An intelligent inquiry method and system based on semantic understanding

Abstract

This invention discloses an intelligent consultation method and system based on semantic understanding, relating to the field of natural language processing technology. The method includes parsing user consultation content to form candidate semantic lattices, updating the semantic matching degree and distinguishing attributes of candidate standard medical concepts based on historical consultation responses, and constructing candidate semantic interpretation branches with valid conditions. It simulates and predicts the impact of responses on interpretation branches, determines the disambiguation benefit of candidate follow-up questions, and selects target follow-up questions based on interaction burden until termination conditions are met. After follow-up questions, it determines or retains the corresponding interpretation branches and valid conditions for candidate concepts, constructs deterministic query branches and parallel candidate query branches respectively, retrieves medical knowledge, and forms shared consultation information, conditional consultation information, and branch difference information. This method can improve the targeting of follow-up question content determination in the semantic understanding-based intelligent consultation process and enhance the accuracy and understandability of medical knowledge retrieval and differentiated output under semantic uncertainty.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an intelligent diagnostic method and system based on semantic understanding. Background Technology

[0002] Existing intelligent consultation technologies are typically applied to scenarios such as online consultations, symptom self-assessment, and intelligent triage. They identify information such as symptoms, body parts, duration, severity, and accompanying manifestations in user input through medical dictionaries, rule bases, classification models, or pre-trained language models. They also combine techniques such as intent recognition, slot filling, contextual association, and medical knowledge graphs to understand the semantics of user consultations. During multi-round interactions, existing methods generally supplement missing information according to a pre-set consultation process, decision tree, or question-and-answer template. Then, based on the identified symptom combinations, disease categories, or departmental tags, they call upon the medical knowledge base to output relevant medical knowledge, risk warnings, or triage suggestions.

[0003] Existing methods, when dealing with colloquial, incomplete, or ambiguous medical consultation expressions, typically require early identification of a single semantic interpretation. This makes it difficult to effectively evaluate the distinguishing role of different follow-up questions among multiple possible interpretations, easily leading to semantic judgment bias or repeated follow-up questions. Furthermore, when the consultation semantics are not yet fully determined, existing methods lack differentiated control over the medical knowledge corresponding to different semantic interpretations, easily resulting in the mixed output of medical information with different or contradictory scopes of application, affecting the accuracy and comprehensibility of intelligent consultation results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent consultation method and system based on semantic understanding, which solves the problems of difficulty in effectively determining the content of follow-up questions when multiple candidate semantic interpretations coexist, and the lack of branched retrieval and differentiated output of medical knowledge corresponding to different semantic interpretations when the consultation semantics are not yet fully determined.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent diagnostic method based on semantic understanding, comprising,

[0008] The system analyzes user consultation content, extracts semantic attributes, identifies multiple candidate standard medical concepts from the medical concept library, and forms a candidate semantic lattice set.

[0009] By combining historical consultation responses to update the semantic matching degree and unconfirmed distinguishing attributes of each candidate standard medical concept, semantically compatible candidate standard medical concepts are combined into candidate semantic interpretation branches with valid conditions.

[0010] Based on the candidate semantic interpretation branches, candidate follow-up questions and predicted answers are generated. The impact of each predicted answer on the candidate semantic interpretation branches is simulated to determine the disambiguation benefit of each candidate follow-up question. The target follow-up question is selected based on the interaction burden of the candidate follow-up questions.

[0011] The execution target is followed up with questions, and the candidate semantic lattice set and candidate semantic interpretation branches are updated based on the follow-up questions and answers, until there are no candidate follow-up questions whose disambiguation benefits are higher than the interaction burden, or the candidate semantic interpretation branches meet the preset interpretation branch termination conditions.

[0012] After the follow-up questioning, candidate standard medical concepts that meet the preset conditions for determining medical concepts will be determined as formal medical concepts, while candidate standard medical concepts that do not meet the preset conditions for determining medical concepts, their respective candidate semantic interpretation branches, and their conditions for establishment will be retained.

[0013] Deterministic query branches are constructed based on formal medical concepts, parallel candidate query branches are constructed based on the retained candidate semantic interpretation branches, and medical knowledge entries corresponding to each query branch are retrieved.

[0014] Based on the applicability of medical knowledge items in deterministic query branches and parallel candidate query branches, common consultation information, conditional consultation information with valid conditions, and branch difference information are generated.

[0015] As a preferred embodiment of the semantic understanding-based intelligent consultation method of the present invention, the step of forming a candidate semantic lattice set includes: dividing the user's consultation content into multiple original consultation expression units, identifying the medical entity expressions and corresponding semantic states in each original consultation expression unit, extracting the semantic attributes associated with the medical entity expressions, and generating expression semantic records; determining multiple candidate standard medical concepts from a medical concept library based on the expression semantic records, determining the initial semantic matching degree, confirmed supporting attributes, confirmed exclusion attributes, confirmed distinguishing attributes, and unconfirmed distinguishing attributes of each candidate standard medical concept according to concept discrimination rules, and associating the expression semantic records with the discrimination results of the corresponding candidate standard medical concepts to form a candidate semantic lattice set.

[0016] As a preferred embodiment of the semantic understanding-based intelligent consultation method of the present invention, the step of updating the semantic matching degree and unconfirmed distinguishing attributes of each candidate standard medical concept by combining historical consultation answers, and combining semantically compatible candidate standard medical concepts into candidate semantic interpretation branches with valid conditions, includes: obtaining historical candidate semantic lattices and historical consultation answers corresponding to the candidate semantic lattice set; associating candidate semantic lattices describing the same medical object across rounds to form a cross-round candidate semantic lattice set; identifying semantic conflicts in the cross-round candidate semantic lattice set; determining semantic attributes without semantic conflicts as valid semantic attributes; and updating the semantic matching degree, confirmed distinguishing attributes, unconfirmed distinguishing attributes, and candidate status of each candidate standard medical concept according to the valid semantic attributes and concept discrimination rules to form an updated candidate semantic lattice set; extracting candidate standard medical concepts in an active state from the updated candidate semantic lattice set; combining semantically compatible candidate standard medical concepts according to the concept relationship rules in the medical concept library; and determining the conditions for fulfillment and the conditions for unconfirmation according to the corresponding concept discrimination results to form a candidate semantic interpretation branch set.

[0017] As a preferred embodiment of the semantic understanding-based intelligent consultation method of the present invention, the step of generating candidate follow-up questions and predicted answers based on candidate semantic interpretation branches refers to: comparing the conditions to be confirmed for each candidate semantic interpretation branch, generating candidate follow-up questions based on unconfirmed distinguishing attributes, semantic conflict information and unconfirmed high-risk attributes, and generating a set of predicted answers based on the unconfirmed semantic attributes and attribute value requirements corresponding to the candidate follow-up questions.

[0018] As a preferred embodiment of the semantic understanding-based intelligent consultation method of the present invention, the step of simulating the impact of each predicted answer on candidate semantic interpretation branches, determining the disambiguation benefit of each candidate follow-up question, and selecting a target follow-up question in combination with the interaction burden of the candidate follow-up question includes: establishing a temporary simulated state for each predicted answer; updating the semantic matching degree, distinguishing attributes, and candidate state of candidate standard medical concepts according to the predicted answer, and reconstructing candidate semantic interpretation branches; comparing the degree of overlap between active candidate standard medical concepts, candidate semantic interpretation branches, pending confirmation items, and branch query features before and after the simulation update, determining the simulated disambiguation effect of each predicted answer, and summarizing the disambiguation benefit of the candidate follow-up question; determining the interaction burden based on the content length of the candidate follow-up question text, the degree of semantic repetition with historical follow-up question text, the number of semantic attributes to be answered, and the number of follow-up question rounds already executed; selecting a target follow-up question from candidate follow-up question propositions whose disambiguation benefit is higher than the interaction burden; and generating a follow-up question end marker when there are no candidate follow-up question propositions that meet the conditions.

[0019] As a preferred embodiment of the semantic understanding-based intelligent consultation method of the present invention, the step of updating the candidate semantic lattice set and candidate semantic interpretation branches according to the follow-up question answers includes: executing the target follow-up question and receiving the follow-up question answers; forming a new round of current candidate semantic lattice set according to the follow-up question answers; associating the new round of current candidate semantic lattice set with historical candidate semantic lattices and historical consultation answers across rounds; updating the semantic state of candidate standard medical concepts and reconstructing candidate semantic interpretation branches; redetermining the target follow-up question according to the updated candidate semantic lattice set and candidate semantic interpretation branches, until a follow-up question end marker is generated or the candidate semantic interpretation branch meets the preset interpretation branch termination condition.

[0020] As a preferred embodiment of the semantic understanding-based intelligent consultation method of the present invention, the step of determining candidate standard medical concepts that meet the preset medical concept determination conditions as formal medical concepts, and retaining candidate standard medical concepts that do not meet the preset medical concept determination conditions, their respective candidate semantic interpretation branches, and their establishment conditions, includes: determining whether candidate standard medical concepts meet the preset medical concept determination conditions based on the number of active candidate standard medical concepts, semantic matching degree difference, key distinguishing attributes, and semantic conflict information in each candidate semantic grid at the end of the follow-up question, and determining candidate standard medical concepts that meet the preset medical concept determination conditions as formal medical concepts; for candidate semantic grids that do not meet the preset medical concept determination conditions, retaining active candidate standard medical concepts, their respective candidate semantic interpretation branches, the established conditions already met, and the established conditions to be confirmed.

[0021] As a preferred embodiment of the semantic understanding-based intelligent consultation method of the present invention, the step of constructing a deterministic query branch based on formal medical concepts, constructing a parallel candidate query branch based on the retained candidate semantic interpretation branches, and retrieving medical knowledge entries corresponding to each query branch includes: constructing a deterministic query branch based on formal medical concepts and associated effective semantic attributes and negative information; combining active candidate standard medical concepts, corresponding establishment conditions, and consistent effective semantic attributes into parallel candidate query branches according to the retained candidate semantic interpretation branches; determining the query branch weight of the parallel candidate query branches based on the semantic matching degree and the proportion of establishment conditions satisfied by the candidate standard medical concepts; associating the deterministic query branches and the parallel candidate query branches to form a consultation query branch set; retrieving and filtering medical knowledge entries from a preset medical information database based on the medical concept field, semantic attribute constraint field, and exclusion condition field in the consultation query branch set; merging and sorting duplicate medical knowledge entries within the same query branch to generate a candidate consultation evidence set.

[0022] As a preferred embodiment of the semantic understanding-based intelligent consultation method of the present invention, the formation of shared consultation information, conditional consultation information with establishment conditions, and branch difference information includes: determining the applicable status of medical knowledge items in the corresponding query branches based on the knowledge application conditions, knowledge exclusion conditions, and establishment conditions of medical knowledge items; determining the medical knowledge content applicable to deterministic query branches or all relevant parallel candidate query branches as shared consultation information; associating the medical knowledge content applicable only to some relevant parallel candidate query branches with the corresponding establishment conditions to form conditional consultation information; and retaining the correspondence between each medical knowledge content and its respective parallel candidate query branch for conflicting medical knowledge content on the same semantic matter in different parallel candidate query branches to form branch difference information.

[0023] Secondly, the present invention provides an intelligent consultation system based on semantic understanding, comprising:

[0024] The consultation semantic parsing module is used to parse the user's consultation content and form a set of candidate semantic lattices;

[0025] The semantic branching module is used to update the semantic state of candidate standard medical concepts and form candidate semantic interpretation branches;

[0026] The follow-up question decision module is used to generate candidate follow-up questions and predict answers, and select target follow-up questions based on disambiguation benefits and interaction burden;

[0027] The follow-up question update module is used to update the candidate semantic lattice set and candidate semantic interpretation branches based on the follow-up question answers, and to determine the formal medical concept;

[0028] The branch retrieval module is used to construct deterministic query branches and parallel candidate query branches, and retrieve the corresponding medical knowledge entries;

[0029] The consultation information generation module is used to generate common consultation information, conditional consultation information, and branch difference information based on the applicability of branches of medical knowledge items.

[0030] The beneficial effects of this invention are as follows: by continuously retaining multiple candidate standard medical concepts and their corresponding candidate semantic interpretation branches, simulating the impact of different predicted answers on the candidate branches, and comprehensively considering the disambiguation benefits and interaction burden to select target follow-up questions, it can reduce premature determination of medical concepts and invalid repeated follow-up questions; after the follow-up questions are completed, for candidate medical concepts that cannot be uniquely determined, their respective interpretation branches and conditions for establishment are retained, and parallel candidate query branches are constructed respectively. Based on the applicability of medical knowledge in each query branch, common consultation information, conditional consultation information with conditions for establishment, and branch difference information are formed, thereby improving the pertinence of semantic disambiguation and the accuracy, clarity, and interpretability of consultation information output. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0032] Fig. 1 This is a flowchart of the semantic understanding-based intelligent consultation method in Example 1.

[0033] Fig. 2 This is a structural diagram of the semantic understanding-based intelligent consultation system in Example 1.

[0034] Fig. 3 This is a schematic diagram of the process of generating candidate follow-up questions and selecting target follow-up questions in Example 1.

[0035] Fig. 4 This is a flowchart illustrating the process of querying branches and consultation information in Example 1. Detailed Implementation

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0038] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0039] Example 1, referring to Figs. 1-4 This is the first embodiment of the present invention, which provides an intelligent consultation method based on semantic understanding, including the following steps:

[0040] S1. Analyze the user's consultation content, extract semantic attributes, identify multiple candidate standard medical concepts from the medical concept library, and form a candidate semantic lattice set;

[0041] S1.1: Receive intelligent consultation request and generate consultation session object;

[0042] The system receives intelligent consultation requests from user terminals, identifies the text content directly input by the user or the text content converted from speech as the current consultation content, assigns a consultation session number to the current intelligent consultation request, and determines the current consultation round. For subsequent inputs in the same multi-round consultation, the corresponding consultation session number is used and the current consultation round is updated. When the current consultation content is a brief answer to the follow-up question in the previous round, it is associated with the medical entity expression or semantic attribute to be confirmed corresponding to the follow-up question in the previous round. The consultation session number, the current consultation round, the current content, and the follow-up question association information are encapsulated to generate a consultation session object.

[0043] S1.2: Parse the current consultation content and generate semantic records;

[0044] Read the current consultation content in the consultation session object and divide it into original consultation expression units according to statement boundaries and independent medical semantics; when the same statement contains multiple independent medical manifestations, they are split separately, and when adjacent statements describe the same medical manifestation, they are merged.

[0045] Identify medical entity expressions used to represent symptoms, abnormal body parts, diseases, drugs, examination items, or examination results from each original consultation expression unit, and determine their affirmative, negative, or uncertain semantic state based on the modification relationship of negative and uncertain expressions; take the medical entity expression as the center, extract body parts, occurrence time, duration, frequency of occurrence, severity, inducing factors, alleviating factors, and accompanying manifestations to form the semantic attributes of the occurrence.

[0046] For semantic attributes that can be standardized, they are converted into corresponding standard attribute values; for semantic attributes whose specific values ​​cannot yet be determined, the original expression is retained and marked as pending confirmation. For brief answers that omit medical entity expressions, their semantic attribution is determined according to the follow-up questions in the consultation conversation.

[0047] The original consultation expression unit, medical entity expression, semantic state, and existing semantic attributes are associated to generate an expression semantic record.

[0048] S1.3: Determine the candidate standard medical concepts and their analytical status based on the semantic records of the expression;

[0049] Based on a pre-defined medical terminology mapping relationship, the medical entity expressions in the semantic records are standardized by synonym and colloquial expressions to obtain standardized medical expressions. Standard concept names and synonyms in a pre-defined medical concept library are then retrieved based on these standardized medical expressions. Simultaneously, a medical text encoding model with fixed input parameters is used to generate a consultation semantic vector by inputting the standardized medical expressions, existing semantic attributes, and original consultation expression units. Semantic-related standard medical concepts are then recalled based on the semantic similarity (cosine similarity) between the consultation semantic vector and the concept semantic description vector.

[0050] The name retrieval results and semantic recall results are merged and deduplicated. Concepts whose medical concept types are inconsistent with the object types represented by medical entity expressions are deleted. Multiple standard medical concepts that meet the preset candidate recall conditions are retained to generate a candidate standard medical concept set.

[0051] For each candidate standard medical concept, the corresponding concept discrimination rules are used to match the existing semantic attributes with concept supporting attributes, concept excluding attributes, and concept distinguishing attributes to determine the confirmed supporting attributes, confirmed excluding attributes, confirmed distinguishing attributes, and unconfirmed distinguishing attributes. The degree of name matching for each candidate standard medical concept is determined according to the order of complete matching of standard concept names, synonym matching, and medical abbreviation or colloquial expression mapping matching. The highest degree of name matching occurs with complete matching of standard concept names, followed by synonym matching, then medical abbreviation or colloquial expression mapping matching, and the lowest degree of name matching occurs when no name matching occurs. This is combined with... The basic matching degree of each candidate standard medical concept is determined by the degree of name matching and the semantic similarity between the consultation semantic vector and the concept semantic description vector. The higher the degree of name matching and the greater the semantic similarity, the higher the corresponding basic matching degree. The basic matching degree is then increased by the support adjustment value set in the concept discrimination rules and decreased by the rejection adjustment value set in the rules, so as to obtain the initial semantic matching degree of each candidate standard medical concept. The support adjustment value and rejection adjustment value are determined based on the difference in the proportion of occurrence of the corresponding semantic attribute in the manually annotated target standard medical concept sample and the non-target standard medical concept sample, and are fixed after review by medical professionals.

[0052] It should be noted that the preset medical terminology mapping relationship is the correspondence between standard medical concept names and their synonyms, medical abbreviations, and common colloquial expressions; the preset medical concept database stores standard medical concepts and their concept names, synonyms, medical concept types, concept semantic descriptions, and concept discrimination rules; the medical text encoding model uses the anonymized and manually annotated historical consultation expressions and the corresponding data of standard medical concepts, fine-tuned to Chinese RoBERTa-wwm-ext, and the model parameters are fixed after fine-tuning; the preset candidate recall condition is that standard medical concepts with exact name matching or synonym matching are directly included in the candidate range, and other standard medical concepts are included in the candidate range when the semantic similarity reaches the preset recall threshold; the preset recall threshold is in the range of 0.60 to 0.85, and is pre-determined by performing Recall@K analysis on the validation data composed of manually annotated historical consultation expressions and their corresponding standard medical concepts.

[0053] The concept discrimination rules include concept support attributes, concept exclusion attributes, and concept differentiation attributes. Among them, concept support attributes are used to support the validity of candidate standard medical concepts, concept exclusion attributes are used to indicate incompatibility with candidate standard medical concepts, and concept differentiation attributes are used to distinguish between multiple similar candidate standard medical concepts.

[0054] S1.4: Form a set of candidate semantic lattices.

[0055] Using each original medical consultation expression unit as a unit, the corresponding expression semantic records, candidate standard medical concepts and their discrimination states are associated to generate candidate semantic lattices.

[0056] Among them, the candidate semantic lattice retains the corresponding medical entity expression and its semantic state, the semantic attributes that have appeared, multiple candidate standard medical concepts, as well as the initial semantic matching degree, confirmed supporting attributes, confirmed exclusion attributes, confirmed distinguishing attributes and unconfirmed distinguishing attributes of each candidate standard medical concept.

[0057] Arrange the candidate semantic grids according to the order of the original consultation expression units, and associate them with the consultation session number and the current consultation round to generate the current candidate semantic grid set.

[0058] S2. Combine historical consultation answers to update the semantic matching degree and unconfirmed distinguishing attributes of each candidate standard medical concept, and combine semantically compatible candidate standard medical concepts into candidate semantic interpretation branches with valid conditions.

[0059] S2.1: Associate the current candidate semantic grid with historical consultation data;

[0060] Read the consultation session number in the current candidate semantic grid set, and obtain the historical candidate semantic grid and historical consultation answers formed in the previous consultation rounds based on the consultation session number.

[0061] When the current consultation content is a response to the previous follow-up question, the current candidate semantic grid is associated with the historical candidate semantic grid corresponding to the previous follow-up question based on the follow-up question association information; when the current consultation content is not a response to a follow-up question, the current candidate semantic grid describing the same medical object is associated with the historical candidate semantic grid based on the correspondence between the medical object represented by the medical entity expression, the candidate standard medical concept, and the semantic attributes.

[0062] Merge the current candidate semantic grid and the historical candidate semantic grid after association. For semantic attributes that are repeated and have the same value in different consultation rounds, retain them only once; for semantic attributes that are more specific in subsequent answers and do not conflict with historical attribute values, use the more specific attribute value as the current valid attribute value; for conflicting attribute values, retain the corresponding attribute value and consultation round respectively; for current candidate semantic grids that do not have a corresponding historical candidate semantic grid, directly add them as new candidate semantic grids to form a cross-round candidate semantic grid set.

[0063] S2.2: Update the semantic state of candidate standard medical concepts;

[0064] The semantic state and semantic attributes of the same medical object in different consultation rounds are compared in the candidate semantic lattice set across rounds. When the same medical object is in an affirmative state and a negative state in different consultation rounds, or when the same semantic attribute has a definite attribute value that cannot be true at the same time, semantic conflict information is generated. Attributes with semantic conflicts are not considered as definite supporting attributes or exclusionary attributes before the conflict is resolved. When the conflicting attribute has previously participated in the semantic matching degree adjustment, the adjustment result generated by the attribute is revoked. The remaining attributes are determined as valid semantic attributes.

[0065] For candidate standard medical concepts that already exist in previous consultation rounds, the corresponding historical semantic matching degree is used as the basis for updating, and the semantic matching degree is updated using the effective semantic attributes newly added in the current consultation round; for candidate standard medical concepts newly added in the current consultation round, the initial semantic matching degree is used as the current semantic matching degree.

[0066] When a newly added valid semantic attribute satisfies the concept support attribute, the semantic matching degree of the corresponding candidate standard medical concept is increased; when a newly added valid semantic attribute satisfies the concept exclusion attribute, the corresponding semantic matching degree is decreased. The same semantic attribute is adjusted only once for the same candidate standard medical concept.

[0067] When a newly added valid semantic attribute provides a definite attribute value for an unconfirmed distinguishing attribute, the corresponding unconfirmed distinguishing attribute will be updated to a confirmed distinguishing attribute; when a confirmed distinguishing attribute has a semantic conflict, the corresponding distinguishing attribute will be restored to an unconfirmed distinguishing attribute.

[0068] When an exclusion attribute is confirmed to match an incompatible exclusion condition in the concept discrimination rules, the corresponding candidate standard medical concept is set to an excluded state. When an exclusion attribute is confirmed to be only a general exclusion attribute, the corresponding semantic matching degree is reduced, and the corresponding candidate standard medical concept remains in an active state. An incompatible exclusion condition is a definite attribute value or combination of attributes that cannot be simultaneously met with the corresponding candidate standard medical concept.

[0069] Valid semantic attributes are matched with preset high-risk semantic rules. When a valid semantic attribute satisfies the combination of a high-risk medical manifestation and its corresponding attribute, a confirmed high-risk attribute is generated. When a high-risk medical manifestation has appeared but the key semantic attributes required for risk assessment are still lacking, the missing key semantic attributes are associated with the corresponding high-risk medical manifestation and recorded as an unconfirmed high-risk attribute.

[0070] The updated semantic matching degree, distinguishing attribute state, and candidate state are written into the corresponding cross-round candidate semantic grid to form the updated candidate semantic grid set.

[0071] It should be noted that the preset high-risk semantic rules include high-risk medical manifestations, key semantic attributes associated with high-risk medical manifestations, trigger values ​​for each key semantic attribute, and attribute combinations used to confirm a high-risk state. High-risk medical manifestations refer to symptoms, signs, or abnormal examinations that are pre-determined by medical professionals and may indicate the need for priority treatment. Key semantic attributes refer to the severity, duration, frequency, and accompanying symptoms used to determine whether a corresponding high-risk medical manifestation reaches a trigger state. Trigger values ​​are preset values ​​or ranges that each key semantic attribute must satisfy. For example, chest pain can be set as a high-risk medical manifestation, and severity, duration, and whether it is accompanied by dyspnea can be set as key semantic attributes. When the relevant attributes meet preset combination conditions, the corresponding high-risk state is confirmed.

[0072] S2.3: Construct candidate semantic interpretation branches with valid conditions.

[0073] Read the updated candidate semantic grid set according to the original consultation expression unit sequence, and extract the active candidate standard medical concepts from each candidate semantic grid.

[0074] Starting with the first candidate semantic lattice containing active candidate standard medical concepts, initial candidate semantic interpretation branches are established for each active candidate standard medical concept in the lattice.

[0075] Read subsequent candidate semantic lattices sequentially. For existing candidate semantic interpretation branches, determine whether each active candidate standard medical concept in the subsequent candidate semantic lattice is compatible with existing candidate standard medical concepts within the branch. When there are no mutual exclusion relationships, temporal relationship conflicts, or valid semantic attribute conflicts, copy the corresponding candidate semantic interpretation branch and add the active candidate standard medical concept to the copied branch to form an extended branch. Only one candidate standard medical concept is retained in the same candidate semantic lattice within the same candidate semantic interpretation branch.

[0076] Based on the conceptual analysis results of each candidate standard medical concept within the branch, confirmed supporting attributes and confirmed distinguishing attributes are identified as conditions that have been met, while unconfirmed distinguishing attributes and their corresponding requirements are identified as conditions to be confirmed. Branch query features are generated based on the candidate standard medical concepts within the branch and the valid semantic attributes consistent with the branch.

[0077] When some candidate semantic lattices cannot be added to the candidate semantic interpretation branch, the corresponding candidate semantic lattices and the reasons for incompatibility are recorded as content to be disambiguated. The resulting candidate semantic interpretation branches are then summarized to generate a candidate semantic interpretation branch set.

[0078] The updated candidate semantic lattice set, candidate semantic interpretation branch set, semantic conflict information, confirmed high-risk attributes, and unconfirmed high-risk attributes are encapsulated to generate a dynamic consultation semantic state object. The candidate semantic interpretation branches retain the candidate standard medical concepts and their corresponding current semantic matching degree, met conditions, pending confirmation conditions, branch query features, and content to be disambiguated.

[0079] S3. Generate candidate follow-up questions and predicted answers based on the candidate semantic interpretation branches, simulate the impact of each predicted answer on the candidate semantic interpretation branches, determine the disambiguation benefit of each candidate follow-up question, and select the target follow-up question based on the interaction burden of the candidate follow-up questions.

[0080] S3.1: Generate candidate follow-up questions and predict answers;

[0081] The system reads the set of candidate semantic interpretation branches, semantic conflict information, and unconfirmed high-risk attributes from the dynamic consultation semantic state object, and extracts the conditions for confirmation and the content to be disambiguated for each candidate semantic interpretation branch; it then compares the conditions for confirmation of different candidate semantic interpretation branches. When different candidate semantic interpretation branches need to be distinguished by different attribute values ​​of the same unconfirmed distinguishing attribute, the corresponding unconfirmed distinguishing attribute is identified as the semantic attribute to be confirmed, and concept-distinguishing candidate follow-up questions are generated. For semantic conflict information, the conflicting medical objects, semantic attributes, and conflicting attribute values ​​are extracted, and conflict-confirming candidate follow-up questions are generated to confirm valid attribute values. For unconfirmed high-risk attributes, the corresponding high-risk medical manifestations and missing key semantic attributes are extracted, and risk-screening candidate follow-up questions are generated to determine the corresponding high-risk status.

[0082] Candidate follow-up propositions with the same semantic attributes to be confirmed and the same disambiguation purpose are merged, and the candidate standard medical concepts and candidate semantic interpretation branches associated with each candidate follow-up proposition are retained to form a set of candidate follow-up propositions.

[0083] The answer type is determined based on the semantic attributes to be confirmed corresponding to the candidate follow-up questions. For state-type semantic attributes, affirmative, negative, and uncertain answers are generated. For semantic attributes with multiple standard values, the confirmation conditions set by each candidate semantic interpretation branch for that semantic attribute are read, and the attribute values ​​or value ranges that can cause different update results from the candidate semantic interpretation branches are determined as representative predicted answers, and uncertain answers are added to form a set of predicted answers. For conflict confirmation-type candidate follow-up questions, the conflict attribute values ​​recorded in the semantic conflict information are determined as predicted answers, and other answers and uncertain answers that do not belong to each conflict attribute value are added. For risk screening-type candidate follow-up questions, based on the trigger values ​​or value ranges of the corresponding key semantic attributes in the preset high-risk semantic rules, predicted answers that meet the trigger conditions, do not meet the trigger conditions, and are uncertain are generated respectively.

[0084] For example, when the semantic attribute to be confirmed for the candidate follow-up question is duration, and the condition for confirmation of the first candidate semantic interpretation branch is duration less than 10 minutes, and the condition for confirmation of the second candidate semantic interpretation branch is duration greater than 30 minutes, “less than 10 minutes”, “10 to 30 minutes”, “more than 30 minutes” and “uncertain” are generated as predicted answers respectively to simulate the impact of different answers on the conditions for the establishment of candidate semantic interpretation branches and branch status.

[0085] S3.2: The impact of simulated predicted answers on candidate semantic interpretation branches;

[0086] For each predicted answer under each candidate follow-up question, copy the current dynamic consultation semantic state object, establish the corresponding temporary simulated state, and convert the predicted answer into the corresponding simulated semantic attribute, and write it into the candidate semantic grid affected by the candidate follow-up question.

[0087] Following the methods of determining effective semantic attributes, adjusting semantic matching degree, updating distinguishing attributes, and determining candidate states in S2.2, the simulated semantic attributes are matched with the concept discrimination rules of relevant candidate standard medical concepts, and the semantic matching degree, confirmed distinguishing attributes, unconfirmed distinguishing attributes, and candidate states of relevant candidate standard medical concepts are simulated and updated.

[0088] Specifically, when a simulated semantic attribute satisfies a concept support attribute, the semantic matching degree of the corresponding candidate standard medical concept is increased; when a simulated semantic attribute satisfies a concept exclusion attribute, the corresponding semantic matching degree is decreased; when a simulated semantic attribute provides a definite attribute value for an unconfirmed distinguishing attribute, the corresponding unconfirmed distinguishing attribute is updated to a confirmed distinguishing attribute; when a simulated semantic attribute hits an incompatible exclusion condition, the corresponding candidate standard medical concept is set to an excluded state; when the predicted answer is an uncertain answer, the semantic matching degree is not adjusted, and the corresponding unconfirmed distinguishing attribute is retained.

[0089] When the predicted answer is used to confirm a semantic conflict, the predicted answer is matched with the corresponding conflict attribute value. If the predicted answer can determine a valid attribute value, the attribute value is determined as a valid semantic attribute in the temporary simulation state, and the corresponding semantic conflict information is removed. If the predicted answer still cannot determine a valid attribute value, the corresponding semantic conflict information is retained.

[0090] When the predicted response is used to confirm an unconfirmed high-risk attribute, the corresponding simulated semantic attribute is matched with the preset high-risk semantic rule. When the simulated semantic attribute and the existing valid semantic attribute together meet the high-risk state triggering condition, a confirmed high-risk attribute is generated in the temporary simulated state, and the corresponding unconfirmed high-risk attribute is removed. When the simulated semantic attribute can determine that the high-risk state triggering condition is not met, the corresponding unconfirmed high-risk attribute is removed. When the simulated semantic attribute still cannot determine the high-risk state, the corresponding unconfirmed high-risk attribute is retained.

[0091] After completing the simulated update of candidate standard medical concepts, following the compatibility judgment and branch construction method in step S2.3, semantically compatible candidate standard medical concepts are recombined from the active candidate standard medical concepts after the simulated update to form simulated candidate semantic interpretation branches. Based on the concept discrimination results and effective semantic attributes after the simulated update, the satisfied conditions, pending conditions, branch query features, and disambiguation content of the simulated candidate semantic interpretation branches are re-determined. For distinguishing attributes confirmed by simulated semantic attributes, the corresponding pending conditions are updated to satisfied conditions; for candidate standard medical concepts set to exclusion, the corresponding candidate standard medical concepts are no longer added to the simulated candidate semantic interpretation branches.

[0092] Compare the active candidate standard medical concepts, candidate semantic interpretation branches, conditions to be confirmed, semantic conflict information, unconfirmed high-risk attributes, and branch query features before and after the simulation update, and determine the reduction ratio of active candidate standard medical concepts, the reduction ratio of candidate semantic interpretation branches, the reduction ratio of matters to be confirmed, and the reduction ratio of the degree of overlap of branch query features.

[0093] The items to be confirmed include confirmation conditions, unresolved semantic conflicts, and unconfirmed high-risk attributes. The degree of overlap in branch query features between two candidate semantic interpretation branches is determined by the proportion of shared branch query features to the total number of branch query features after merging and deduplicating the two candidate semantic interpretation branches. The degree of overlap in branch query features between any two different candidate semantic interpretation branches in the candidate semantic interpretation branch set is calculated separately, and the average of these overlaps is determined as the degree of overlap in branch query features for the candidate semantic interpretation branch set. If there are fewer than two candidate semantic interpretation branches, the corresponding degree of overlap in branch query features is set to 0. The reduction ratio of the degree of overlap in branch query features is determined by the proportion of the difference in the degree of overlap in branch query features before and after the simulated update to the degree of overlap in branch query features before the simulated update. If the degree of overlap in branch query features before the simulated update is 0, the corresponding reduction ratio is set to 0.

[0094] The four ratios mentioned above are summed according to preset benefit weights to obtain the simulated disambiguation effect corresponding to each predicted answer. Then, based on the occurrence ratio of the corresponding predicted answer in historical consultation answers, the simulated disambiguation effects corresponding to each predicted answer under the same candidate follow-up question are weighted and summed to obtain the disambiguation benefit of the candidate follow-up question. When corresponding historical statistical data is unavailable, all predicted answers are given the same weight.

[0095] It should be noted that the preset benefit weights correspond to the reduction ratio of active candidate standard medical concepts, the reduction ratio of candidate semantic interpretation branches, the reduction ratio of pending confirmation items, and the reduction ratio of branch query feature overlap, with the sum of each weight being 1. Different weight combinations were tested using historical consultation verification data with labeled target candidate semantic interpretation branches. Under the condition that the target branch retention rate meets the preset requirements, the weight combination with the lowest average number of follow-up questions and the lowest number of unresolved disambiguation items was selected and consistently used in actual consultations.

[0096] The reduction ratios for active candidate standard medical concepts, candidate semantic interpretation branches, and pending confirmation items are determined based on the proportion of the difference between the corresponding quantities before and after the simulation update to the corresponding quantities before the simulation update. If the corresponding quantities before the simulation update are 0, or if the corresponding quantities after the simulation update do not decrease, the corresponding reduction ratio is set to 0. If the overlap of branch query features after the simulation update is not lower than the overlap of branch query features before the simulation update, the corresponding reduction ratio is set to 0.

[0097] S3.3: Determine the interaction burden and select the target follow-up question.

[0098] Fill the candidate follow-up questions and their corresponding semantic attributes to be confirmed into the preset follow-up question template to form the candidate follow-up question text.

[0099] Based on the content length of the candidate follow-up question text, the degree of semantic repetition with the historical follow-up question text, the number of semantic attributes that require user answers, and the number of follow-up question rounds already executed, the corresponding burden value is determined according to the preset grading rules, and then summarized according to the preset burden weight to obtain the interaction burden of the candidate follow-up question proposition.

[0100] The disambiguation benefit and interaction burden are converted to the same numerical range. When there is an unconfirmed high-risk attribute and there are risk-screening candidate follow-up questions with a disambiguation benefit higher than the interaction burden, the candidate follow-up question with the largest difference between the disambiguation benefit and the interaction burden is selected first from the corresponding risk-screening candidate follow-up questions; when there are no risk-screening candidate follow-up questions that meet the conditions, the candidate follow-up question with the largest corresponding difference is selected from all candidate follow-up questions with a disambiguation benefit higher than the interaction burden.

[0101] The selected candidate follow-up questions and their corresponding texts are identified as target follow-up questions; if there are no candidate follow-up questions whose disambiguation benefits outweigh the interaction burden, a follow-up question end marker is generated.

[0102] It should be noted that the preset grading rules divide each burden factor into three burden levels: low, medium, and high, and set the corresponding burden values ​​to 0, 0.5, and 1, respectively. As one implementation method, the number of characters contained in the generated candidate follow-up text is counted. A number of characters not exceeding 20, exceeding 20 but not exceeding 40, and exceeding 40 correspond to low-burden, medium-burden, and high-burden levels, respectively. The semantic similarity (cosine similarity) between the candidate follow-up text and historical follow-up text is normalized to 0 to 1, and the maximum semantic similarity is taken. A maximum semantic similarity below 0.4, not below 0.4 but not below 0.7, and not below 0.7 correspond to low-burden, medium-burden, and high-burden levels, respectively. The absence of historical follow-up text corresponds to the low-burden level. The number of semantic attributes requiring user responses (1, 2, or more) corresponds to low-burden, medium-burden, and high-burden levels, respectively. The number of follow-up rounds executed (not exceeding 1, 2-3, or 4 or more) corresponds to low-burden, medium-burden, and high-burden levels, respectively. These grading boundaries are pre-validated based on the correspondence between each burden factor and the effective response rate in historical consultation data and remain fixed within the same consultation process.

[0103] The sum of the preset burden weights corresponding to content length, semantic repetition degree (using the aforementioned medical text encoding model to encode candidate follow-up texts and historical follow-up texts respectively, calculating the cosine similarity between corresponding semantic vectors, and determining the normalized maximum cosine similarity as the semantic repetition degree), number of semantic attributes, and number of follow-up rounds is 1. As one implementation method, the four preset burden weights are 0.2, 0.3, 0.3, and 0.2 respectively. Each burden value is multiplied by its corresponding preset burden weight and then summed to obtain the interaction burden, which ranges from 0 to 1. The preset burden weights can be pre-adjusted based on the influence of each burden factor on the effective response rate in historical consultation verification data; an effective response refers to a response from which at least one effective semantic attribute can be determined from the follow-up response.

[0104] S4. Execute follow-up questions, update the candidate semantic lattice set and candidate semantic interpretation branches based on the follow-up question answers, until there are no candidate follow-up questions whose disambiguation benefits are higher than the interaction burden, or the candidate semantic interpretation branches meet the preset interpretation branch termination conditions.

[0105] S4.1: Follow up on the execution goal;

[0106] When a target follow-up question is generated, the target follow-up question and its corresponding text determined in step S3.3 are read, and the target follow-up question text is sent to the user terminal; the user's follow-up question answer is received, and the follow-up question answer is used as the content of a new round of user consultation under the same consultation session number; when a follow-up question end marker is generated, the follow-up question is no longer executed, and the current dynamic consultation semantic state object is determined as the dynamic consultation semantic state object at the end of the follow-up question.

[0107] S4.2: Update the dynamic consultation semantic state object based on the answers to follow-up questions;

[0108] Following the user consultation content parsing and candidate semantic grid formation rules in step S1, follow-up questions and answers are processed to form a new set of current candidate semantic grids.

[0109] Using the new set of current candidate semantic lattices as input for the current round, steps S2.1 to S2.3 are executed sequentially: The new set of current candidate semantic lattices is cross-round associated with historical candidate semantic lattices and historical consultation answers to form a cross-round candidate semantic lattice set; the semantic matching degree, distinguishing attributes, and candidate status of candidate standard medical concepts are updated based on the newly added valid semantic attributes in the current round, and semantic conflict information, confirmed high-risk attributes, and unconfirmed high-risk attributes are updated simultaneously; candidate semantic interpretation branches are reconstructed based on the updated active candidate standard medical concepts, updating the corresponding establishment conditions, branch query features, and content to be disambiguated. Following the encapsulation method in step S2.3, the above update results are encapsulated into an updated dynamic consultation semantic state object and written into the current consultation session. This update is executed based on the user's actual follow-up questions and answers.

[0110] S4.3: Execute the follow-up question in a loop and terminate the process after determining the appropriate option.

[0111] Re-enter the updated dynamic consultation semantic state object into steps S3.1 to S3.3 to regenerate candidate follow-up questions and predicted answers, and determine the disambiguation benefit, interaction burden, and target follow-up question. When a target follow-up question is generated and the candidate semantic interpretation branch does not meet the preset interpretation branch termination condition, continue to execute the new target follow-up question. When a follow-up question end marker is generated or the candidate semantic interpretation branch meets the preset interpretation branch termination condition, end the follow-up question, record the reason for stopping the follow-up question, and determine the current dynamic consultation semantic state object as the dynamic consultation semantic state object at the end of the follow-up question.

[0112] It should be noted that the preset termination conditions for the interpretation branch include that the number of candidate semantic interpretation branches does not exceed one, or that the follow-up questions and answers in consecutive preset rounds do not change the active candidate standard medical concept, candidate semantic interpretation branch, or conditions to be confirmed; wherein, the preset rounds are preferably 2 to 3 rounds, and are determined in advance based on historical consultation verification data.

[0113] S5. After the follow-up questioning, the candidate standard medical concepts that meet the preset medical concept determination conditions are determined as formal medical concepts, and the candidate standard medical concepts that do not meet the preset medical concept determination conditions, their respective candidate semantic interpretation branches, and their establishment conditions are retained.

[0114] S5.1: Determine whether the candidate standard medical concepts meet the preset medical concept determination conditions;

[0115] Read the dynamic consultation semantic state object at the end of the follow-up questioning, and process the active candidate standard medical concepts one by one according to the candidate semantic grid. The preset medical concept determination conditions include any of the following situations: there is only one active candidate standard medical concept in the candidate semantic grid, and there are no unresolved semantic conflicts directly related to the candidate standard medical concept; there are multiple active candidate standard medical concepts in the candidate semantic grid, the difference between the candidate standard medical concept with the highest semantic matching degree and the candidate standard medical concept with the second highest semantic matching degree reaches the preset concept differentiation threshold, the key differentiation attribute used to distinguish the two has been identified, and there are no unresolved semantic conflicts directly related to the candidate standard medical concept with the highest semantic matching degree.

[0116] It should be noted that the preset concept differentiation threshold is preferably set to 0.10–0.30 after the semantic matching degree has been normalized to 0–1. This threshold is determined after comprehensive verification based on the accuracy and candidate retention rates of medical concepts corresponding to different thresholds in historical labeled consultation verification data. The key differentiation attribute is the concept differentiation attribute used in the concept discrimination rules to distinguish between the candidate standard medical concept with the highest semantic matching degree and the candidate standard medical concept with the second highest semantic matching degree.

[0117] S5.2: Define a formal medical concept and retain undetermined candidates;

[0118] For candidate standard medical concepts that meet the preset conditions for determining medical concepts, the corresponding candidate standard medical concepts are determined as formal medical concepts. For candidate semantic lattices that do not meet the preset conditions for determining medical concepts, a unique formal medical concept is not mandatory. The active candidate standard medical concepts are retained, along with the candidate semantic interpretation branches to which each candidate standard medical concept belongs, the conditions for establishment that have been met, and the conditions for establishment that are yet to be confirmed. An indeterminate concept marker is added.

[0119] S5.3: Generate a complete consultation semantic state object.

[0120] By associating formal medical concepts, candidate standard medical concepts marked with conceptual uncertainty, their respective candidate semantic interpretation branches and conditions for establishment with the effective semantic attributes, semantic conflict information, confirmed high-risk attributes, unconfirmed high-risk attributes and reasons for stopping the inquiry at the end of the inquiry, a complete consultation semantic state object is generated.

[0121] S6. Construct deterministic query branches based on formal medical concepts, construct parallel candidate query branches based on the retained candidate semantic interpretation branches, and retrieve the medical knowledge entries corresponding to each query branch.

[0122] S6.1: Construct deterministic query branches and parallel candidate query branches;

[0123] Extract medical entity expressions and attribute values ​​with negative semantic states from valid semantic attributes to form negative information; attributes with unresolved semantic conflicts are not used as query constraints and are retained in the complete consultation semantic state object.

[0124] For each formal medical concept, the formal medical concept is written into the medical concept field, the valid semantic attributes and confirmed high-risk attributes associated with the formal medical concept are written into the semantic attribute constraint field, the negative information associated with the same medical object is written into the exclusion condition field, a deterministic query branch is generated, and the query branch weight of the deterministic query branch is set to 1.

[0125] Read each candidate standard medical concept marked with conceptual uncertainty one by one according to the retained candidate semantic interpretation branches. Write the candidate standard medical concepts in the same candidate semantic interpretation branch into the medical concept field. Write the valid semantic attributes consistent with the candidate semantic interpretation branch and the confirmed high-risk attributes into the semantic attribute constraint field. Write the corresponding negative information into the exclusion condition field to generate parallel candidate query branches. The met conditions and the conditions to be confirmed are retained as the conditions for the parallel candidate query branches. The conditions to be confirmed are not written into the semantic attribute constraint field.

[0126] The initial weights are determined based on the average current semantic matching degree and the proportion of fulfillment of the establishment conditions for each candidate standard medical concept in the parallel candidate query branches. The proportion of fulfillment of the establishment conditions is the ratio of the number of fulfilled establishment conditions to the total number of fulfilled and pending establishment conditions; if no establishment conditions exist, the proportion of fulfillment of the establishment conditions is set to 1. The average current semantic matching degree and the proportion of fulfillment of the establishment conditions are weighted and summarized according to preset weights, and the initial weights of each parallel candidate query branch are normalized to obtain the corresponding query branch weights; when all initial weights are 0, the same weight is set for each parallel candidate query branch.

[0127] It should be noted that the sum of the current average semantic matching degree and the preset weights corresponding to the proportion of conditions met is 1, and is predetermined based on the hit rate of valid medical knowledge entries in historical medical information retrieval verification data.

[0128] S6.2: Generate a set of consultation query branches;

[0129] Assign query branch numbers to each deterministic query branch and parallel candidate query branch, and associate the query branch number, query branch type, medical concept field, semantic attribute constraint field, exclusion condition field, query branch weight, satisfied condition and pending condition to generate a set of consultation query branches.

[0130] S6.3: Retrieve medical knowledge entries according to the query branch;

[0131] Read the consultation query branch set one by one according to the query branch number, and retrieve the corresponding medical knowledge entries from the preset medical information database according to the standard medical concept identifier in the medical concept field. Then, summarize the medical knowledge entries retrieved from the same query branch.

[0132] The medical knowledge entries in the pre-set medical information database include knowledge entry identifiers, standard medical concept identifiers, applicable conditions, exclusion conditions, knowledge sources, knowledge source reliability levels, and knowledge content.

[0133] S6.4: Filter medical knowledge entries and determine the search matching degree;

[0134] The recalled medical knowledge entries are matched with the corresponding query branches. If the standard medical concept identifier of the medical knowledge entry matches at least one standard medical concept identifier in the medical concept field, the applicable conditions of the medical knowledge entry do not conflict with the exclusion condition field, and the exclusion condition of the medical knowledge entry is not matched by the semantic attribute constraint field, the corresponding medical knowledge entry is retained; otherwise, the corresponding medical knowledge entry is excluded.

[0135] The search matching degree is determined based on the proportion of standard medical concepts matching the medical knowledge entries to the total number of standard medical concepts in the medical concept field, and the proportion of the applicable conditions of the medical knowledge entries being satisfied by the semantic attribute constraint field. When there are no applicable conditions for a medical knowledge entry, the proportion of applicable conditions satisfied is determined to be 1.

[0136] The retained medical knowledge entries are associated with their corresponding query branch numbers, query branch types, query branch weights, retrieval matching degrees, knowledge sources, applicable conditions, and exclusion conditions. When the same medical knowledge entry is retrieved from different query branches, the association between the medical knowledge entry and each query branch is retained separately.

[0137] It should be noted that the knowledge applicability conditions refer to the semantic attribute conditions that a medical knowledge entry must meet to be applicable to the corresponding medical object, including body part, time attribute, severity, inducing factors, accompanying manifestations, or other limiting attributes; the knowledge exclusion conditions refer to the semantic attribute conditions that a medical knowledge entry is not applicable to the corresponding medical object, including negative information, attribute values, or attribute combinations that are incompatible with the medical knowledge content.

[0138] S6.5: Merge duplicate entries and perform a preliminary sort.

[0139] Within the same query branch, medical knowledge entries with the same knowledge entry identifier or the same standardized knowledge content are merged, while retaining the corresponding knowledge source and applicable conditions before merging.

[0140] The knowledge source reliability level is converted into a source reliability score within the range of 0 to 1, and a preliminary ranking score is determined based on the retrieval matching degree and the source reliability score. As one implementation method, the source reliability scores corresponding to the high, medium, and low knowledge source reliability levels are set to 1, 0.7, and 0.4, respectively; the retrieval matching degree and source reliability score are weighted and summed according to a preset ranking weight to obtain the preliminary ranking score.

[0141] The sum of the preset sorting weights is 1, and the sorting position of the valid medical knowledge entries in the historical medical information retrieval verification data is predetermined.

[0142] The medical knowledge items in each query branch are sorted according to the preliminary ranking scores. The sorted medical knowledge items and the corresponding query branch information are then encapsulated to generate a candidate consultation evidence set.

[0143] S7. Based on the applicability of medical knowledge items in deterministic query branches and parallel candidate query branches, generate common consultation information, conditional consultation information with valid conditions, and branch difference information.

[0144] Specifically, the process involves reading the candidate consultation evidence set and the complete consultation semantic state object. Each medical knowledge item is associated with its corresponding deterministic query branch or parallel candidate query branch according to the query branch number. The applicability and exclusion conditions of each medical knowledge item are matched with the semantic attribute constraints, exclusion conditions, and validity conditions of the corresponding query branch to determine the applicability status of each medical knowledge item within the corresponding query branch. Medical knowledge items in the applicability status are grouped according to medical objects and semantic matters. When a medical knowledge item originates from a deterministic query branch, or when the same medical knowledge content is in the applicability status in all relevant parallel candidate query branches corresponding to the same medical object, the corresponding medical knowledge content is identified as shared consultation information. When medical knowledge content is only in one or more of the same medical object, the grouping process is repeated. When a relevant parallel candidate query branch is in an applicable state, the corresponding medical knowledge content, its corresponding parallel candidate query branch, and the corresponding met and pending conditions for establishment are associated to form conditional consultation information with establishment conditions. When medical knowledge entries in different parallel candidate query branches form medical knowledge content that cannot be established simultaneously for the same semantic matter, the different medical knowledge content, their corresponding parallel candidate query branches, and their corresponding conditions for establishment are retained to form branch difference information, and no medical knowledge content is determined as a definitive consultation conclusion applicable to all relevant parallel candidate query branches. The shared consultation information, conditional consultation information, and branch difference information are associated with the corresponding medical knowledge entries and query branches to form consultation information results after cross-branch semantic control.

[0145] This embodiment also provides an intelligent consultation system based on semantic understanding, including:

[0146] The consultation semantic parsing module is used to parse the user's consultation content and form a set of candidate semantic lattices;

[0147] The semantic branching module is used to update the semantic state of candidate standard medical concepts and form candidate semantic interpretation branches;

[0148] The follow-up question decision module is used to generate candidate follow-up questions and predict answers, and select target follow-up questions based on disambiguation benefits and interaction burden;

[0149] The follow-up question update module is used to update the candidate semantic lattice set and candidate semantic interpretation branches based on the follow-up question answers, and to determine the formal medical concept;

[0150] The branch retrieval module is used to construct deterministic query branches and parallel candidate query branches, and retrieve the corresponding medical knowledge entries;

[0151] The consultation information generation module is used to generate common consultation information, conditional consultation information, and branch difference information based on the applicability of branches of medical knowledge items.

[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An intelligent consultation method based on semantic understanding, characterized in that, include: The system analyzes user consultation content, extracts semantic attributes, identifies multiple candidate standard medical concepts from the medical concept library, and forms a candidate semantic lattice set. By combining historical consultation responses to update the semantic matching degree and unconfirmed distinguishing attributes of each candidate standard medical concept, semantically compatible candidate standard medical concepts are combined into candidate semantic interpretation branches with valid conditions. Based on the candidate semantic interpretation branches, candidate follow-up questions and predicted answers are generated. The impact of each predicted answer on the candidate semantic interpretation branches is simulated to determine the disambiguation benefit of each candidate follow-up question. The target follow-up question is selected based on the interaction burden of the candidate follow-up questions. The execution target is followed up with questions, and the candidate semantic lattice set and candidate semantic interpretation branches are updated based on the follow-up questions and answers, until there are no candidate follow-up questions whose disambiguation benefits are higher than the interaction burden, or the candidate semantic interpretation branches meet the preset interpretation branch termination conditions. After the follow-up questioning, candidate standard medical concepts that meet the preset conditions for determining medical concepts will be determined as formal medical concepts, while candidate standard medical concepts that do not meet the preset conditions for determining medical concepts, their respective candidate semantic interpretation branches, and their conditions for establishment will be retained. Deterministic query branches are constructed based on formal medical concepts, parallel candidate query branches are constructed based on the retained candidate semantic interpretation branches, and medical knowledge entries corresponding to each query branch are retrieved. Based on the applicability of medical knowledge items in deterministic query branches and parallel candidate query branches, common consultation information, conditional consultation information with valid conditions, and branch difference information are generated.

2. The intelligent consultation method based on semantic understanding as described in claim 1, characterized in that: The process of forming a candidate semantic lattice set includes: dividing the user's consultation content into multiple original consultation expression units, identifying the medical entity expressions and corresponding semantic states in each original consultation expression unit, extracting the semantic attributes associated with the medical entity expressions, and generating expression semantic records; determining multiple candidate standard medical concepts from a medical concept library based on the expression semantic records, determining the initial semantic matching degree, confirmed supporting attributes, confirmed exclusion attributes, confirmed distinguishing attributes, and unconfirmed distinguishing attributes of each candidate standard medical concept according to concept discrimination rules, and associating the expression semantic records with the discrimination results of the corresponding candidate standard medical concepts to form a candidate semantic lattice set.

3. The intelligent consultation method based on semantic understanding as described in claim 2, characterized in that: The process of updating the semantic matching degree and unconfirmed distinguishing attributes of each candidate standard medical concept by combining historical consultation answers, and combining semantically compatible candidate standard medical concepts into candidate semantic interpretation branches with valid conditions, includes: obtaining historical candidate semantic lattices and historical consultation answers corresponding to the candidate semantic lattice set; associating candidate semantic lattices describing the same medical object across rounds to form a cross-round candidate semantic lattice set; identifying semantic conflicts in the cross-round candidate semantic lattice set; determining semantic attributes without semantic conflicts as valid semantic attributes; and updating the semantic matching degree, confirmed distinguishing attributes, unconfirmed distinguishing attributes, and candidate status of each candidate standard medical concept according to valid semantic attributes and concept discrimination rules to form an updated candidate semantic lattice set; extracting active candidate standard medical concepts from the updated candidate semantic lattice set; combining semantically compatible candidate standard medical concepts according to concept relationship rules in the medical concept base; and determining the conditions for fulfillment and the conditions for unconfirmation according to the corresponding concept discrimination results to form a candidate semantic interpretation branch set.

4. The intelligent consultation method based on semantic understanding as described in claim 3, characterized in that: The process of generating candidate follow-up questions and predicted answers based on candidate semantic interpretation branches refers to: comparing the conditions for confirmation of each candidate semantic interpretation branch, generating candidate follow-up questions based on unconfirmed distinguishing attributes, semantic conflict information, and unconfirmed high-risk attributes, and generating a set of predicted answers based on the unconfirmed semantic attributes and attribute value requirements corresponding to the candidate follow-up questions.

5. The intelligent consultation method based on semantic understanding as described in claim 4, characterized in that: The simulation of the impact of each predicted answer on the candidate semantic interpretation branches determines the disambiguation benefit of each candidate follow-up question and selects the target follow-up question based on the interaction burden of the candidate follow-up question. This includes: establishing a temporary simulated state for each predicted answer; updating the semantic matching degree, distinguishing attributes, and candidate state of the candidate standard medical concepts based on the predicted answers, and reconstructing the candidate semantic interpretation branches; comparing the degree of overlap between the active candidate standard medical concepts, candidate semantic interpretation branches, pending confirmation items, and branch query features before and after the simulation update to determine the simulated disambiguation effect of each predicted answer and summarizing the disambiguation benefit of the candidate follow-up question; determining the interaction burden based on the content length of the candidate follow-up question text, the degree of semantic repetition with the historical follow-up question text, the number of semantic attributes to be answered, and the number of follow-up question rounds already executed; selecting the target follow-up question from the candidate follow-up question propositions whose disambiguation benefit is higher than the interaction burden; and generating a follow-up question end marker when there are no candidate follow-up question propositions that meet the conditions.

6. The intelligent consultation method based on semantic understanding as described in claim 5, characterized in that: The step of updating the candidate semantic lattice set and candidate semantic interpretation branches based on follow-up questions and answers includes: executing the target follow-up question and receiving the follow-up question answer; forming a new round of current candidate semantic lattice set based on the follow-up question answer; associating the new round of current candidate semantic lattice set with historical candidate semantic lattices and historical consultation answers across rounds; updating the semantic state of candidate standard medical concepts and reconstructing candidate semantic interpretation branches; redetermining the target follow-up question based on the updated candidate semantic lattice set and candidate semantic interpretation branches, until a follow-up question end marker is generated or the candidate semantic interpretation branch meets the preset interpretation branch termination condition.

7. The intelligent consultation method based on semantic understanding as described in claim 6, characterized in that: The process of determining candidate standard medical concepts that meet the preset medical concept determination conditions as formal medical concepts, and retaining candidate standard medical concepts that do not meet the preset medical concept determination conditions, their respective candidate semantic interpretation branches, and their establishment conditions, includes: determining whether candidate standard medical concepts meet the preset medical concept determination conditions based on the number of active candidate standard medical concepts, semantic matching difference, key distinguishing attributes, and semantic conflict information in each candidate semantic grid at the end of the follow-up question, and determining candidate standard medical concepts that meet the preset medical concept determination conditions as formal medical concepts; for candidate semantic grids that do not meet the preset medical concept determination conditions, retaining active candidate standard medical concepts, their respective candidate semantic interpretation branches, the establishment conditions that have been met, and the establishment conditions that are yet to be confirmed.

8. The intelligent consultation method based on semantic understanding as described in claim 7, characterized in that: The process of constructing deterministic query branches based on formal medical concepts, constructing parallel candidate query branches based on retained candidate semantic interpretation branches, and retrieving medical knowledge entries corresponding to each query branch includes: constructing deterministic query branches based on formal medical concepts and associated effective semantic attributes and negation information; combining active candidate standard medical concepts, corresponding establishment conditions, and consistent effective semantic attributes into parallel candidate query branches according to the retained candidate semantic interpretation branches; determining the query branch weights of parallel candidate query branches based on the semantic matching degree and the proportion of establishment conditions satisfied by the candidate standard medical concepts; associating the deterministic query branches and parallel candidate query branches to form a consultation query branch set; retrieving and filtering medical knowledge entries from a preset medical information database based on the medical concept fields, semantic attribute constraint fields, and exclusion condition fields in the consultation query branch set; merging and sorting duplicate medical knowledge entries within the same query branch to generate a candidate consultation evidence set.

9. The intelligent consultation method based on semantic understanding as described in claim 8, characterized in that: The process of forming shared consultation information, conditional consultation information with conditions for establishment, and branch difference information includes: determining the applicability of medical knowledge items in corresponding query branches based on their applicability conditions, exclusion conditions, and establishment conditions; identifying medical knowledge content applicable to deterministic query branches or all relevant parallel candidate query branches as shared consultation information; associating medical knowledge content applicable only to some relevant parallel candidate query branches with their corresponding establishment conditions to form conditional consultation information; and retaining the correspondence between each medical knowledge content and its respective parallel candidate query branch for conflicting medical knowledge content on the same semantic matter in different parallel candidate query branches to form branch difference information.

10. A semantic understanding-based intelligent consultation system, based on the semantic understanding-based intelligent consultation method according to any one of claims 1 to 9, characterized in that, include: The consultation semantic parsing module is used to parse user consultation content and form a set of candidate semantic lattices; The semantic branching module is used to update the semantic state of candidate standard medical concepts and form candidate semantic interpretation branches; The follow-up question decision module is used to generate candidate follow-up questions and predicted answers, and select target follow-up questions based on disambiguation benefits and interaction burden; The follow-up question update module is used to update the candidate semantic lattice set and candidate semantic interpretation branches based on the follow-up question answers, and to determine the formal medical concept; The branch retrieval module is used to construct deterministic query branches and parallel candidate query branches, and retrieve the corresponding medical knowledge entries; The consultation information generation module is used to generate common consultation information, conditional consultation information, and branch difference information based on the applicability of branches of medical knowledge items.

Images