Question and answer method and system based on knowledge chain training artificial intelligence model

Abstract

The invention provides a question and answer method and system based on a knowledge chain training artificial intelligence model, and the method comprises the steps: organizing domain fact data to generate a knowledge chain set based on a knowledge tracing triggering mechanism which is used for starting a hierarchical tracing process from an answer to basic data in a knowledge chain methodology; the knowledge chain set comprises a plurality of knowledge units and traceability association paths among the knowledge units; calling a support relationship judgment standard in a knowledge chain methodology, and executing a traceability support operation on each knowledge unit to obtain a corresponding support type; inputting the knowledge chain set and the corresponding support type into an artificial intelligence training model for associated training to generate a knowledge output screening rule; performing adaptive optimization on the knowledge output screening rule through the field verification data to obtain a trained artificial intelligence model based on the knowledge chain; and in response to inputting the target question to the artificial intelligence model, obtaining a question and answer result for the target question.

Description

Technical Field

[0001] This disclosure relates to the field of computer processing technology, and in particular to a question-answering method and system based on training an artificial intelligence model using a knowledge chain. Background Technology

[0002] Currently, training artificial intelligence generally adopts a "one-dimensional" training model, directly utilizing data for training and hoping that large models will discover patterns on their own. However, this model has significant drawbacks. It lacks in-depth exploration of the inherent logic and hierarchical structure of knowledge, resulting in low efficiency during model training and difficulty in accurately grasping the relationships between knowledge. Especially when facing complex domain problems, due to the lack of a systematic knowledge organization and tracing mechanism, the model's output is often inaccurate and incomplete, failing to meet the needs of practical applications. Summary of the Invention

[0003] In view of the aforementioned problems, this disclosure provides a question-answering method and system for training artificial intelligence models based on knowledge chains, aiming to improve training efficiency and question-answering quality.

[0004] In conjunction with the first aspect of the present invention, an embodiment of the present invention provides a question-answering method based on training an artificial intelligence model using a knowledge chain, comprising: Based on the knowledge tracing triggering mechanism in the knowledge chain methodology, which is used to initiate the hierarchical tracing process from the answer to the basic data, a knowledge chain set is generated by organizing domain fact data. The knowledge chain set includes multiple knowledge units and tracing association paths between knowledge units. By invoking the support relationship determination criteria in the knowledge chain methodology, a traceability support operation is performed on each knowledge unit in the knowledge chain set to obtain the support type corresponding to the knowledge unit. The knowledge chain set and the corresponding support type are input into the artificial intelligence training model for association training to generate knowledge output filtering rules. The knowledge output filtering rules are adapted and optimized using domain validation data to obtain a trained knowledge chain-based artificial intelligence model. In response to inputting a target question into the artificial intelligence model, the artificial intelligence model outputs a question-and-answer result for the target question.

[0005] In conjunction with a second aspect of the present invention, an embodiment of the present invention provides a question-answering system based on a knowledge chain-trained artificial intelligence model, comprising: A memory having a computer program stored thereon; a processor for executing the computer program in the memory to implement the steps of the method of any one of the first aspects.

[0006] Through the above-described technical solution, this disclosure can achieve at least the following effective effects: The knowledge tracing trigger mechanism organizes domain factual data in an orderly manner, generating a knowledge chain set including multiple knowledge units and tracing paths. This provides a structured knowledge framework for model training, supporting relationship determination criteria in performing tracing support operations on knowledge units, clarifying support types, and strengthening the internal logic of knowledge. The knowledge output filtering rules generated through associated training make the model output more accurate and targeted. After domain validation data adaptation and optimization, the model can better adapt to real-world scenarios, significantly improving question-answering efficiency and accuracy. Faced with a target question, it can quickly output high-quality question-answering results, not only providing users with reliable information but also improving training efficiency and question-answering quality.

[0007] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0008] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a question-answering method based on knowledge chain training of an artificial intelligence model provided in an embodiment of the present invention.

[0009] Figure 2 This is one implementation provided by an embodiment of the present invention. Figure 1 The flowchart for step S11.

[0010] Figure 3 This is one implementation provided by an embodiment of the present invention. Figure 1 The flowchart for step S12.

[0011] Figure 4 This is one implementation provided by an embodiment of the present invention. Figure 1 The flowchart for step S13.

[0012] Figure 5 This is a schematic diagram of exemplary hardware and software components of a question-answering system based on a knowledge chain-trained artificial intelligence model, provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0015] This invention provides a question-answering method based on training an artificial intelligence model using a knowledge chain. See also... Figure 1 As shown, the question-answering method based on training an artificial intelligence model using a knowledge chain includes the following steps.

[0016] In step S11, based on the knowledge tracing triggering mechanism in the knowledge chain methodology used to initiate the hierarchical tracing process from the answer to the basic data, domain fact data is organized to generate a knowledge chain set, which includes multiple knowledge units and tracing association paths between knowledge units; The knowledge tracing triggering mechanism is a reverse reasoning engine based on the knowledge chain methodology. It initiates a hierarchical tracing process from the answer to the original data through preset semantic association rules, ensuring that the generation of knowledge units conforms to the domain's logical chain. The knowledge chain set is a networked collection of knowledge units formed by the structured organization of domain factual data. It includes knowledge units (such as concepts, facts, and rules) and the tracing association paths between them (such as causal relationships and dependency relationships).

[0017] In this embodiment, natural language processing techniques (such as dependency parsing and semantic role labeling) are used to parse the original text data and extract entity, attribute, and relation triples. For example, in the medical field, from the sentence "diabetic patients need to control their blood sugar," the triple "diabetic patient (entity) - need to control (relation) - blood sugar (entity)" can be extracted. Subsequently, using hierarchical tracing rules in the knowledge chain methodology (such as tracing back from the conclusion to the premise and from the result to the cause), the triples are expanded into a multi-level knowledge chain. For example, further tracing "control blood sugar (entity) - dependency (relation) - insulin injection (entity)" forms a complete chain of "diabetic patient → control blood sugar → insulin injection." This process uses a graph database (such as Neo4j) to store knowledge units and associated paths, ensuring the integrity and traceability of the knowledge chain. The final generated knowledge chain set contains tens of thousands of knowledge units and millions of associated paths.

[0018] In step S12, the support relationship determination criteria in the knowledge chain methodology are invoked to perform a traceability support operation on each knowledge unit in the knowledge chain set to obtain the support type corresponding to the knowledge unit; Among them, the criteria for determining support relationships are logical rules defined in the knowledge chain methodology, used to quantify the strength of support between knowledge units (such as direct support, indirect support, and no support). The support tracing operation involves performing reverse reasoning on knowledge units based on the criteria to identify their dependent upstream knowledge units and support types (such as theoretical support, experimental support, and case support).

[0019] In this embodiment, firstly, a rule engine (such as Drools) applies a judgment criterion: if the generation of knowledge unit B depends on knowledge unit A (e.g., a mathematical theorem depends on an axiom), it is judged as "direct support"; if B depends on C, and C depends on A, it is judged as "indirect support". For example, in the field of physics, "Newton's second law (F=ma)" directly supports the "kinetic energy theorem", while the "law of conservation of mass" indirectly supports the kinetic energy theorem. Secondly, a pre-trained language model (such as BERT) is used to calculate the semantic similarity between knowledge units to assist in correcting the judgment of the rule engine. For example, if the cosine similarity of the embedding vectors of two knowledge units exceeds a threshold (e.g., 0.8), there may be implicit support relationships not covered by the rules. Finally, each knowledge unit is labeled as "strong support", "weak support", or "no support", forming a support type tag library.

[0020] In step S13, the knowledge chain set and the corresponding support type are input into the artificial intelligence training model for association training to generate knowledge output filtering rules; Among them, associative training involves inputting a set of knowledge chains (structured knowledge) and supporting types (logical constraints) into an artificial intelligence model, and simultaneously optimizing knowledge representation and logical reasoning capabilities through a multi-task learning framework. Knowledge output filtering rules: Decision rules generated after model training are used to select the knowledge units that best match the supporting types as answer candidates in question-answering scenarios.

[0021] In this embodiment, a Transformer-based encoder-decoder model can be used. The input layer receives a set of knowledge chains (serialized knowledge units and associated paths) and supporting type labels, capturing long-distance dependencies through a self-attention mechanism. For example, the model can learn that knowledge units with "direct support" should be prioritized in answer generation. During training, the loss function consists of two parts: a knowledge representation loss (such as contrastive learning loss, ensuring that the embedding vectors of similar knowledge units are close) and a logical constraint loss (such as cross-entropy loss, forcing the model output to conform to the supporting type labels). Through joint optimization, the model learns to balance knowledge relevance and logical rationality when generating answers. For example, when a user asks "How to treat diabetes?", the model will prioritize knowledge units of the "direct support" type (such as "insulin injection") rather than "indirect support" (such as "exercise advice").

[0022] In step S14, the knowledge output filtering rules are adapted and optimized using domain verification data to obtain a trained knowledge chain-based artificial intelligence model. Domain validation data consists of labeled datasets from the target domain (such as law or finance), including real user queries and corresponding standard answers, used to evaluate model performance. Adaptation optimization involves adjusting model parameters using domain adaptation techniques (such as parameter fine-tuning and data augmentation) to optimize metrics such as accuracy and recall in the target domain.

[0023] In this embodiment, firstly, the current performance of the model (e.g., F1 score) is calculated using domain validation data. If the metric is below a threshold (e.g., 0.85), fine-tuning is initiated. During fine-tuning, the underlying parameters of the model (e.g., the word embedding layer in BERT) are fixed, and only the parameters of the top-level classifier (e.g., the weights of the fully connected layer) are adjusted to avoid overfitting. Secondly, if the amount of validation data is insufficient (e.g., less than 1000 records), data augmentation techniques (e.g., back-translation, synonym replacement) are used to expand the dataset. For example, "diabetic patients need to control their blood sugar" is back-translated as "blood sugar control is crucial for diabetic patients," generating semantically similar but differently expressed samples. Furthermore, a domain adaptation algorithm (e.g., CORAL) is introduced to improve the model's generalization ability by minimizing the feature distribution differences between the source domain (general knowledge) and the target domain (professional knowledge).

[0024] In step S15, in response to inputting a target question into the artificial intelligence model, the artificial intelligence model outputs a question-and-answer result for the target question.

[0025] The Retrieval Enhanced Generation (RAG) framework is employed. First, Elasticsearch is used to retrieve semantically similar knowledge units (such as Top-5 related fragments) from the knowledge base. For example, for "What to do about diabetes?", knowledge units such as "insulin injection" and "dietary control" are retrieved. Second, the search results are concatenated with the original question to form a prompt, which is then input into a generation model (such as GPT-3.5) to generate candidate answers. For example, the prompt might be: "Question: What to do about diabetes? Related knowledge: 1. Insulin injection; 2. Dietary control. Please generate the answer:". Finally, the candidate answers are ranked using knowledge output filtering rules, and the answer with the highest support type ("direct support") and knowledge unit coverage is selected as the final output. For example, the model might generate "Diabetes requires insulin injection to control blood sugar, combined with a low-sugar diet."

[0026] The aforementioned technical solution, through a knowledge tracing trigger mechanism, can systematically organize domain factual data, generating a knowledge chain set including multiple knowledge units and tracing association paths. This provides a structured knowledge framework for model training, thereby supporting relationship determination criteria in performing tracing support operations on knowledge units, clarifying support types, and strengthening the internal logic of knowledge. The knowledge output filtering rules generated through associated training make the model output more accurate and targeted. After domain validation data adaptation and optimization, the model can better adapt to real-world scenarios, significantly improving question-answering efficiency and accuracy. Faced with a target question, it can quickly output high-quality question-answering results, not only providing users with reliable information but also improving training efficiency and question-answering quality.

[0027] In a preferred embodiment, see Figure 2 As shown, in step S11, the knowledge tracing triggering mechanism based on the knowledge chain methodology, which initiates the hierarchical tracing process from answer to basic data, organizes domain fact data to generate a knowledge chain set, including: In step S111, the core association dimensions corresponding to the knowledge tracing triggering mechanism in multiple tracing directions are extracted; Among them, the core association dimension is a multi-dimensional indicator defined in the knowledge tracing triggering mechanism, used to quantify the strength of association between knowledge units, including semantic similarity, logical dependency, causal relationship strength, etc., used to initiate the hierarchical tracing process.

[0028] In this embodiment, firstly, a pre-trained language model (such as Sentence-BERT) is used to calculate the semantic similarity between any two texts in the domain factual data, generating a numerical score of 0-1 (1 indicating complete similarity). Secondly, dependency parsing is used to parse the text structure, identify logical dependencies (such as "if-then" and "because-therefore"), and count the frequency of occurrence of these dependencies as the logical dependency score. For example, in the medical field, the causal strength of "hypertension leads to heart disease" can be calculated by combining co-occurrence frequency (such as the number of times "hypertension" and "heart disease" appear together in the literature) and temporal order (hypertension was diagnosed earlier than heart disease). Finally, the semantic similarity, logical dependency, and causal strength are weighted and fused (the weights are trained using domain expert-annotated data) to form a core association dimension vector.

[0029] In step S112, based on the independence of knowledge content and the matching degree of the correlation dimension, the domain fact data is split according to the core correlation dimension to obtain multiple independent knowledge units. Each independent knowledge unit includes complete knowledge content, source identifier and application scenario description. Among them, an independent knowledge unit is the smallest knowledge entity with complete knowledge content, a clear source, and an application scenario, and it is the basic building block of a knowledge chain set.

[0030] In this embodiment, firstly, a similarity matrix (e.g., cosine similarity) is calculated for each text segment in the domain fact data based on the core association dimension vector, and the DBSCAN clustering algorithm is applied to group texts with similarity scores higher than a threshold (e.g., 0.7) into the same category. For example, multiple text segments describing "diabetes symptoms" are clustered into a "symptom category" knowledge unit. Secondly, independence is verified for each clustering result: if the association dimension scores of a certain text segment with other texts are all lower than the threshold, it is split into independent units.

[0031] For example, while "diabetes requires blood sugar control" and "methods for controlling blood sugar" are related, the latter can exist independently, so they are separated into two units. Finally, each independent unit is labeled with a source identifier (such as document ID, webpage URL) and an application scenario description (such as "clinical diagnosis" or "patient education") to ensure knowledge traceability and scenario adaptability.

[0032] In step S113, content parsing is performed on each independent knowledge unit to extract feature elements used to characterize the core attributes of the knowledge, wherein the extraction scope of the feature elements covers the entire content of the independent knowledge unit; Among them, feature elements are key information used to characterize the core attributes of an independent knowledge unit, including entities, attributes, relationships, rules, etc.

[0033] In this embodiment, firstly, the BiLSTM-CRF model is used to identify entities (such as disease names and drug names) and attributes (such as symptoms and dosages) within independent knowledge units. For example, from the statement "diabetic patients need to inject 10 units of insulin daily," entities and attributes such as "diabetes (disease)," "insulin (drug)," and "10 units (dosage)" are extracted. Secondly, the OpenIE algorithm is used to extract relationships between entities (such as "treatment" and "side effects"), forming triples of "diabetes-treatment-insulin." Finally, the triples are semantically normalized (e.g., unifying "treatment" and "medication" as the same relationship type), generating a structured feature element library. For example, all triples involving "diabetes-treatment" are categorized into the same feature set.

[0034] In step S114, based on the feature elements, starting from the derivation basis of any of the independent knowledge units, the corresponding core viewpoints or data source vouchers in other independent knowledge units are matched to form a directional association and establish a traceability association path between multiple independent knowledge units. Among them, the traceability path is a logical chain between independent knowledge units based on the reasoning basis, core viewpoints or data sources, which is used to support the hierarchical traceability of the knowledge chain.

[0035] In this embodiment, firstly, starting from the derivation basis of any independent knowledge unit (e.g., "according to clinical study A"), other units containing the same derivation basis or core viewpoint (e.g., "clinical study A proves insulin is effective") are searched in the feature element database. For example, starting from the unit "insulin dosage needs to be adjusted according to blood glucose", the unit "blood glucose monitoring is the basis for adjusting dosage" is retrieved, forming a preliminary path of "dosage adjustment → blood glucose monitoring". Secondly, subsequent units are recursively matched using a path expansion algorithm (e.g., breadth-first search) until no new association can be found or the path length limit (e.g., 5 units) is reached. Finally, the path is logically verified: if there is a contradiction in the path (e.g., "insulin increases blood glucose" and "insulin decreases blood glucose"), it is marked as an invalid path and removed. The final generated path must satisfy the following requirements: the association dimension type must be consistent (e.g., both are causal relationships) and the feature element matching degree must be higher than a threshold (e.g., 0.8).

[0036] In step S115, the knowledge chain set for each domain category is determined based on the association dimension type, matched feature elements, and path length of each traceability path, wherein the path length is determined based on the number of associated independent knowledge units. In this embodiment, firstly, paths are grouped according to the type of association dimension (such as causal relationship, dependency relationship), for example, all causal paths are grouped into a "causal chain" set. Secondly, the overlap of feature elements in each group of paths is calculated (such as the frequency of "insulin" appearing in multiple paths). If the overlap is higher than a threshold (such as 0.6), they are merged into the same knowledge chain set. For example, a "diabetes treatment chain" may contain multiple sub-paths such as "insulin treatment" and "oral medication treatment". Finally, core knowledge chains are selected based on path length (number of associated units): paths with a length of 3-5 are retained (paths that are too short lack depth, while paths that are too long are prone to introducing noise), forming the final knowledge chain set. For example, in the medical field, a complete chain of "diabetes symptoms → diagnosis → treatment → complication prevention" may be generated.

[0037] In step S116, an intra-chain index directory is established for each knowledge chain set, and a cross-domain index directory is established for the knowledge chain sets classified by each domain according to the association dimension, thereby generating the knowledge chain set. The index directory includes an independent knowledge unit identifier, a summary of core feature elements, and information on the starting node of the tracing association path.

[0038] The index directory is used to quickly retrieve the metadata structure of the knowledge chain collection, including intra-chain indexes (retrieval of units within a single collection) and cross-domain indexes (retrieving associations between collections in multiple domains).

[0039] In this embodiment, for intra-chain indexes, an inverted index is generated for each independent unit in the knowledge chain set, recording the unit identifier, a summary of core feature elements (e.g., "Insulin: Treatment of diabetes, dosage 10 units"), and information on the starting node of the tracing association path (e.g., "Tracing from the 'Diabetes Diagnosis' unit to the 'Symptoms' unit"). For example, when a user queries "insulin treatment," the inverted index can directly locate the relevant unit. For cross-domain indexes, an association graph is constructed between knowledge chain sets: if two sets share feature elements (e.g., "hypertension" appears in both the "Cardiovascular Disease Chain" and the "Metabolic Syndrome Chain"), an edge connection is established, and the association dimension type is labeled (e.g., "Comorbidity"). Finally, the index directory supports millisecond-level retrieval response, such as quickly locating the "Insulin Dosage Adjustment" unit in the "Diabetes Treatment Chain" in a question-and-answer scenario, improving model efficiency.

[0040] In a preferred embodiment, step S115, determining the knowledge chain set for each domain category based on the association dimension type, matched feature elements, and path length of each traceability path, includes: In step S1151, an initial knowledge chain structure is formed based on the association dimension type, matched feature elements, and path length of each traceability path. The initial knowledge chain structure is a preliminary knowledge network prototype constructed based on the association dimension type, matching feature elements and path length of the tracing association path. It has not yet verified the logical coherence and includes the topological relationship between independent knowledge units and their association paths.

[0041] In this embodiment, firstly, the tracing paths are classified according to the type of association dimension (such as causal relationship, dependency relationship, co-occurrence relationship). For example, "diabetes → insulin treatment" is classified as a causal relationship path, and "hypertension → cardiovascular disease" is classified as a co-occurrence relationship path. Secondly, matching feature elements (such as entities, attributes, rules) in the path are extracted to generate feature vectors (e.g., the entity vector of "diabetes" includes dimensions such as "symptoms", "treatment", and "complications"). Then, the path length (the number of associated independent knowledge units) is calculated. If the path length exceeds a threshold (e.g., 8 units), it is split into multiple sub-paths (e.g., "diabetes → symptoms → diagnosis → treatment" is split into "diabetes → symptoms" and "symptoms → diagnosis → treatment"). Finally, the independent knowledge units in the path are used as nodes, and the association relationships are used as edges to construct an initial knowledge chain of a directed graph structure. For example, in the medical field, a linear chain of "symptoms → diagnosis → treatment" or a mesh chain of "comorbidity → complications" is formed.

[0042] In step S1152, starting from any of the independent knowledge units, the basic data knowledge units are deduced in reverse along the traceability path, the logical coherence in the deduction process is checked, and the traceability path verification result of the initial knowledge chain structure is obtained. Among them, the traceability path verification result is the evaluation result of checking the logical coherence in the initial knowledge chain structure through reverse deduction, and marking whether there are semantic breaks, logical contradictions or missing deduction basis in the path.

[0043] In this embodiment, starting from the end node of the initial knowledge chain (e.g., "insulin treatment"), the derivation proceeds backward along the associated path to the basic data node (e.g., "diabetes symptoms"). For example, the derivation process is "insulin treatment ← dosage adjustment ← blood glucose monitoring ← diabetes symptoms". During the derivation process, the completeness of the derivation basis for each step is checked: if the derivation basis for a certain step (e.g., "blood glucose monitoring results") is not clearly marked in the path or contradicts the context (e.g., "blood glucose monitoring" is confused with "blood pressure monitoring"), it is marked as a logical breakpoint. At the same time, the rationality of the derivation is verified using a natural language reasoning model (e.g., RoBERTa-large): inputting "Can insulin treatment be deduced from diabetes symptoms?", the model outputs a confidence score (e.g., 0.85). If the score is lower than the threshold (e.g., 0.7), it is judged as a logical contradiction. Finally, a verification result report is generated, marking the location and type of the problem (e.g., "derivation basis for step 3 is missing").

[0044] In step S1153, multiple independent knowledge units are selected from the initial knowledge chain structure as verification starting points, and the selection range covers independent knowledge units of different domain classifications and different levels. In this embodiment, firstly, independent knowledge units are grouped according to domain classification labels (such as ICD-10 medical codes, financial industry classification codes), for example, the medical domain is divided into three categories: "symptoms," "diagnosis," and "treatment." Secondly, within each group, stratified sampling is performed according to knowledge levels (basic concepts, intermediate rules, advanced applications): high-frequency entities (such as "diabetes") are selected at the basic concept level, typical relationships (such as "hypertension leads to heart disease") are selected at the intermediate rule level, and complex scenarios (such as "multi-drug combination therapy") are selected at the advanced application level. Finally, it is ensured that the number of validation starting points meets the statistical significance requirement (such as selecting at least 5 units per group) and covers key nodes in the initial knowledge chain (such as path starting points and branch points). For example, in the medical domain, "diabetes symptoms" (basic), "insulin dosage calculation" (intermediate), and "diabetes complication management" (advanced) are selected as validation starting points.

[0045] In step S1154, for the independent knowledge unit that serves as the starting point of the verification, the sequence of upstream knowledge units along all the traceability paths is determined, wherein the sequence of upstream knowledge units is arranged from near to far in the tracing direction. In this embodiment, the verification starting point (e.g., "insulin therapy") is used as the root node. All incoming edges (i.e., associated paths pointing to this node) are traversed in the initial knowledge chain structure. For example, two paths are found: "dosage adjustment → insulin therapy" and "blood glucose monitoring → dosage adjustment". Then, the upstream nodes of each path are recursively traversed until the basic data node (e.g., "diabetes symptoms") is reached, generating the sequence "insulin therapy ← dosage adjustment ← blood glucose monitoring ← diabetes symptoms". Simultaneously, redundancy is eliminated from the sequence: if a node appears repeatedly in multiple paths (e.g., "blood glucose monitoring" appears in two paths simultaneously), only the first occurrence is retained. Finally, the sequence is sorted according to the tracing direction (from the verification starting point to the basic data), forming a sequence of upstream knowledge units from near to far. For example, a sequence length of 4 indicates that 3 steps of derivation are required to reach the basic data.

[0046] In step S1155, the upstream knowledge unit sequence is reversed, and the core viewpoints and derivation basis of each upstream knowledge unit are extracted in turn and compared with the derivation basis and core viewpoints of the next level knowledge unit. In this embodiment, firstly, starting from the end of the upstream knowledge unit sequence (e.g., "diabetes symptoms"), its core viewpoints (e.g., "excessive thirst, excessive urination") and derivation basis (e.g., "patient self-reported symptoms") are extracted. Then, the process moves forward to the next level unit (e.g., "blood glucose monitoring"), extracting its core viewpoints (e.g., "fasting blood glucose ≥ 7.0 mmol / L") and derivation basis (e.g., "laboratory test results"). Next, the semantic similarity between the two levels of units is calculated: the core viewpoints are encoded into vectors using the Sentence-BERT model, and the cosine similarity is calculated (e.g., the similarity between "excessive thirst, excessive urination" and "elevated blood glucose" is 0.6). If the similarity is lower than a threshold (e.g., 0.5), it is determined to be a semantic break. Simultaneously, the logical relationship of the derivation basis is checked: if the basis for "blood glucose monitoring" is "laboratory test," while the basis for "diabetes symptoms" is "patient self-report," and the two are not directly related, it is marked as a logical contradiction. Finally, a comparison report is generated, marking the location of the problem (e.g., "mismatch between derivation basis from level 2 to level 3").

[0047] In step S1156, the comparison results are used to characterize the locations where semantic breaks or logical contradictions occur. By supplementing and extracting feature elements and reconstructing logical relationships, the target tracing and association path is obtained. In this embodiment, for problematic locations in the comparison results (such as semantic breaks or logical contradictions), the missing feature elements are first supplemented: if the association between "blood glucose monitoring" and "diabetes symptoms" is missing, then "hyperglycemia is a core indicator of diabetes" is extracted from the domain knowledge base as a supplementary feature. Secondly, the logical relationship is reconstructed: the original path "diabetes symptoms → blood glucose monitoring → dosage adjustment → insulin therapy" is modified to "diabetes symptoms (core indicator: hyperglycemia) → blood glucose monitoring (basis: laboratory test) → dosage adjustment (rule: based on blood glucose value) → insulin therapy (goal: blood glucose control)", and the type of derivation basis is marked at each step (such as "core indicator", "laboratory test", "rule"). Finally, a logic verification model (such as GPT-4) is used to verify the rationality of the reconstructed path: when the input is "Is the derivation from diabetes symptoms to insulin therapy reasonable?", and the model outputs "reasonable" with a confidence level ≥ 0.9, it is determined to be a target tracing association path.

[0048] In step S1157, each path in the target tracing and association path is extracted from the starting point to the ending point in turn to determine the knowledge chain set of each domain category.

[0049] In this embodiment, firstly, the target tracing paths are grouped according to domain classification labels (such as medical, financial). For example, the medical domain includes "diabetes treatment chain" and "hypertension management chain". Secondly, each group of paths is clustered: the overlap of feature elements between paths is calculated (e.g., "diabetes treatment chain" and "insulin treatment chain" share the "insulin" feature). If the overlap is ≥0.7, they are merged into the same knowledge chain set. Then, a unique identifier (e.g., "MED-001" represents the first chain in the medical domain) and metadata (e.g., the number of independent knowledge units and the average path length) are generated for each set. Finally, the sets are encapsulated into structured data (e.g., JSON format), including fields such as "domain classification", "knowledge chain ID", "list of independent units", and "list of associated paths", to generate a knowledge chain set.

[0050] In a preferred embodiment, the support type includes multi-source support type and single-source support type, see [link to relevant documentation]. Figure 3 As shown, in step S12, the support relationship determination criteria in the knowledge chain methodology are invoked to perform a traceability support operation on each knowledge unit in the knowledge chain set, thereby obtaining the support type corresponding to the knowledge unit, including: In step S121, the support relationship determination criteria are parsed, and the core reference items for determining the support type are extracted. The core reference items include the number of traceable related paths, the matching degree of basic data, and the completeness of case support. Each reference item has a clear determination basis. Among them, the core reference items are key indicators for determining the support type, including the number of traceable related paths, the matching degree of basic data, and the completeness of case support, which are used to quantitatively evaluate the reliability of the support source of knowledge units.

[0051] In this embodiment of the disclosure, firstly, structured rules are extracted from the support relationship determination standard document of the knowledge chain methodology, such as "if the number of traceable related paths is ≥3 and the basic data matching degree is ≥0.8, then it is determined to be multi-source support".

[0052] Secondly, the rules are broken down into quantifiable core reference items: Number of upstream related paths: The total number of upstream related paths of the target knowledge unit is counted, reflecting the diversity of supporting sources; Basic data matching degree: The overlap rate of feature elements between upstream knowledge units and target knowledge units is calculated to measure data consistency; Case support completeness: The coverage of related cases to the target knowledge unit is evaluated, reflecting the sufficiency of practical verification.

[0053] Finally, define the criteria for each reference item: for example, the number of paths is set to "≥3" as the multi-source threshold, the matching degree is set to "≥0.8" as the high confidence threshold, and the case completeness is set to "covering more than 80% of the feature elements" as the sufficient support standard. The criteria are then encoded into executable logical conditions through a rule engine (such as Drools).

[0054] In step S122, each independent knowledge unit in the knowledge chain set is traversed, the current independent knowledge unit is taken as the target knowledge unit, all traceability paths of the target knowledge unit are extracted, and the upstream knowledge unit pointed to by each path is determined. In this embodiment, graph traversal and node localization techniques are employed. First, the knowledge chain set is modeled as a directed graph structure, where nodes are independent knowledge units and edges are tracing paths (e.g., "diabetes treatment → insulin dosage adjustment").

[0055] Then, traverse each node in the graph: Node localization: Use the depth-first search (DFS) algorithm to locate the current target knowledge unit (such as "insulin dose adjustment"); Path extraction: Starting from the target node, traverse backwards along the incoming edges (edges pointing to the node) to collect all upstream knowledge units (such as "blood glucose monitoring" and "diabetes diagnosis") and their corresponding associated paths; Path storage: Encapsulate the path information into structured data (such as JSON format), including fields such as "path ID", "start point", "end point", and "associated feature elements".

[0056] In step S123, the number of traceability paths of the target knowledge unit is counted. The count of the number of paths is limited to the valid paths that have been verified, and invalid paths that are not logically coherent are excluded. Among them, the effective path is the traceability path that has been confirmed to have no semantic breaks or contradictions after logical verification.

[0057] In this embodiment, path filtering and counting techniques are employed. First, valid paths are filtered from the path list: Logical verification marking: The generated verification result report is called, and the "logical coherence" field (e.g., "passed" or "broken") of each path is extracted; Invalid path exclusion: Paths marked "broken" or "contradictory" are filtered out, retaining only "passed" paths; Path counting: The number of remaining valid paths is counted. For example, the target knowledge unit "insulin dose adjustment" has 3 valid paths ("blood glucose monitoring → dose adjustment", "symptom assessment → dose adjustment", "complication risk → dose adjustment"). Finally, the number of paths is stored as an attribute of the target knowledge unit.

[0058] In step S124, the feature elements of the upstream knowledge unit of each traceability path are checked, and the basic data matching degree between the upstream knowledge unit and the target knowledge unit is calculated. Among them, the basic data matching degree is the degree of overlap between upstream knowledge units and target knowledge units at the feature element level, reflecting the consistency and reliability of data sources.

[0059] In this embodiment of the disclosure, feature vector similarity calculation technology is used. First, feature elements (such as entities, attributes, rules) of the target knowledge unit and the upstream knowledge units of each effective path are extracted. For example, the feature elements of the target unit "insulin dose adjustment" are: ["dose rule", "blood glucose threshold", "patient age"]; the feature elements of the upstream unit "blood glucose monitoring" are: ["blood glucose value", "detection time", "patient ID"].

[0060] Then, calculate the intersection and union of the two sets of feature elements: intersection: ["blood glucose value"] (overlapping elements); union: ["dosage rule", "blood glucose threshold", "patient age", "test time", "patient ID"] (all elements).

[0061] Finally, calculate the matching score. If weighted calculation is required (e.g., key elements have higher weights), TF-IDF or attention mechanisms can be used to adjust the element importance. The final average matching score of all upstream units is taken as the base data matching score of the target unit.

[0062] In step S125, the associated case index is extracted from the feature elements of the target knowledge unit, the corresponding case knowledge unit is retrieved, and the completeness of the support of the case knowledge unit for the target knowledge unit is determined. In this embodiment, case retrieval and coverage analysis techniques are employed. First, the associated case index (e.g., "Case ID-001") is extracted from the feature elements of the target knowledge unit, and the corresponding case unit (e.g., "Patient A's insulin treatment record") is retrieved from the case knowledge base. Then, the coverage of the case to the target feature elements is analyzed: Element mapping: The entities / attributes in the case are aligned with the target feature elements (e.g., "blood glucose value = 8.0 mmol / L" in the case corresponds to the target element "blood glucose threshold"); Coverage calculation: The ratio of the number of target elements covered by the case to the total number of elements is calculated. For example, if there are 5 target elements ("dosage rule", "blood glucose threshold", "patient age", "complication risk", "adjustment frequency"), and the case covers 3 ("blood glucose threshold", "patient age", "adjustment frequency"), then the coverage is 3 / 5 = 0.6; Completeness grading: The case support completeness is determined according to the coverage threshold (e.g., ≥0.8 is "complete", 0.5-0.8 is "partially complete", <0.5 is "incomplete").

[0063] Finally, the highest coverage of all related cases is taken as the case support completeness of the target unit. For example, if the coverage of 3 cases is 0.6, 0.7 and 0.9 respectively, then the completeness is 0.9.

[0064] In step S126, the number of traceable related paths, the matching degree of basic data, and the completeness of case support are compared with the judgment criteria in the support relationship judgment criteria to determine whether the knowledge unit corresponds to the multi-source support type or the single-source support type.

[0065] In this embodiment, a multi-condition decision tree technique is employed. First, the calculated indicators (number of paths, matching degree, and completeness) are compared with the thresholds in the support relationship determination criteria: Multi-source support condition: number of paths ≥ 3, matching degree ≥ 0.8, and completeness ≥ 0.8; Single-source support condition: if the multi-source condition is not met, and the number of paths ≥ 1, matching degree ≥ 0.5, and completeness ≥ 0.5, then it is determined to be single-source support; No support condition: if the number of paths = 0 or the matching degree / completeness is lower than the threshold, then it is marked as "no support" (further verification is required).

[0066] Then, a decision tree model (such as scikit-learn's DecisionTreeClassifier) ​​is used to automatically determine the support type: input the indicator value (e.g., [3, 0.47, 0.9]) and output the support type (e.g., "multi-source support"). Finally, the determination result is written into the metadata of the knowledge unit (e.g., support_type="multi-source support"), completing the support type labeling.

[0067] In a preferred embodiment, in step S125, the extraction of associated case indexes from the feature elements of the target knowledge unit, retrieval of corresponding case knowledge units, and determination of the completeness of the support of the case knowledge units for the target knowledge unit include: In step S1251, the associated case index is extracted from the feature elements of the target knowledge unit. The associated case index includes the identification information of the case knowledge unit and the description of the associated dimension. The description of the associated dimension indicates the association direction between the case and the target knowledge unit. In this embodiment, structured data parsing technology is employed. First, predefined related case fields (such as `related_cases`) are extracted from the metadata of the target knowledge unit. These fields are typically stored in JSON or XML format. Then, each case entry is parsed to extract `case_id` (unique identifier) ​​and `dimension` (related dimension description). The related dimension description must adhere to a predefined classification system (such as "application scenario verification," "logical deduction verification," and "data corroboration") to ensure that the correlation between the case and the target unit can be quantified and analyzed. Finally, a related case index list is generated.

[0068] In step S1252, based on the identification information in the associated case index, the corresponding case knowledge unit is retrieved in the knowledge chain set to obtain the complete feature elements of the case knowledge unit; In this embodiment, knowledge graph query and data fusion technology is employed. First, based on the extracted case_id, node queries are performed in the knowledge chain set (usually stored in a graph database such as Neo4j), ultimately generating a set of feature elements for the case knowledge unit.

[0069] In step S1253, the application scenario description and core viewpoint of the target knowledge unit are extracted, and the case background of the case knowledge unit is semantically compared with the application scenario description of the target knowledge unit to obtain the scenario fit parameter. In this embodiment, semantic similarity calculation technology from Natural Language Processing (NLP) is employed. First, the application scenario description (e.g., "suitable for dose adjustment for patients with large blood glucose fluctuations in type 2 diabetes") and case background (e.g., "Patient A, 50 years old, type 2 diabetes, large blood glucose fluctuations") of the target knowledge unit are extracted. These are then converted into vector representations using a pre-trained language model (e.g., BERT). Next, the cosine similarity between the two vectors is calculated. Finally, the similarity is normalized to the [0,1] interval (e.g., (similarity + 1) / 2) to generate a scenario fit parameter. If the similarity is ≥0.8, it is considered "highly fitting"; 0.5-0.8 is "moderately fitting"; and <0.5 is "lowly fitting".

[0070] In step S1254, the conclusions and viewpoints of the case knowledge unit are correlated with the core viewpoints of the target knowledge unit to determine the logical relationship between the two. In this embodiment, logical reasoning and rule matching techniques are employed. First, the core viewpoints of the target knowledge unit (such as "dynamic dose adjustment can improve the blood glucose control target achievement rate") and case conclusions (such as "after adopting the dynamic dose adjustment scheme, the target achievement rate increases to 85%) are extracted and converted into logical propositions: Target proposition: Dynamic adjustment → Improved target achievement rate; Case proposition: Dynamic adjustment ∧ Target achievement rate = 85%.

[0071] Then, the rule engine matches logical relationships: Direct support: The case proposition contains all the premises of the target proposition and the conclusion is consistent (e.g., in the case, "dynamic adjustment" leads to "increased achievement rate"); Indirect support: The case proposition partially supports the target proposition (e.g., the case only verifies the feasibility of "dynamic adjustment" and does not mention the achievement rate); Contradiction: The case proposition negates the target proposition (e.g., the case shows that "the achievement rate decreased after dynamic adjustment"); Irrelevant: The case has no logical connection with the target proposition (e.g., the case discusses "dietary intervention" rather than "dosage adjustment").

[0072] Finally, generate logical relation type labels (such as logic_relation="direct support").

[0073] In step S1255, it is checked whether the result data of the case knowledge unit can corroborate the derivation basis of the target knowledge unit. The data corroboration judgment is based on the correlation between the result data and the derivation basis and the matching degree of data types. In this embodiment, data association analysis and type matching techniques are employed. First, the derivation basis of the target knowledge unit (e.g., "If blood glucose level > 7.0 mmol / L, then increase the dose by 2U") and case result data (e.g., {"metric": "blood glucose level", "value": "7.2 mmol / L", "action": "increase the dose by 2U"}) are extracted. Then, a two-step verification is performed: Relevance verification: Check whether the case data directly relates to the key indicators in the derivation basis (such as blood glucose level, dosage adjustment rules). For example, the "7.2 mmol / L" in the case is directly related to the ">7.0 mmol / L" in the derivation basis. Type matching verification: Confirm that the data type of the case (such as continuous value, category label) is consistent with the requirements of the derivation basis. For example, if the derivation basis requires "blood glucose value (continuous value)" and the case provides "7.2 mmol / L (continuous value)", then the match is successful.

[0074] If both steps pass, the data is deemed "valid"; otherwise, it is deemed "invalid". The final data supporting result is then generated.

[0075] In step S1256, the case support evaluation results, which include detailed descriptions from multiple dimensions, are generated by taking into account the scenario fit parameters, logical relationship types, and data corroboration. A multi-dimensional scoring and weighted aggregation technique is employed. First, weights are assigned to each dimension (e.g., scenario fit 0.4, logical relationship 0.3, data corroboration 0.3). Then, weighted scores are calculated based on the results of steps S1253-S1255: scenario fit parameter (e.g., 0.9 → 90 points); logical relationship type mapping score (e.g., "direct support" → 100 points, "indirect support" → 70 points); data corroboration result mapping score (e.g., True → 100 points, False → 0 points). Finally, an evaluation result description is generated based on total_score: 92.

[0076] In step S1257, based on the case support evaluation results and referring to the requirements for case support completeness in the support relationship determination criteria, the degree of support completeness of the case knowledge unit for the target knowledge unit is determined.

[0077] In this embodiment, a threshold determination and level classification technique is adopted. First, the total_score is compared with the threshold in the support relationship determination standard: complete support: total_score≥90; partial support: 70≤score<90; weak support: 50≤score<70; no support: score<50.

[0078] Then, based on the judgment results, a support completeness label (e.g., support_level="Complete Support") is generated, and the judgment criteria are recorded (e.g., "Scenario fit 0.9, direct logical support, valid data evidence"). Finally, the results are written into the metadata of the target knowledge unit (e.g., case_support_level="Complete Support"), completing the case support completeness judgment.

[0079] In a preferred embodiment, the determination criteria for the support relationship include at least one of the following: the number of traceable related paths reaches a preset number, the basic data matching degree reaches a preset matching degree, and the case support completeness reaches a preset completeness.

[0080] In a preferred embodiment, see Figure 4 As shown, in step S13, the step of associating the knowledge chain set with the corresponding support type input into the artificial intelligence training model for training, and generating knowledge output filtering rules, includes: In step S131, the knowledge chain set and its corresponding support type are divided into training data group and verification data group according to the number of knowledge units and the proportion of support types by domain classification. In this embodiment, stratified sampling and data distribution analysis techniques are employed. First, knowledge units are divided into different domain subsets (e.g., medicine, engineering) based on their metadata (e.g., the domain field). Then, two levels of statistics are performed on each subset: quantity statistics: calculating the total number of knowledge units in each domain (e.g., the medical domain contains 1000 units); support type percentage statistics: calculating the percentage of different support types (e.g., "experimental verification," "theoretical derivation," "case evidence") in each domain (e.g., "experimental verification" accounts for 60% in the medical domain). Finally, based on the statistical results, training data (70%) and validation data (30%) are proportionally divided to ensure that the two sets of data are consistent with the original set in terms of domain distribution and support type percentage, avoiding data bias from affecting the model's generalization ability.

[0081] In step S132, the training data set is input into the artificial intelligence training model, the knowledge association weights of the artificial intelligence training model are initialized, and the association learning process of the artificial intelligence training model is started. The learning objective of the artificial intelligence training model is to establish a stable mapping relationship between knowledge unit features and support types. In this embodiment, neural network initialization and supervised learning techniques are employed. First, a suitable model architecture (such as a multilayer perceptron (MLP) or graph neural network (GNN) is selected for structured data association analysis, and the network weights are initialized (e.g., random initialization or Xavier initialization). Then, the training data set is input into the model, with knowledge unit features (such as application scenarios, core viewpoints, and data types) as input and supporting types (such as "experimental verification" and "theoretical derivation") as output, and the weights are adjusted through a backpropagation algorithm.

[0082] In step S133, multi-dimensional feature extraction is performed on the independent knowledge units in the training data group through an artificial intelligence training model to generate a feature-supported association model. In this embodiment, feature engineering and model fusion techniques are employed. First, multi-dimensional feature extraction is performed on independent knowledge units in the training data set: semantic features: semantic vectors of application scenario description and core viewpoints are extracted using a pre-trained language model (such as BERT); structural features: metadata of the knowledge unit (such as the number of associated cases and data type distribution) is parsed to generate structured feature vectors; statistical features: the position of the knowledge unit in the knowledge chain (such as depth and centrality) is calculated as a network feature.

[0083] Then, the three types of features are concatenated into a comprehensive feature vector, which is then input into an association learning model (such as a random forest or a Transformer). The model captures the interaction relationships between features through an attention mechanism, generating a feature-supported association model. This model can output the probability distribution of knowledge units belonging to each support type (e.g., "experimental verification" probability 0.8, "theoretical derivation" probability 0.2).

[0084] In step S134, the verification data set is input into the feature support association model to obtain the output predicted support type. Based on the predicted support type and the actual support type in the verification data set, the prediction accuracy and type confusion rate are calculated. Cross-validation and confusion matrix analysis techniques are employed. First, the validation data set is input into the feature-supported association model to obtain the predicted support type for each knowledge unit (e.g., predicted as "experimental verification"). Then, the predicted type is compared with the actual type to calculate the prediction accuracy. For type confusion rate, a confusion matrix is ​​constructed, and the number of misclassifications for each support type is counted (e.g., the number of times "theoretical derivation" is misclassified as "case evidence"), and the confusion rate is calculated. Finally, the model performance is comprehensively evaluated using both accuracy and confusion rate to identify support types that need optimization (e.g., types with high confusion rates require enhanced feature discrimination).

[0085] In step S135, the feature weight parameters of the feature support association model are adjusted based on the prediction accuracy and the type confusion rate. In this embodiment, gradient descent and regularization techniques are employed. First, based on the prediction accuracy and confusion rate, the direction for optimization is determined (e.g., improving the accuracy of the "Experimental Validation" type). Then, the gradient of the loss function with respect to the weights of each feature is calculated, and through multiple rounds of adjustments, the model focuses more on features that are key to the prediction of the supporting type (e.g., the weight of the "Data Type" feature in the "Experimental Validation" type).

[0086] In step S136, the model training, validation and parameter adjustment steps are repeated until the prediction accuracy of the feature support association model reaches a stable state. In this embodiment, early stopping and convergence determination techniques are employed. First, a maximum number of training epochs (e.g., 100 epochs) and a convergence threshold (e.g., accuracy fluctuation <1%) are set. Then, after each training epoch, the validation set accuracy is recorded, and the difference from the previous epoch is calculated: ΔAccuracy = |Accuracy t -Accuracy t-1 If ΔAccuracy < θ (e.g., θ = 0.01) and this condition persists for n rounds (e.g., n = 5), then the model is considered to have reached a stable state, and training is stopped. Otherwise, steps S132-S135 are executed until the convergence condition is met. This process ensures that the model fully learns the correlation between knowledge unit features and support types while avoiding overfitting.

[0087] In step S137, based on the stabilized feature-supported association model and combined with the output priority rules in the knowledge chain methodology, the core screening conditions for knowledge output are extracted. In this embodiment, a rule engine and priority mapping technology are employed. First, the probability distribution of support type for each knowledge unit is obtained from the stabilized feature support association model (e.g., "experimental verification" 0.8, "theoretical derivation" 0.15). Then, according to the output priority rules (e.g., priority: experimental verification > theoretical derivation > case evidence), probability thresholds are set (e.g., "experimental verification" must be ≥0.7, "theoretical derivation" must be ≥0.5).

[0088] Finally, the support types that meet the threshold and have the highest priority are extracted as the core screening criteria (e.g., knowledge units must meet "experimental verification probability ≥ 0.7" to be output), ensuring that the output results meet the rigor requirements of the knowledge chain methodology.

[0089] In step S138, the core screening conditions are converted into rule statements that the model can execute. The rule statements clearly specify the output order of different support types and the minimum standard of feature matching degree, thereby generating the knowledge output screening rules.

[0090] In this embodiment, a domain-specific language (DSL) design and code generation technique is employed. First, the core selection criteria are broken down into logical atoms (e.g., "experimental verification probability ≥ 0.7" is condition 1, and "theoretical derivation probability < 0.3" is condition 2). Then, atomic conditions are combined according to output priority to generate rule statements (e.g., "IF condition 1 AND NOT condition 2 THEN output is experimental verification"). Finally, the rule statements are converted into executable code for the model (e.g., Python's if-else statement or SQL's WHERE clause) to generate knowledge output selection rules.

[0091] In a preferred embodiment, step S133, which involves extracting multi-dimensional features from independent knowledge units in the training data set using an artificial intelligence training model to generate a feature-supported association model, includes: In step S1331, the content of independent knowledge units is segmented using an artificial intelligence training model based on a domain-specific vocabulary database to obtain multiple semantic vocabulary units. In this embodiment, a fusion technique of dictionary-based rule-based word segmentation and statistical word segmentation is employed. First, a predefined domain-specific vocabulary (e.g., a medical vocabulary containing terms like "gene editing" and "clinical trials") is loaded as the base dictionary for word segmentation rules. Then, forward maximum matching word segmentation is performed on the text content of independent knowledge units (e.g., application scenario descriptions, core viewpoints): starting from the beginning of the text, the longest word in the dictionary is matched character by character; if a match is successful, the text is segmented; otherwise, the next longest word is used. Simultaneously, a statistical model (e.g., a Hidden Markov Model, HMM) is used to process out-of-vocabulary words (e.g., newly emerging professional terms), optimizing the word segmentation boundaries by calculating the transition probabilities between characters. Finally, multiple semantic vocabulary units (e.g., "gene editing technology" and "clinical trial data") are generated.

[0092] In step S1332, each semantic word unit is transformed into a vector. The transformation process uses a pre-trained semantic vector model to map the semantic word unit into a word vector of fixed dimension. In this embodiment, word embedding technology is employed. First, a pre-trained model suitable for the domain text (such as BioBERT for the medical field) is selected, and its trained weight parameters are loaded. Then, the semantic vocabulary units generated in step S1331 are input into the model, and each word is converted into a fixed-dimensional vector (e.g., 300-dimensional) through the model's internal embedding layer. The conversion process utilizes the word co-occurrence relationships learned by the model during the pre-training stage: if two words frequently co-occur in the corpus (e.g., "gene" and "edit"), their vectors are close in the vector space (e.g., cosine similarity > 0.8). Finally, multiple word vectors are generated (e.g., "gene editing" → [0.2, -0.5, 0.7, ...]), providing quantized input for semantic aggregation.

[0093] In step S1333, based on the semantic correlation between word vectors, multiple word vectors are aggregated to generate semantic feature vectors that can represent the overall connotation of independent knowledge units. In this embodiment, a weighted average and attention mechanism fusion technique is employed. First, the cosine similarity between all pairwise word vectors is calculated to construct a similarity matrix (e.g., an n×n matrix, where n is the number of words). Then, weights are assigned to each word vector based on the similarity matrix: if word A and B have a high similarity (e.g., >0.7), then B's aggregation contribution weight to A is larger (e.g., 0.6). Next, a weighted average aggregation is performed on all word vectors.

[0094] To further capture key semantics, an attention mechanism (such as the self-attention layer of the Transformer) can be introduced to dynamically adjust weights to highlight important words (such as core domain terms). This ultimately generates a semantic feature vector (e.g., 300-dimensional) representing the overall meaning of the knowledge unit.

[0095] In step S1334, the position information of the independent knowledge unit in the knowledge chain set is extracted, and position feature parameters are generated. The position information includes the level of the independent knowledge unit in its domain classification and the number of connections with upstream and downstream independent knowledge units. In this embodiment, graph structure analysis and statistical techniques are employed. First, the knowledge chain set is modeled as a directed graph, where nodes are independent knowledge units and edges represent relationships (such as "reference" or "support"). Then, for each knowledge unit, the following calculations are performed: Level: Determine the depth of a node in the graph using breadth-first search (BFS) (e.g., the root node has a depth of 0, and the child nodes have a depth of 1). Connection count: The in-degree (how many upstream units it is associated with) and out-degree (how many downstream units it is associated with) of a node are counted and combined into the total connection count (e.g., in-degree + out-degree).

[0096] Finally, the number of levels and connections is normalized (e.g., scaled to the [0,1] interval) to generate positional feature parameters (e.g., [0.2, 0.8] represents 20% depth and 80% number of connections), quantifying the structural importance of knowledge units in the knowledge chain.

[0097] In step S1335, the traceability and association path information of the independent knowledge unit is statistically analyzed, and an association feature vector is generated. The traceability and association path information includes the number of paths, the path length, and the association dimension type of each path. In this embodiment, path search and feature statistics techniques are employed. First, starting from the target knowledge unit, a depth-first search (DFS) or breadth-first search (BFS) is performed to traverse all possible reverse paths (e.g., "A→B→C" indicates that A depends on B, and B depends on C). Then, the following statistics are calculated: number of paths: total number of paths (e.g., 3 paths); path length: number of nodes in each path (e.g., path 1 has a length of 2); association dimension type: type of association in each path (e.g., path 1 includes "theoretical derivation" and "experimental verification").

[0098] Finally, the number of paths, average path length, and the proportion of each dimension type (e.g., "theoretical derivation" accounts for 60%) are encoded into vectors (e.g., [3, 2.5, 0.6, 0.4]) to generate associated feature vectors and quantify the traceability dependencies of knowledge units.

[0099] In step S1336, vectors are concatenated according to preset weights to fuse the semantic feature vector, the positional feature parameter, and the associated feature vector to generate a comprehensive feature of knowledge units that includes multi-dimensional information. In this embodiment, a weighted concatenation and dimensionality reduction technique is employed. First, weights are assigned to the three types of features (e.g., semantic features 0.6, location features 0.2, and association features 0.2) to reflect their contribution to the prediction of the support type. Then, the vectors are concatenated according to their weights: semantic feature vector (300 dimensions) × 0.6 → 180 dimensions; location feature parameters (2 dimensions) × 0.2 → 0.4 dimensions (actually retaining 2 dimensions); association feature vector (4 dimensions) × 0.2 → 0.8 dimensions (actually retaining 4 dimensions).

[0100] After concatenation, a comprehensive feature is generated (e.g., 180+2+4=186 dimensions). To reduce computational complexity, dimensionality can be further reduced (e.g., to 100 dimensions) using principal component analysis (PCA) or linear discriminant analysis (LDA) while retaining key feature information.

[0101] In step S1337, a corresponding independent knowledge unit identifier and domain classification label are added to the comprehensive features of each independent knowledge unit to generate target multidimensional features; In this embodiment, structured data encapsulation technology is employed. First, a unique identifier (such as a UUID) is added to the comprehensive features of each knowledge unit to ensure a one-to-one correspondence between features and units. Then, based on the domain classification results of step S131, domain labels (such as "medicine") are attached to the features. Finally, the identifier, labels, and comprehensive features are encapsulated into structured data (such as JSON format). The structured data provides labeled input to the feature-supported association model, supporting supervised learning.

[0102] In step S1338, the target multidimensional features are associated with the corresponding support types to generate the feature support association model.

[0103] In this embodiment, supervised learning and probabilistic modeling techniques are employed. First, a suitable model for structured data classification (such as random forest, XGBoost, or neural network) is selected. Then, the target multidimensional features are used as input, and the supporting types (such as "experimental verification" and "theoretical derivation") are used as output to train the model to learn the mapping relationship between features and types. During training, the model adjusts its parameters by optimizing the loss function (such as cross-entropy loss) to make the predicted probability distribution close to the true label. Finally, a feature-support association model is generated, which can take the comprehensive features of a new knowledge unit as input and output the probability of it belonging to each supporting type (such as "experimental verification" probability 0.8, "theoretical derivation" probability 0.2).

[0104] In a preferred embodiment, step S14, which involves adapting and optimizing the knowledge output filtering rules using domain validation data to obtain a trained knowledge chain-based artificial intelligence model, includes: In step S141, the domain verification data is input into the artificial intelligence training model, the injected knowledge chain methodology is invoked, and the knowledge traceability triggering mechanism is activated. The domain verification data includes independent knowledge units that have not participated in training and their corresponding annotation support types. The annotation support types are formed through cross-comparison of multi-source data. In this embodiment, domain validation data (such as fragments of medical research papers not used in training) is first loaded and input into a pre-trained artificial intelligence model (such as a Transformer-based text classification model). Logical rules of the knowledge chain methodology (such as "if the conclusion depends on clinical trial data, then the support type is experimental validation") are injected into the model, and a knowledge tracing trigger mechanism is initiated: for each knowledge unit, the model extracts key entities (such as "clinical trial" and "sample size") and relationships (such as "A depends on B") from its text content, constructing a local knowledge graph. If a path related to the support type determination exists in the graph (such as "conclusion → experimental data → statistical method"), subsequent support tracing operations are triggered to ensure that the determination process is based on complete knowledge chain dependencies.

[0105] In step S142, based on the knowledge traceability triggering mechanism, a traceability support operation is performed on each independent knowledge unit in the domain verification data, the support type of the independent knowledge unit is determined according to the support relationship determination criteria, and a model is generated to determine the support type. In this embodiment, a path search and rule matching technique based on a knowledge graph is employed. First, the model extracts all incoming edges (i.e., dependent upstream nodes) from the local knowledge graph of the knowledge unit. For example, if the knowledge unit is "Drug A is effective," its upstream nodes might include "Clinical Trial B" and "Molecular Mechanism C." Then, the model assigns a support type to each upstream node based on support relationship determination criteria (such as a predefined rule base: "Clinical Trial → Experimental Verification" and "Molecular Mechanism → Theoretical Derivation"). If the same knowledge unit is supported by multiple upstream nodes (e.g., simultaneously dependent on experiments and theories), the model uses a weighted voting mechanism: weights are calculated based on the upstream node's authority (e.g., journal impact factor) or association strength (e.g., text co-occurrence frequency), ultimately generating the model-determined support type (e.g., "Experimental Verification (weight 0.7) + Theoretical Derivation (weight 0.3)").

[0106] In step S143, the knowledge output filtering rules are invoked to filter and sort the independent knowledge units after the model determines the support type, and a verification output result including a list of independent knowledge units sorted by priority and the corresponding support type description is generated. In this embodiment, a rule engine and ranking algorithm are integrated. First, the model loads predefined knowledge output filtering rules (e.g., prioritizing experimental verification knowledge in the medical field), using the model-determined support type generated in step S142 as input. Then, the rule engine calculates a priority score for each knowledge unit based on priority logic in the rule base (e.g., "if the support type is experimental verification, then priority +2; theoretical derivation +1"). Next, the model performs a ranking operation: arranging knowledge units in descending order of score to generate a priority ranking list (e.g., ["Experimental verification: Drug A is effective", "Theoretical derivation: Molecular mechanism C"]). Simultaneously, the model adds a support type description (e.g., "dependent on clinical trial B") to each unit, forming a structured verification output result.

[0107] In step S144, the verification output result is compared with the labeled support type and the expected output order to generate deviation reasons including support type misjudgment and / or priority order not in accordance with expectations; In this embodiment, comparative analysis and attribution reasoning techniques are employed. First, the model compares the verification output (such as a priority ranking list) with the labeled support types and expected output order (such as expert ranking) item by item, and counts two types of deviations: misjudgment of support type: the model's type determination is inconsistent with the label (e.g., the model judges it as "theoretical derivation" while the label is "experimental verification"); priority ranking does not meet expectations: the difference between the model ranking and the expert ranking exceeds a threshold (e.g., the model ranks low-priority units at the top).

[0108] Then, the model uses attribution reasoning to pinpoint the cause of the bias: if the support type is misjudged, it analyzes whether it is due to insufficient feature extraction (e.g., failure to capture key experimental terms) or flaws in the rule logic (e.g., failure to cover newly emerging support relationships); if the priority ranking is biased, it checks whether the rule weights are reasonable (e.g., whether there is an overemphasis on theoretical derivation). Finally, a bias cause report is generated (e.g., "Misjudgment due to failure to identify the keyword 'randomized controlled trial'").

[0109] In step S145, the parameters of the feature support association model are adjusted to address the misjudgment of the support type, the dimensional weights of feature extraction and the application logic of the support relationship judgment criteria are optimized, and / or the priority weights in the knowledge output filtering rules are adjusted in combination with domain characteristics for priority sorting. In this embodiment, parameter optimization and dynamic rule adjustment techniques are employed. To address misjudgments of support types, the model adjusts the parameters of the feature support association model through grid search or Bayesian optimization. For example, if a misjudgment is due to excessively low semantic feature weights, their weights are increased (e.g., from 0.5 to 0.7), and the application logic of the support relationship determination criteria is optimized (e.g., the cosine similarity threshold is reduced from 0.75 to 0.7 to capture more weak associations). To address priority ranking biases, the model adjusts the priority weights of the knowledge output filtering rules based on domain characteristics (e.g., the medical field places greater emphasis on experimental evidence). For example, the weight of experimental verification is increased from 0.6 to 0.8, and theoretical derivation is reduced from 0.3 to 0.2, ensuring that the ranking results meet domain requirements.

[0110] In step S146, the steps of domain verification data input, model processing, deviation analysis and rule adjustment are repeated until the degree of agreement between the verification output and the expected output meets the domain application standards, and a trained knowledge chain-based artificial intelligence model is obtained.

[0111] In this embodiment, iterative optimization and convergence determination techniques are employed. The model repeatedly executes steps S141-S145: each iteration inputs new domain validation data, generates validation output results, compares with annotations, and adjusts parameters / rules until the fit meets the domain application standards. Specifically, after each iteration, the model calculates the current accuracy (e.g., support type matching rate increases from 85% to 92%) and ranking correlation coefficient (e.g., Kendall's Tau increases from 0.7 to 0.85). If the fluctuation of the indicators is less than a threshold (e.g., ±0.01) for n consecutive iterations (e.g., n=3), the model is determined to have converged. Finally, the model outputs a trained knowledge chain-based artificial intelligence model. This model meets the domain requirements in both support type determination and priority ranking, and can be directly deployed in practical application scenarios (e.g., medical knowledge base construction).

[0112] This invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the methods described in the foregoing embodiments.

[0113] This invention provides a question-answering system based on a knowledge chain-trained artificial intelligence model, comprising: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of any of the methods described in the foregoing embodiments.

[0114] Figure 5The question-answering system 100 based on a knowledge chain-trained artificial intelligence model, as shown, includes a processor 1001 and a memory 1003. The processor 1001 and memory 1003 are connected, for example, via a bus 1002. Optionally, the question-answering system 100 may further include a communication component 1004, which can be used for data interaction between the device 100 and other devices, such as sending and / or receiving data. It should be noted that in actual scheduling, the communication component 1004 is not limited to one, and the structure of this question-answering system 100 based on a knowledge chain-trained artificial intelligence model does not constitute a limitation on the embodiments of this application.

[0115] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0116] Bus 1002 may include a pathway for transmitting information between the aforementioned components. Bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 1002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0117] The memory 1003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium capable of carrying or storing program code and capable of being read by a computer, without limitation herein.

[0118] The memory 1003 is used to store program code for executing embodiments of the present disclosure, and its execution is controlled by the processor 1001. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing embodiments of the question-answering method for training an artificial intelligence model based on a knowledge chain.

[0119] The preferred embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings. However, the present disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present disclosure, various changes, modifications, substitutions and variations can be made to these embodiments, and all such changes, modifications, substitutions and variations fall within the protection scope of the present disclosure.

[0120] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction, and such combinations should also be considered as part of this disclosure. To avoid unnecessary repetition, this disclosure will not further describe the various possible combinations. The technical scope of this application is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A question-answering method based on training an artificial intelligence model using knowledge chains, characterized in that, The method comprises the following steps: Based on the knowledge chain methodology for starting the answer to the hierarchical traceability process of the basic data, the knowledge traceability trigger mechanism organizes the field fact data to generate a knowledge chain set, which includes multiple knowledge units and traceability association paths between the knowledge units; Call the supporting relationship judgment standard in the knowledge chain methodology, and perform a traceability support operation on each knowledge unit in the knowledge chain set to obtain the corresponding support type of the knowledge unit; Input the knowledge chain set and the corresponding support type into an artificial intelligence training model for associated training to generate a knowledge output screening rule; Through domain verification data, the knowledge output screening rule is adapted and optimized to obtain a trained knowledge chain-based artificial intelligence model; In response to inputting a target question into the artificial intelligence model, the artificial intelligence model outputs a question and answer result for the target question. 2.The method of claim 1, wherein the knowledge chain comprises a plurality of knowledge nodes, and each of the knowledge nodes comprises a plurality of knowledge pieces. The knowledge traceability trigger mechanism based on the knowledge chain methodology for starting the answer to the hierarchical traceability process of the basic data, and the field fact data is used to generate a knowledge chain set, which comprises: Extract the core association dimension corresponding to the knowledge traceability trigger mechanism in multiple traceability directions; According to the independence of knowledge content and the matching degree of association dimension, the field fact data is split according to the core association dimension to obtain multiple independent knowledge units, each of which includes complete knowledge content, source identification and application scenario description; Each of the independent knowledge units is analyzed for content, and feature elements representing knowledge core attributes are extracted, wherein the extraction range of the feature elements covers all the contents of the independent knowledge units; Based on the feature elements, the derivation basis of any independent knowledge unit is taken as the starting point, and the corresponding core viewpoints or data source credentials in other independent knowledge units are matched to form a directional association and establish a traceability association path between the multiple independent knowledge units; According to the association dimension type, the matched feature elements and the path length of each traceability association path, the knowledge chain set of each field classification is determined, wherein the path length is determined according to the number of associated independent knowledge units; An index directory is established for each knowledge chain set, and a cross-field index directory is established for the knowledge chain set of each field classification according to the association dimension to generate the knowledge chain set, wherein the index directory includes independent knowledge unit identification, core feature element abstract and traceability association path starting node information. 3.The method of claim 2, wherein, According to the association dimension type, the matched feature elements and the path length of each traceability association path, the knowledge chain set of each field classification is determined, which comprises: According to the association dimension type, the matched feature elements and the path length of each traceability association path, an initial knowledge chain structure is formed; Starting from any independent knowledge unit, the traceability association path is reversely deduced to the basic data knowledge unit, and the logical coherence in the deduction process is checked to obtain the traceability path verification result of the initial knowledge chain structure; selecting multiple independent knowledge units from the initial knowledge chain structure as a verification starting point, the selection range covering independent knowledge units of different domain classifications and different levels; determining, for the independent knowledge unit as the verification starting point, an upstream knowledge unit sequence of all the traceable association paths thereof, wherein the upstream knowledge unit sequence is arranged in a traceable direction from near to far; deriving in reverse along the upstream knowledge unit sequence, and sequentially extracting a core viewpoint and a derivation basis of each upstream knowledge unit, and comparing the derivation basis and the core viewpoint with a derivation basis and a core viewpoint of a next-level knowledge unit; characterizing a position where a semantic break or a logical contradiction occurs according to a comparison result, and obtaining a target traceable association path through supplementary extraction of feature elements and reconstruction of logical relationships; sequentially extracting each path in the target traceable association path from a starting point to a terminal point, and determining a knowledge chain set of each domain classification. 4.The method of claim 1, wherein the knowledge chain is constructed based on a plurality of questions and answers, and the knowledge chain is constructed based on a plurality of questions and answers. The support types include a multi-source support type and a single-source support type, and the support relationship determination standard in the knowledge chain methodology is called to perform a traceable support operation on each knowledge unit in the knowledge chain set to obtain a support type corresponding to the knowledge unit, including: analyzing the support relationship determination standard to extract a core reference item for support type determination, the core reference item including a traceable association path quantity, a basic data matching degree, and a case support completeness; traversing each independent knowledge unit in the knowledge chain set, taking a current independent knowledge unit as a target knowledge unit, extracting all traceable association paths of the target knowledge unit, and determining an upstream knowledge unit pointed to by each path; counting the traceable association path quantity of the target knowledge unit, the path quantity counting range being valid paths that have passed verification, and excluding invalid paths that are not logically coherent; checking feature elements of the upstream knowledge unit of each traceable association path, and calculating a basic data matching degree between the upstream knowledge unit and the target knowledge unit; extracting an associated case index in the feature elements of the target knowledge unit, retrieving a corresponding case knowledge unit, and judging a support completeness of the case knowledge unit to the target knowledge unit; comparing the traceable association path quantity, the basic data matching degree, and the case support completeness with determination standards in the support relationship determination standard to determine that the knowledge unit is of the multi-source support type or the single-source support type. 5.The method of claim 4, wherein the knowledge chain-based training of the artificial intelligence model is performed by using a knowledge chain-based question and answer model. The extraction of the associated case index in the feature elements of the target knowledge unit, the retrieval of the corresponding case knowledge unit, and the judgment of the support completeness of the case knowledge unit to the target knowledge unit include: extracting the associated case index from the feature elements of the target knowledge unit, the associated case index including identification information and an associated dimension description of the case knowledge unit, and the associated dimension description indicating an association direction between the case and the target knowledge unit; based on the identification information in the associated case index, retrieving the corresponding case knowledge unit in the knowledge chain set to obtain complete feature elements of the case knowledge unit; extracting an application scenario description and a core viewpoint of the target knowledge unit, and performing semantic comparison between a case background of the case knowledge unit and the application scenario description of the target knowledge unit to obtain a scenario fit degree parameter; Correlation analysis is performed on the conclusion view of the case knowledge unit and the core view of the target knowledge unit to determine the logical relationship therebetween; It is checked whether the result data of the case knowledge unit can support the derivation basis of the target knowledge unit, and the data support judgment is based on the relevance of the result data and the derivation basis and the data type matching degree; The case support evaluation result including specific description of multiple dimensions is generated by comprehensively considering the scene fitting degree parameter, the logical relationship type and the data support condition; Based on the case support evaluation result, the support completeness of the case knowledge unit to the target knowledge unit is determined by referring to the requirements of the case support completeness in the support relationship judgment standard. 6.The method of claim 4, wherein the knowledge chain-based training of the artificial intelligence model is performed by using a knowledge chain-based question and answer model. The judgment standard in the support relationship judgment standard includes at least one of the following: The number of traceable correlation paths reaches a preset number, the basic data matching degree reaches a preset matching degree, and the case support completeness reaches a preset completeness. 7.The method of claim 1, wherein the knowledge chain comprises a plurality of knowledge nodes, and each of the knowledge nodes comprises a plurality of knowledge pieces. The knowledge chain set and the corresponding support type are input into an artificial intelligence training model for correlation training to generate a knowledge output screening rule, including: The knowledge chain set and the corresponding support type are classified according to the field to count the number of knowledge units and the proportion of support types, and are divided into a training data group and a verification data group; The training data group is input into the artificial intelligence training model to initialize the knowledge correlation weight of the artificial intelligence training model, start the correlation learning process of the artificial intelligence training model, and the learning goal of the artificial intelligence training model is to establish a stable mapping relationship between the knowledge unit features and the support type; The independent knowledge units in the training data group are subjected to multi-dimensional feature extraction by the artificial intelligence training model to generate a feature support correlation model; The verification data group is input into the feature support correlation model to obtain the predicted support type, and the prediction accuracy and type confusion rate are calculated according to the predicted support type and the actual support type in the verification data group; Based on the prediction accuracy and the type confusion rate, the feature weight parameters of the feature support correlation model are adjusted; The model training, verification and parameter adjustment steps are repeatedly executed until the prediction accuracy of the feature support correlation model reaches a stable state; Based on the stable feature support correlation model, the core screening condition of the knowledge output is extracted in combination with the output priority rule in the knowledge chain methodology; The core screening condition is converted into a rule statement executable by the model, and the rule statement clearly specifies the output order of different support types and the minimum standard of feature matching degree, thereby generating the knowledge output screening rule.

8. The question-answering method based on knowledge chain training of an artificial intelligence model according to claim 7, characterized in that, The independent knowledge units in the training data group are subjected to multi-dimensional feature extraction by the artificial intelligence training model to generate a feature support correlation model, including: The content of the independent knowledge unit is subjected to word segmentation processing based on the field professional vocabulary library by the artificial intelligence training model to obtain a plurality of semantic vocabulary units; Each semantic vocabulary unit is subjected to vector conversion, and the conversion process adopts a pre-trained semantic vector model to map the semantic vocabulary unit to a fixed-dimensional word vector. The semantic feature vector capable of representing the overall connotation of the independent knowledge unit is generated by aggregating a plurality of vocabulary vectors based on the semantic correlation between the vocabulary vectors; Position information of the independent knowledge unit in the knowledge chain set is extracted to generate position feature parameters, and the position information includes the level of the independent knowledge unit in the domain classification and the number of connections with upstream and downstream independent knowledge units; The trace association path information of the independent knowledge unit is counted to generate an association feature vector, and the trace association path information includes the number of paths, the length of the paths, and the association dimension type of each path; The semantic feature vector, the position feature parameter, and the association feature vector are fused by vector splicing according to a preset weight to generate a knowledge unit comprehensive feature including multi-dimensional information; The comprehensive feature of each independent knowledge unit is added with a corresponding independent knowledge unit identifier and a domain classification label to generate a target multi-dimensional feature; The target multi-dimensional feature is associated with the corresponding support type to generate the feature support association model. 9.The method of Claim 1-8, wherein, The knowledge output filtering rule is adapted and optimized by the domain verification data to obtain a trained artificial intelligence model based on the knowledge chain, including: The domain verification data is input into the artificial intelligence training model, the injected knowledge chain methodology is called, and the knowledge trace triggering mechanism is started, wherein the domain verification data includes independent knowledge units that do not participate in training and corresponding labeled support types, and the labeled support types are formed by cross comparison of multi-source data; Based on the knowledge trace triggering mechanism, the trace support operation is performed on each independent knowledge unit in the domain verification data, the support type of the independent knowledge unit is determined according to the support relationship determination standard, and the model determination support type is generated; The knowledge output filtering rule is called to filter and sort the independent knowledge units after the model determination support type, and a verification output result including an independent knowledge unit list sorted by priority and a corresponding support type description is generated; The verification output result is compared with the labeled support type and the expected output sequence to generate a deviation reason including support type misjudgment and / or priority sorting not meeting expectations; For the support type misjudgment, the parameters of the feature support association model are adjusted, the dimension weight of feature extraction and the application logic of the support relationship determination standard are optimized, and / or for the priority sorting, the priority weight in the knowledge output filtering rule is adjusted in combination with the domain characteristics; The steps of domain verification data input, model processing, deviation analysis, and rule adjustment are repeatedly performed until the consistency of the verification output result and the expected output meets the domain application standard to obtain the trained artificial intelligence model based on the knowledge chain. 10.A question and answer system for training an artificial intelligence model based on a knowledge chain, the question and answer system comprising: It includes: A memory having a computer program stored thereon; A processor configured to execute the computer program in the memory to implement the steps of the method of any one of claims 1-9.

Images