliability, and transparency of automated healthcare services.

[0019] To achieve the above objectives, the present invention provides the following technical solution: The healthcare service system based on multi-granularity semantic parsing and knowledge reasoning engine includes: an input receiving module, a multi-granularity semantic parsing module, a hybrid knowledge base, a dynamic probabilistic reasoning engine, a reasoning fusion and verification module, and an output generation module.

[0020] The input receiving module is used to receive unstructured, conversational dialogue data from the user.

[0021] The multi-granularity semantic parsing module is communicatively connected to the input receiving module and is used to parse the dialogue data to generate structured entity data.

[0022] The hybrid knowledge base includes a probabilistic knowledge graph (PKG) and a generative large-scale language model (LLM) pre-trained in the medical field.

[0023] The dynamic probabilistic reasoning engine is communicatively connected to the multi-granularity semantic parsing module and the hybrid knowledge base. It is used to launch two reasoning paths in parallel based on the structured entity data: one generative reasoning path using the LLM to generate a first set of candidate diagnostic hypotheses; and the other symbolic reasoning path that performs multi-hop probability calculation on the PKG to generate a second set of candidate diagnostic paths.

[0024] The reasoning fusion and verification module is used to cross-verify the first set of candidate diagnostic hypotheses generated by the generative reasoning pathway and the second set of candidate diagnostic paths generated by the symbolic reasoning pathway, so as to confirm one or more verified final conclusions.

[0025] The output generation module is used to present a health service response containing the final conclusion to the user, which preferably includes a natural language explanation derived from the symbolic reasoning pathway.

[0026] Preferably, the multi-granularity semantic parsing module includes a coarse-grained context analyzer and a fine-grained entity extractor; the coarse-grained context analyzer uses a first language model to perform global analysis on the dialogue data to identify user intent and dialogue topics; the fine-grained entity extractor is used to accurately identify medical entities and their associated state attributes from the dialogue data, thereby generating the structured entity data.

[0027] More preferably, the fine-grained entity extractor employs a two-stage generation framework, in which the first stage generates medical terms mentioned in the dialogue, and the second stage combines predefined pattern information to generate the corresponding state for each term.

[0028] In a further preferred embodiment, the probabilistic knowledge graph (PKG) uses concepts from a standard medical ontology as nodes and relationships between concepts as edges. Each relationship edge is labeled with a probability value representing the confidence level of the relationship. This design primarily aims to explicitly model and address the uncertainty prevalent in medical knowledge.

[0029] Furthermore, in a more preferred embodiment, the relation edges in the PKG, in addition to being labeled with the probability values, also include information about the evidence sources used to generate those probability values, thereby ensuring the traceability and auditability of each piece of knowledge in the graph. The probabilistic knowledge graph is constructed through statistical analysis of multi-source heterogeneous data, including clinical guidelines, medical literature, and anonymized electronic health records (EHRs), and can be validated and corrected using assessments from human medical experts.

[0030] More preferably, when executing the symbolic reasoning path, the dynamic probabilistic reasoning engine employs a graph traversal algorithm, which can aggregate the probability values ​​of each edge along the reasoning path to calculate the total end-to-end path probability.

[0031] More preferably, when performing cross-validation, the inference fusion and verification module uses a weighted scoring function to determine the alignment between the hypothesis of the generative inference path and the path of the symbolic inference path. This function combines the confidence score of the LLM's hypothesis output and the path probability score calculated by the PKG, and compares the final score with a preset threshold to determine whether the verification passes.

[0032] More preferably, when generating the natural language interpretation, the output generation module inputs a structured representation of the verified symbolic reasoning path, such as a sequence of triples, into the LLM and instructs it to translate the logic chain into a coherent and easily understood natural language sentence.

[0033] Compared with the prior art, the present invention has the following advantages: 1. Through an innovative dual-path parallel reasoning and cross-validation mechanism, this invention can use the rigorous logic of symbolic reasoning to examine and filter the potential "illusions" of generative models, while using the powerful associative ability of generative models to compensate for the incompleteness of knowledge graphs. As a result, the final output conclusion is double-validated, and the reliability is significantly improved.

[0034] 2. This invention not only provides the final conclusion, but also generates a natural language explanation that is easy for humans to understand, based on a verified and transparent symbolic reasoning path, clarifying the logical basis for the decision to the user. This greatly overcomes the "black box" problem of traditional AI systems, enhances user trust in the system, and meets the urgent need for explainable XAI in the medical field.

[0035] 3. By introducing a probabilistic knowledge graph (PKG), this invention can explicitly model and compute the inherent uncertainty in medical knowledge, making the reasoning process closer to the complexity and probabilistic nature of clinical reality, and avoiding the overly rigid conclusions of traditional deterministic systems.

[0036] 4. With the help of a multi-granularity semantic parsing module, this invention can deeply understand the colloquial, ambiguous, or even incomplete natural language input by users, accurately extract key medical information, and lay a solid foundation for subsequent high-quality reasoning.

[0037] In summary, this invention, through a series of innovative designs, effectively integrates the advantages of different AI technologies, solves the core pain points of existing technologies, and provides a safer, more reliable, intelligent, and transparent automated healthcare service solution. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a detailed system architecture diagram of a healthcare service system based on multi-granularity semantic parsing and knowledge reasoning engine. Figure 2 This is a comprehensive flowchart of a healthcare service system based on multi-granularity semantic parsing and a knowledge reasoning engine. Figure 3 A conceptual diagram illustrating the core innovation process of the medical and health service system based on multi-granularity semantic parsing and knowledge reasoning engine of this invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0041] Please see Figure 1 As shown, Figure 1 This paper illustrates the overall architecture of a healthcare service system according to a preferred embodiment of the present invention. This system can be flexibly deployed on one or more servers. The server hardware configuration typically includes, but is not limited to: at least one central processing unit (CPU) or graphics processing unit (GPU), main memory (RAM), non-volatile storage media such as hard disk drives (HDDs) or solid-state drives (SSDs), and a network interface. The various functions of the system are implemented by the processor executing pre-stored software module instructions in memory. These modules logically cooperate to form a complete service process. Specifically, these modules include: The input receiving module 110 is the front-end interface for system-user interaction and the entry point for all services. In its specific implementation, it can be an application programming interface (API) for other applications to call; it can be a WebSocket connection for real-time web or mobile chat; or it can be any other technology interface suitable for receiving communication data. Its core responsibility is to receive unstructured input from the user, which is usually spoken natural language text, such as dialogue content directly entered by the user through the chat window, or text converted from the user's speech through speech recognition technology. This module ensures that the original information input by the user can be transmitted completely and accurately to the subsequent processing modules.

[0042] The multi-granularity semantic parsing module 120 is the key to achieving deep understanding of user input. Its performance directly affects the subsequent processing quality of the entire system. It receives raw dialogue data from the input receiving module 110 and performs a carefully designed two-stage parsing process, mainly to extract structured information that is crucial for subsequent reasoning. This parsing process embodies the core idea of ​​"multi-granularity".

[0043] The coarse-grained context analyzer 121 focuses on the macro level of the dialogue. It uses a pre-trained language model similar to the CMed-GPT model to perform a preliminary, global analysis of the entire dialogue segment. Its main tasks are intent detection and topic classification. For example, it can accurately determine whether the user is describing symptoms, asking for drug information, reporting test results, or conducting a routine health consultation. Through this coarse-grained analysis, the system can understand the overall context of the dialogue and the user's core needs, setting the correct direction and scope for subsequent precise inference.

[0044] The fine-grained entity extractor 122 is the core and most challenging part of the parsing process, responsible for information mining at the micro level. This invention preferably employs an advanced sequence-to-sequence model, whose design concept draws on a two-stage generative framework. This extractor is specifically designed to accurately extract medical entities and their related attributes from text. It does not merely identify keywords such as "chest pain," but rather strives to generate a structured, information-rich tuple. For example, given the input "I've been feeling a dull ache in my chest for the past three days," the fine-grained entity extractor can generate similar structured data. This highly structured output, compared to the original natural language text, can be utilized more directly and accurately by the subsequent inference engine, greatly reducing information loss and ambiguity during transmission.

[0045] Hybrid Knowledge Base 130 is the knowledge hub of the entire system and the source of knowledge for the system to make intelligent decisions. It consists of two different but complementary knowledge sources that work together to support reasoning: The probabilistic knowledge graph PKG131 is a core innovation of this invention. It primarily addresses the deterministic shortcomings of existing knowledge graphs, thereby more realistically reflecting the essence of medical knowledge. The construction process of the PKG is rigorous and complex, and preferably includes the following steps: Nodes and edges: Nodes represent standardized medical concepts from authoritative medical ontology libraries such as the Unified Medical Language System (UMLS), such as "pneumonia," "fever," and "amoxicillin," ensuring the standardization and authority of knowledge; edges represent the relationships between these concepts, such as "is_a_symptom_of" or "may_be_treated_by," etc.

[0046] The key innovation of PKG in probabilistic labeling lies in assigning a structured representation to each edge, such as source node, target node, relation type, probability value, and source of evidence. The probability value is a numerical value between 0 and 1, quantifying the confidence or likelihood of the relation holding true. This probability value is not subjectively set, but objectively calculated through large-scale statistical analysis using algorithms such as Bayesian inference and co-occurrence frequency analysis on massive and diverse medical data sources, including clinical guidelines, medical literature databases like PubMed, and rigorously anonymized electronic health records (EHRs).

[0047] Traceability: To achieve complete transparency and auditability, the evidence source field stores the original document identifiers, data record batches, or other traceable identifiers that support the relationship and its probability values. This makes every piece of knowledge in the graph traceable and auditable, greatly enhancing the system's transparency and reliability in professional scenarios. During the construction process, an initial entity and relationship extraction can be performed using LLM to form a draft graph, which is then rigorously reviewed, calibrated, and verified by a team of human medical experts to ensure the final accuracy of the knowledge.

[0048] The generative large language model LLM132 integrates a powerful, medically optimized LLM, such as CMed-GPT or similar models, which was chosen for its superior natural language understanding and generation capabilities. It is mainly responsible for handling nuances at the language level, understanding complex contexts, and generating human-readable natural language text in the final stage.

[0049] Please see Figure 3 As shown, the dynamic probabilistic reasoning engine 140 is the brain of the system, responsible for performing actual and complex reasoning tasks. After receiving structured entity data generated by the multi-granularity semantic parsing module 120, it initiates two in parallel reasoning paths with different functions but the same goal: Generative reasoning pathway 141 fully leverages the advantages of LLM. It formats the parsed structured entity data, such as symptom lists, signs, and personal histories, into a carefully crafted prompt, which is then input into LLM 132. LLM utilizes its vast, implicit knowledge base to perform association, induction, and deduction, thereby generating a list of candidate diagnostic hypotheses ranked by probability. For example, the output might be [('pneumonia', confidence: 0.75), ('bronchitis', confidence: 0.68),...]. This pathway is particularly adept at handling complex, ambiguous, and incomplete inputs, and can provide insightful initial hypotheses.

[0050] Symbolic reasoning path 142 performs formal, transparent, probabilistic logical reasoning. It uses parsed entities such as "chest pain" as starting nodes in the probabilistic knowledge graph PKG131. Then, it executes an efficient probabilistic graph traversal algorithm, such as a bundle search algorithm with cumulative probabilities that typically involve multiplication operations on the path, or a more complex belief propagation algorithm. This algorithm explores the PKG for the most probable reasoning paths that lead from these starting nodes to potential disease nodes. Its output is a set of structured reasoning paths, each with an end-to-end cumulative probability. For example, the output might be [('chest pain'->is_a_symptom_of(0.6)->'pneumonia'), cumulative probability: 0.6]. This path provides a solid, verifiable, and quantitative logical framework for the system's decision-making, which is fundamentally different from the existing approach of merely retrieving static, deterministic paths.

[0051] Please see Figure 3 As shown, the inference fusion and verification module 150 is the essence of the architecture of this invention and the core guarantee for ensuring the reliability of the final output of the system. It is responsible for the thoughtful fusion and rigorous verification of the results of the two independent paths mentioned above.

[0052] The core function of this module is to compare and verify the outputs of two independent pathways. It recognizes that LLM may produce "illusions" and PKG may miss some rare disease knowledge due to incomplete data. By requiring the results of the two pathways to verify each other, the system can greatly reduce the risk of errors in a single model and achieve the error correction and enhancement effect of "1+1>2".

[0053] Specifically, the module receives a list of candidate hypotheses from the generative pathway and a list of inference paths from the symbolic pathway. It checks whether the highest-confidence hypothesis given by the LLM, such as "pneumonia," also appears at the end of a high-probability inference path in the symbolic pathway. If "pneumonia" is indeed the end of one or more high-probability paths, then the hypothesis is preliminarily considered validated, and the system can further calculate a fusion score, for example, a weighted sum. in and It is an adjustable weighting coefficient. Only when this fusion score exceeds a pre-set threshold is the hypothesis finally confirmed as a credible conclusion. Conversely, if the LLM gives a conclusion with an extremely low probability in the PKG, or the PKG arrives at a high-probability conclusion that the LLM has not mentioned at all, the system will determine that the current information is insufficient to draw a reliable conclusion and may proactively ask the user questions to obtain more key information, forming a dynamic interactive closed loop.

[0054] The output generation module 160 is responsible for presenting the conclusions to the user in a clear, user-friendly and responsible manner after the reasoning fusion and verification module 150 confirms one or more verified conclusions.

[0055] For a verified conclusion, the module will once again leverage the powerful natural language generation capabilities of LLM132, but this time, the input to LLM is no longer a vague user question, but rather a verified, structured, symbolic reasoning path as evidence.

[0056] For example: Chest pain -> is_a_symptom_of(0.6) -> pneumonia. At the same time, the LLM will receive a clear instruction, namely, "Please translate this logic chain into a natural language explanation that is easy for ordinary people to understand." The response generated by the LLM might be: "Based on the symptoms you mentioned, such as 'chest pain,' the system analysis believes that there is a possibility of 'pneumonia.' This is because in a large amount of medical data statistics, 'chest pain' is a common associated symptom of 'pneumonia,' with an association confidence level of approximately 60%." This approach provides crucial interpretability (XAI), not only giving a conclusion but also clearly demonstrating the basis for that conclusion, thereby establishing the user's trust in the system's conclusions.

[0057] Please see Figure 2 As shown, Figure 2 This invention provides a detailed explanation of the complete process of a medical and health service method, which is preferably performed by... Figure 1 The system shown will be used to execute this process, and the specific steps are described below: Step S210: Receive user input. The system receives user input dialogue data from various possible clients such as web pages and mobile applications through its front-end input receiving module 110. This data is usually unstructured and conversational.

[0058] Step S220: Perform multi-granularity semantic parsing. The received raw data is sent to the multi-granularity semantic parsing module 120 for in-depth processing. First, the coarse-grained context analyzer 121 identifies the overall intent of the dialogue, such as asking for a diagnosis or consultation. Then, the fine-grained entity extractor 122 accurately extracts all relevant medical entities and their state attributes from the text to form a structured data record.

[0059] Step S230: Parallel execution of dual-path inference. After receiving the structured entity data, the dynamic probabilistic inference engine 140 immediately starts two inference paths in parallel: Step S231, Generative Pathway: The structured entity data is encoded into an optimized cue, which is then input into LLM132. The LLM performs inference and generates a set of candidate diagnostic hypotheses with associated confidence levels.

[0060] Step S232: Symbolic pathways. Core entities such as symptoms in the structured entity data are used as entry nodes of PKG131. The probabilistic graph reasoning algorithm is executed to calculate the reasoning paths from these entry nodes to different conclusions, such as diseases, and their end-to-end cumulative probabilities.

[0061] Step S240: Inference fusion and verification of inference results. The inference fusion and verification module 150 receives the results from the two paths and performs rigorous cross-validation. It compares the assumptions of LLM and the probability paths of PKG to determine whether the two are logically consistent.

[0062] Step S250: Determine whether the verification is successful. The module makes a final judgment based on the preset fusion rules and thresholds. If one or more hypotheses are strongly supported by the high-probability path of PKG, the verification is successful, and the process proceeds to step S260. If all hypotheses fail to pass the verification, it may indicate that the current information is insufficient or there is a conflict. The process can be transferred to step S270, or it can be terminated directly and the user can be informed that a reliable conclusion cannot be drawn at present.

[0063] Step S260: Generate and output the verified response. For the verified conclusion, the output generation module 160 will generate the final response. The response is composite and contains at least two parts: first, clear health advice or conclusion; second, an easy-to-understand natural language explanation derived from the symbolic path of verifying the conclusion, clarifying the basis for the decision.

[0064] Step S270: Request supplementary information. If the verification fails, the system can intelligently analyze the conflict points or missing information points and generate a targeted question. For example, if "coughing" is a key node in multiple high-probability paths but the user did not mention it, the system will ask the user through output module 160: "Do you have coughing symptoms?" Then the process returns to step S210 to receive a new round of input from the user, thus forming a dynamic, multi-round dialogue loop to obtain more sufficient information to make a judgment.

[0065] To illustrate the working method and technical advantages of the present invention in more detail, an exemplary use case is provided below.

[0066] A user entered a vague and complex description into the system: "I've been feeling tired lately, sometimes with chest tightness and a dry cough, but I don't have a fever. What could be the cause?" Step S220: Parsing. The multi-granularity semantic parsing module 120 parses this passage into structured entity data: {(term:'fatigue', state:'existence'), (term:'chest tightness', state:'existence', attribute:'intermittent'), (term:'dry cough', state:'existence'), (term:'fever', state:'non-existent')}.

[0067] Step S230: Inference. S231: Generative Pathway. After receiving these symptoms, LLM132 may associate them with multiple possibilities and generate a hypothesis list [('myocarditis', 0.7), ('anxiety', 0.5), ('atypical pneumonia', 0.45),...]. LLM may assign a high confidence level to "myocarditis" due to symptoms such as "fatigue" and "chest tightness". S232: Symbolic Pathway. In PKG131, dynamic probabilistic inference engine 140 uses "fatigue", "chest tightness", and "dry cough" as starting points and the absence of "fever" as a constraint to perform probabilistic inference. The calculation results may show that the path from "fatigue" and "chest tightness" to "myocarditis" has a high cumulative probability. At the same time, the combination of "dry cough" and "no fever" also makes the path to "atypical pneumonia" have a certain probability. Although "anxiety" may cause similar symptoms, it may lack a strong physiological symptom association path in PKG, and its probability value is low.

[0068] Steps S240-S250, Fusion and Validation: The reasoning fusion and validation module 150 found that the primary hypothesis of LLM, "myocarditis," and the secondary hypothesis, "atypical pneumonia," were both supported by pathways with significant probability values ​​in the symbolic pathway. Therefore, these two hypotheses were validated. However, "anxiety disorder," although it may have been proposed by LLM, could not be validated in PKG due to the lack of strong physiological symptom association pathways and was therefore filtered out.

[0069] Step S260, Output: The output generation module 160 generates a cautious and informative multi-hypothesis response: "Based on your description of symptoms such as 'fatigue,' 'chest tightness,' and 'dry cough,' and the absence of fever, the system analysis suggests that it may be related to multiple conditions, including the possibility of 'myocarditis' or 'atypical infection.' Given the symptom of 'chest tightness,' it is strongly recommended that you seek professional medical evaluation in a timely manner to rule out any serious conditions. The system draws this conclusion based on the fact that your symptom combination has a strong statistical correlation with these diseases in the medical knowledge base." To clearly demonstrate the significant advancements of this invention over existing technologies, the following table compares the key features of different technical solutions: Table 1 Comparison of Technical Solutions As can be clearly seen from the table above, this invention, through its unique dual-path reasoning, probabilistic knowledge graph, and closed-loop verification mechanism, successfully overcomes the limitations of various existing technologies and achieves comprehensive advancements in multiple dimensions such as comprehension ability, transparency, uncertainty handling, reliability, and output quality, thereby providing a truly intelligent service solution that meets the stringent requirements of the medical field.

[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A healthcare service system based on multi-granularity semantic parsing and knowledge reasoning engine, characterized in that: include: At least one processor; And the instructions stored in non-volatile memory and executed by the at least one processor, the instructions being configured to implement the following modules: The input receiving module is used to receive unstructured dialogue data from the user; A multi-granularity semantic parsing module, which is communicatively connected to the input receiving module, is used to parse the dialogue data to generate structured entity data; A hybrid knowledge base comprising a probabilistic knowledge graph (PKG) and a generative large language model (LLM) pre-trained in the medical field, wherein relation edges in the PKG are labeled with probability values ​​representing confidence levels; The dynamic probabilistic reasoning engine, which is communicatively connected to the multi-granularity semantic parsing module and the hybrid knowledge base, is used to execute in parallel based on the structured entity data: a generative reasoning path using the LLM to generate a first set of candidate diagnostic hypotheses; and a symbolic reasoning path that performs multi-hop probability calculation on the PKG to generate a second set of candidate diagnostic paths. The reasoning fusion and verification module is used to cross-validate the first set of candidate diagnostic hypotheses and the second set of candidate diagnostic paths to confirm one or more validated final conclusions; and An output generation module is used to present a health service response that includes the final conclusion and a natural language explanation derived from the symbolic reasoning pathway.

2. The medical and health service system based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 1, characterized in that, The multi-granularity semantic parsing module is specifically used to: perform coarse-grained context analysis to identify user intent; And perform fine-grained entity extraction to identify medical entities and their states, thereby generating the structured entity data.

3. The medical and health service system based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 1, characterized in that, The inference fusion and verification module is specifically used to perform cross-validation by comparing whether the candidate diagnostic hypothesis generated by the LLM exists at the endpoint of the high-probability candidate diagnostic path calculated by the PKG.

4. The medical and health service system based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 1, characterized in that, The output generation module is specifically used to: input the structured representation of the verified symbolic reasoning path into the LLM, so that the LLM can translate it into the natural language interpretation.

5. The medical and health service system based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 1, characterized in that, In addition to the probability values, the relationship edges in the probability knowledge graph are also labeled with the evidence source information used to generate those probability values.

6. A medical and health service method based on multi-granularity semantic parsing and knowledge reasoning engine, characterized in that, Includes the following steps: Step S210: Receive unstructured, conversational dialogue data from the user; Step S220: Perform multi-granularity semantic parsing to parse the dialogue data to generate structured entity data; Step S230: On a hybrid knowledge base containing a probabilistic knowledge graph (PKG) and a generative large-scale language model (LLM), inference is performed in parallel based on the structured entity data. This step includes: Step S231: Execute a generative reasoning path to generate a first set of candidate diagnostic hypotheses using the LLM; and Step S232: Execute a symbolic reasoning path and perform multi-hop probabilistic reasoning on the PKG to calculate the second set of candidate diagnostic paths and their probability scores; Step S240: The results of the generative reasoning pathway and the symbolic reasoning pathway are fused and verified. When the candidate diagnostic hypothesis of the LLM aligns with the high-probability path of the PKG, a final, verified health service response is confirmed; and Step S260: Generate and output the final, verified health service response, wherein the response includes a natural language interpretation derived from the verified symbolic reasoning path.

7. A medical and health service method based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 6, characterized in that, The fusion and verification in step S240 is to determine the degree of alignment through a weighted scoring function. This function combines the confidence score of the LLM's hypothesis output and the path probability score calculated by the PKG, and compares the final score with a preset threshold.

8. A medical and health service method based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 6, characterized in that, The multi-hop probabilistic inference in step S232 employs a graph traversal algorithm, which aggregates the probability values ​​of each edge along the inference path to calculate the total end-to-end path probability.

9. A medical and health service method based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 6, characterized in that, The natural language interpretation in step S260 is generated by inputting the structured triple sequence of the verified symbolic reasoning path into the LLM, which then translates it into a coherent natural language sentence.

Images