Mental health assessment field knowledge graph construction method and storage medium

Abstract

The invention discloses a psychological health assessment field knowledge graph construction method and a storage medium. The method comprises the following steps: S1, data preprocessing and structuring: processing an original document of a mental health assessment scale, and converting the original document into scale descriptive text and structured data; s2, entity extraction and seed entity set construction: extracting measurement factors and dimension entities in a single scale range, and constructing a standardized seed entity set; s3, relation triad extraction and deduplication: extracting structural and semantic triads under the constraint of a single scale, and performing semantic discrimination and deduplication by using LLM; and S4, multi-level map fusion and dynamic mapping are carried out, a global three-level semantic skeleton is constructed, entities are normalized and mapped to a standard concept, and problem access and map completion are realized. According to the method, the problems of cross-scale semantic confusion, inconsistent entity granularity and lack of hierarchical logic in the prior art are solved, and the specialty, accuracy and traceability of the knowledge graph are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, knowledge engineering, and mental health technology, specifically to a method for automated construction, entity alignment, and graph fusion of knowledge graphs in the mental health field based on Large Language Model (LLM). Background Technology

[0002] With the increasingly widespread application of artificial intelligence technology in the field of mental health, constructing high-quality mental health knowledge graphs has become a key foundation for realizing intelligent consultation, assisted diagnosis, and personalized intervention. However, existing mental health knowledge graph construction technologies face many challenges in practical applications. First, mental health scales exhibit significant heterogeneity, with their knowledge highly scattered across PDF documents and textbooks, making them difficult to process uniformly. Second, early construction methods mainly relied on human experts to define rules and ontology, which is costly and difficult to scale. While methods based on traditional supervised learning models have achieved automation to some extent, they heavily depend on large-scale labeled data and lack generalization ability when facing complex long-tail concepts and few-sample scenarios in the field of psychology. Existing GraphRAG technology is not only computationally expensive but also lacks control over entity granularity. Finally, directly using a general large language model (LLM) for triple extraction often produces illusions, resulting in generated entities and relations that do not conform to professional psychological standards, producing a large number of irrelevant entities or poorly granular nodes, affecting the usability of the graph.

[0003] In the process of cross-scale knowledge extraction and fusion, challenges such as semantic drift and contextual fragmentation exist. The same term can have completely different psychological meanings in different scales. For example, "Anxiety" refers to symptoms in the SCL-90 scale, while it refers to an emotional mechanism in the STAI scale. If the original theoretical framework of the items is removed, the boundaries of constructs will become blurred. For example, "Tension" refers to a personality dimension and a transient emotional state in the 16PF and STAI scales, respectively. If these are merged indiscriminately, the graph logic will be confused. At the same time, traditional semantic similarity calculation methods (such as Word2Vec and BERT) are difficult to distinguish between "synonymous relationships" and "correlated relationships" in psychology. For example, a general model may incorrectly merge "insomnia" and "sleep disorder" because they are highly similar, but they have different diagnostic meanings in clinical practice and have a hierarchical relationship. Such subtle semantic differences are blind spots for general embedding models and directly affect the accuracy of the graph.

[0004] Therefore, there is an urgent need for a knowledge graph construction method that can combine the powerful semantic understanding capabilities of large language models with strict process and hierarchical control to solve semantic drift and ensure professionalism and hierarchy. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing a knowledge graph in the field of mental health based on LLM and multi-level fusion. This method ensures the professionalism, accuracy and traceability of the graph through the strategy of "extraction within a single scale - semantic deduplication - multi-level mapping".

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for constructing a knowledge graph in the field of mental health assessment includes the following steps:

[0008] S1: Data preprocessing and structuring: Process the original documents of the mental health assessment scales and convert them into descriptive text and structured data.

[0009] S2: Entity extraction and seed entity set construction: Extract measurement factors and dimension entities within a single scale and construct a standardized seed entity set;

[0010] S3: Relational triple extraction and deduplication. Under single-scale constraints, structural and semantic triples are extracted, and semantic discrimination and deduplication are performed using LLM.

[0011] S4: Multi-level graph fusion and dynamic mapping constructs a global three-level semantic skeleton, normalizes entities and maps them to standard concepts, and enables the access of questions and graph completion.

[0012] Step S1 also includes obtaining the original document of the mental health assessment scale, performing text cleaning and format standardization, and separating it into two types of basic data: descriptive text and questionnaire assessment questions; using a large language model to convert the unstructured questionnaire questions into a structured data sequence containing question number, question stem, question source, options and associated factors.

[0013] Converting unstructured questionnaire questions into structured data sequences using a large language model includes: constructing prompts containing preset task instructions and output format requirements, driving the large language model to output the question number, question stem, question source, options, and associated factors in a structured JSON format.

[0014] Step S2 also includes using a seed entity generation strategy to extract measurement factors and dimension entities within a single scale; performing synonym merging, hierarchical processing, and standardized mapping on the extracted entities to construct a standardized seed entity set Q.

[0015] The seed entity generation strategy in step S2 includes:

[0016] S201: Single Scale Entity Extraction: For each scale's descriptive text, accurately extract the scale's explicitly defined measurement factors and assessment dimensions;

[0017] S202: Concept Fusion and Standardization: The extracted entities are merged with synonyms, processed hierarchically, and normalized for variants. The terminology is standardized according to standards such as DSM-5 to form a seed entity set Q.

[0018] Step S3 includes extracting structural and semantic relationships within the scale under the constraints of a single scale context; using a large language model to perform semantic discrimination based on logical reasoning; and combining this with a manual verification mechanism to eliminate semantic redundancy and retain independent relationships, thereby forming a high-quality set of triples.

[0019] In step S3, when using the large language model for semantic discrimination, a logical reasoning discrimination method based on Prompt constraints is adopted. The input is the triple pair to be compared, and the large language model is required to analyze whether their semantics are semantically equivalent, hierarchically contained, or mutually independent, and perform deduplication operation based on the judgment result.

[0020] Step S4 includes constructing a global three-level semantic skeleton containing a domain layer, a category layer, and a standard concept layer using the seed entity set Q; normalizing and mapping the underlying entities extracted from each single scale to the standard concepts of the global skeleton through a hierarchical mapping mechanism to resolve local semantic conflicts; and at the same time, integrating the questionnaire questions into the knowledge graph to construct a full-link mapping from questions to global concepts, thereby achieving global fusion and dynamic completion of the knowledge graph.

[0021] The dynamic access and hybrid completion strategy for questionnaire questions in step S4 is as follows: if the concept measured by the question already exists in the three-level skeleton, it is directly connected to the skeleton node; if the concept belongs to the fine-grained constructs specific to the scale and is not included in the skeleton, a new concept node is automatically created, and the large language model is used to dynamically map it to the upper-level node with the closest semantics.

[0022] A computer storage medium storing a computer program that, when executed by a processor, implements the method.

[0023] The advantages of this invention are as follows: By employing a strategy of first extracting from a single scale and then globally fusing the results, semantic stability is improved, effectively avoiding confusion between terms under different theoretical frameworks and preserving the original theoretical background of psychological concepts. The deep semantic understanding and logical reasoning capabilities of the Large Language Model (LLM) replace traditional BERT vector similarity calculations, improving data quality and consistency. This effectively distinguishes between synonyms and related words in psychology, avoiding erroneous merging caused by general model comprehension biases, while reducing semantic repetition and conflict in traditional automated extraction. By introducing standards such as DSM-5 and APA to align entities, standardization and normalization of entities are achieved, ensuring that the graph content conforms to medical and psychological standards, reducing repetition, conflict, and ambiguous concepts. By constructing a multi-level mapping system, bottom-up (from question to concept) dynamic reasoning is supported, providing interpretable logical support for intelligent psychological assessment, thereby realizing dynamic assessment and source tracing functions. Attached Figure Description

[0024] The following is a brief explanation of the contents of each of the accompanying drawings and the markings in the drawings:

[0025] Figure 1 This is a flowchart illustrating the overall method of this invention, showing the entire process from raw data preprocessing to final knowledge graph generation.

[0026] Figure 2 This is a system architecture diagram of an example of the present invention, which shows the hierarchical relationship of the system's functional modules and the data flow. Detailed Implementation

[0027] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and the description of the preferred embodiments.

[0028] This embodiment provides a method for constructing a knowledge graph in the field of mental health based on LLM and multi-level fusion, including the following steps:

[0029] S1: Data Preprocessing and Structuring: Obtain the original documents of the mental health assessment scale, perform text cleaning and format standardization, and separate the document content into two types of basic data: descriptive text and questionnaire assessment questions; use a large language model to convert the unstructured questionnaire questions into a structured data sequence containing question number, question stem, question source, options and associated factors;

[0030] S2: Entity Extraction and Seed Entity Set Construction: A seed entity generation strategy is adopted to extract measurement factors and dimension entities within a single scale; the extracted entities are then subjected to synonym merging, hierarchical processing, and standardized mapping to construct a standardized seed entity set Q;

[0031] S3: Relational Triple Extraction and Deduplication: Under the contextual constraints of a single scale, structural and semantic relations within the scale are extracted separately; for the extracted triples, the traditional vector similarity calculation is abandoned, and a large language model is used to perform semantic discrimination based on logical reasoning. Combined with a manual verification mechanism, semantic redundancy is eliminated and independent relations are retained to form a high-quality triple set.

[0032] S4: Multi-level graph fusion and dynamic mapping: A global three-level semantic skeleton containing a domain layer, a category layer, and a standard concept layer is constructed using the seed entity set Q; through a hierarchical mapping mechanism, the underlying entities extracted from each single scale are normalized and mapped to the standard concepts of the global skeleton, resolving local semantic conflicts; at the same time, questionnaire questions are integrated into the graph to construct a full-link mapping from questions to global concepts, realizing global fusion and dynamic completion of the knowledge graph.

[0033] In S1, the unstructured questionnaire questions are converted into a structured data sequence using a large language model. Specifically, prompts containing preset task instructions and output format requirements are constructed to drive the large language model to output the question number, question stem, question source, options, and associated factors in a structured JSON format.

[0034] The seed entity generation strategy in S2 specifically includes:

[0035] S201: Single Scale Entity Extraction: For each scale's descriptive text, accurately extract the scale's explicitly defined measurement factors and assessment dimensions;

[0036] S202: Concept Fusion and Standardization: The extracted entities are merged with synonyms, processed hierarchically, and normalized for variants. The terminology is standardized according to standards such as DSM-5 to form a seed entity set Q.

[0037] When using a large language model for semantic discrimination in S3, a logical reasoning discrimination method based on Prompt constraints is adopted. The input is a pair of triples to be compared, and the large language model is required to analyze whether their semantics are semantically equivalent, hierarchically inclusive, or mutually independent, and perform deduplication operation based on the judgment result.

[0038] The specific steps for constructing a global three-level semantic skeleton in S4 include:

[0039] S401: Semantic clustering is performed based on the seed entity set Q, and entities with similar semantics are aggregated into the same cluster;

[0040] S402: Construct the most representative three-level concepts for each cluster, and further summarize them into two-level concepts and one-level domains based on psychological standards to form a three-level skeleton system.

[0041] The dynamic access and hybrid completion strategy for questionnaire questions in S4 is as follows: if the concept measured by the question already exists in the three-level skeleton, it is directly connected to the skeleton node; if the concept belongs to a fine-grained construct unique to the scale and is not included in the skeleton, a new concept node is automatically created, and the large language model is used to dynamically map it to the upper-level node with the closest semantics.

[0042] like Figure 1 , Figure 2 As shown, by leveraging the powerful semantic understanding capabilities of large language models and employing a strategy of "intra-scale extraction - semantic deduplication - multi-level mapping," the professionalism, accuracy, and traceability of the graph are ensured. The methods include:

[0043] S1: Data Preprocessing and Structuring: Obtain the original documents of the mental health assessment scale, perform text cleaning and format standardization, and separate the document content into two types of basic data: descriptive text and questionnaire assessment questions; use a large language model to convert the unstructured questionnaire questions into a structured data sequence containing question number, question stem, question source, options and associated factors;

[0044] S2: Entity Extraction and Seed Entity Set Construction: A seed entity generation strategy is adopted to extract measurement factors and dimension entities within a single scale; the extracted entities are then subjected to synonym merging, hierarchical processing, and standardized mapping to construct a standardized seed entity set Q;

[0045] S3: Relational Triple Extraction and Deduplication: Under the contextual constraints of a single scale, structural and semantic relations within the scale are extracted separately; for the extracted triples, the traditional vector similarity calculation is abandoned, and a large language model is used to perform semantic discrimination based on logical reasoning. Combined with a manual verification mechanism, semantic redundancy is eliminated and independent relations are retained to form a high-quality triple set.

[0046] S4: Multi-level graph fusion and dynamic mapping: A global three-level semantic skeleton containing a domain layer, a category layer, and a standard concept layer is constructed using the seed entity set Q; through a hierarchical mapping mechanism, the underlying entities extracted from each single scale are normalized and mapped to the standard concepts of the global skeleton, resolving local semantic conflicts; at the same time, questionnaire questions are integrated into the graph to construct a full-link mapping from questions to global concepts, realizing global fusion and dynamic completion of the knowledge graph.

[0047] like Figure 2 The diagram shown is a general framework diagram of a knowledge graph construction method for the mental health assessment domain based on LLM and multi-level fusion. Based on this framework diagram, this solution specifically includes the following steps:

[0048] S1: Obtain the original documents of the mental health assessment scale. This example addresses the issue of multiple sources, multiple versions, and heterogeneous structures in the design of mental health assessment scales. The first step is data cleaning and structuring, specifically including:

[0049] S101: Use the PyMuPDF processing library to process the document, convert the PDF file into text, and then clean the text, including removing redundant formatting, correcting characters, and standardizing terminology.

[0050] S102: Separate the cleaned text content into a descriptive text file and a scale assessment problem file; the descriptive text file includes scale introduction, theoretical background, and factor definitions, etc.; the scale assessment problem file includes specific assessment questions and options.

[0051] S103: Before structuring the questionnaire questions, construct a set of prompt constraints based on mental health assessment and knowledge graph construction to standardize the processing of questionnaire items and the expression of results. Specifically, this includes extracting relevant information from each assessment item, including the unique identifier of the item (question_id), the semantic content of the item (question_text), the scale or source information to which the item belongs (question_source), the options (question_options), and the psychological dimension or factor corresponding to the item, thereby achieving a unified structured expression of assessment items from different sources of questionnaires.

[0052] S104: Use Large Language Model (LLM) in conjunction with prompt to transform unstructured questionnaire questions into structured JSON data.

[0053] This step transforms unstructured text into structured JSON data and stores it in a database.

[0054] S2: Entity Extraction and Seed Entity Set Construction. This example addresses the issue of noise generation during entity extraction in a zero-shot environment using LLM, employing a seed entity generation strategy, specifically including:

[0055] S201: First, perform single-scale entity extraction on the descriptive text file obtained from S1. That is, for each scale's descriptive text, accurately extract the measurement factors and assessment dimensions that are clearly defined by the scale.

[0056] S202: Conceptual fusion and standardization of the entities obtained in S201. The core entities that are retained are subjected to (1) merging of synonyms, (2) hierarchical processing, that is, identifying the inclusion relationship between entities, such as merging State Anxiety and Trait Anxiety under the concept of Anxiety, (3) variant normalization, that is, converting the adjective form (such as "Anxious") into the noun form ("Anxiety"), (4) standardization, according to psychological standards such as DSM-5, non-standard terms are uniformly mapped to standard medical terms to form the seed entity set Q.

[0057] S3: Extraction of relational triples within a single scale. This example first defines a relation type schema, constructing triples only within the relation type schema. To avoid cross-contextual ambiguity, the descriptive text file obtained in S1 is first processed by extracting structural and semantic relations within the single scale, then the triples are integrated and deduplicated. Specifically, this includes:

[0058] S301: Structural Relationship Extraction. This involves constructing a set of prompt constraints based on mental health assessment and knowledge graph construction. The aim is to standardize the processing of descriptive text for single mental health assessment scales and the expression of results. Specifically, this includes: the subordinate relationships between scales and factors, the hierarchical or membership relationships between factors within a scale, and standardizing the extraction of results using structured triples.<h,r,t> In this formal representation, h represents the head entity, r represents the relation type, and t represents the tail entity.

[0059] Finally, LLM combined with Prompt constraints was used to extract the scale's structural information.

[0060] S302: Semantic Relation Extraction. This involves constructing a set of Prompt constraints based on psychological mechanisms and knowledge graphs to standardize the processing methods and expression formats of implicit psychological mechanisms in the descriptive text of the scale. Specifically, this includes: identifiable psychological relation types, including correlation, causation, symptom-related relations, and risk-indicating relations; and standardizing the extraction results as semantic relation triples.<h,r,t> In formal representation, each triple contains a head entity, a relation type, and a tail entity. Finally, semantic relations are extracted using LLM combined with Prompt constraints.

[0061] S303: Dual deduplication mechanism:

[0062] (1) Precise deduplication: For identical triples<h,r,t> Where h represents the head entity, r represents the relationship between the connected entities, t represents the tail entity, and the record with the highest confidence score is retained.

[0063] (2) Semantic discrimination and manual collaborative deduplication based on LLM: Abandoning the traditional vector similarity calculation method, the Large Language Model (LLM) is used for logical reasoning discrimination based on Prompt constraints. First, a Prompt containing psychological knowledge background is constructed. The triple pairs to be compared are input, and LLM is required to analyze whether their semantics are (i) semantically equivalent, that is, the two triples in the triple pair point to the same psychological measurement construct, and there is no essential difference in measurement dimension, core connotation or application scenario; (ii) hierarchical inclusion, that is, there is a clear hierarchical relationship between the constructs of the two triples in the triple pair, one of which is a superordinate concept and the other is a subordinate concept; (iii) mutually independent, that is, the domain is related but they are different constructs or independent sub-dimensions, and their core connotation and measurement target are clearly distinguished. Finally, triples with high similarity (semantic equivalence) are automatically merged; those with medium similarity are manually filtered; and those with low similarity are retained as independent relationships.

[0064] For example, the triple <Physical Self-Care Ability Scale, is a subscale, Activities of Daily Living Scale> and the triple <Physical Self-Care Ability Scale, belongs to, Activities of Daily Living Scale> are determined to be semantically equivalent, and the one with higher confidence is retained. The triple <Competitiveness, is a factor, Type A behavior> and the triple <Competitiveness, is a dimension, Type A behavior> have medium similarity and await manual verification. Triples with low similarity or those determined to be hierarchically contained are retained as independent relations. Finally, the high-quality triples are stored in the database.

[0065] S4: Multi-level graph fusion and dynamic mapping. This embodiment describes in detail how to integrate discrete single-scale subgraphs into a multi-level graph with global consistency.

[0066] S401: Based on seed entities, a bottom-up, multi-level framework is constructed. Using the standardized seed entity set Q generated in S2 as a foundation, LLM is used to semantically cluster the seed entities. For example, semantically similar seed entities, such as sociability and excitability, are grouped into the same cluster. For each cluster, a representative tertiary concept is constructed, such as extraversion. Based on psychological standards, the tertiary concepts are further subdivided into secondary concepts, such as personality traits, and primary domains, such as personality domains, thus forming a three-level hierarchical framework.

[0067] S402: Entity Normalization Mapping. This establishes a mapping from the original scale entities to the tertiary skeleton. Specifically, it establishes a triple <original entity, mapped to, tertiary concept>. This step normalizes heterogeneous terms discrete across various scales to standard concepts, eliminating local conflicts and achieving semantic alignment.

[0068] S403: Dynamic access and hybrid completion of questionnaire questions. The questionnaire question (Item) JSON data obtained from S1 serves as the bottom-level node of the graph, constructing the <Question, Measurement, Concept> relationship. A hybrid connection strategy is executed for the concept measured by the question. If the concept already exists in the third-level skeleton, it is directly connected to the skeleton node. If the concept is a fine-grained construct specific to the scale and not included in the skeleton, the system automatically creates a new concept node and uses LLM to dynamically map it to the semantically closest upper-level node in the skeleton (such as a second-level concept), thereby achieving adaptive expansion of the graph while maintaining skeleton stability.

[0069] Through the above steps, a multidimensional knowledge graph is finally constructed that includes both fine-grained scale structures and general psychological concept levels, effectively supporting cross-scale comprehensive analysis and intelligent question answering.

[0070] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described in the above embodiments.

[0071] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0072] Obviously, the specific implementation of this invention is not limited to the above-described methods. Any non-substantial improvements made using the inventive concept and technical solution of this invention are within the protection scope of this invention.

Claims

1. A method for constructing a knowledge graph in the field of mental health assessment, characterized in that: Includes the following steps: S1: Data preprocessing and structuring: Process the original documents of the mental health assessment scales and convert them into descriptive text and structured data. S2: Entity extraction and seed entity set construction: Extract measurement factors and dimension entities within a single scale and construct a standardized seed entity set; S3: Relational triple extraction and deduplication. Under single-scale constraints, structural and semantic triples are extracted, and semantic discrimination and deduplication are performed using LLM. S4: Multi-level graph fusion and dynamic mapping constructs a global three-level semantic skeleton, normalizes entities and maps them to standard concepts, and enables the access of questions and graph completion.

2. The method for constructing a knowledge graph in the field of mental health assessment as described in claim 1, characterized in that: Step S1 also includes obtaining the original document of the mental health assessment scale, performing text cleaning and format standardization, and separating it into two types of basic data: descriptive text and questionnaire assessment questions; using a large language model to convert the unstructured questionnaire questions into a structured data sequence containing question number, question stem, question source, options and associated factors.

3. The method for constructing a knowledge graph in the field of mental health assessment as described in claim 2, characterized in that: Converting unstructured questionnaire questions into structured data sequences using a large language model includes: constructing prompts containing preset task instructions and output format requirements, driving the large language model to output the question number, question stem, question source, options, and associated factors in a structured JSON format.

4. The method for constructing a knowledge graph in the field of mental health assessment as described in claim 1, characterized in that: Step S2 also includes using a seed entity generation strategy to extract measurement factors and dimension entities within a single scale; performing synonym merging, hierarchical processing, and standardized mapping on the extracted entities to construct a standardized seed entity set Q.

5. The method for constructing a knowledge graph in the field of mental health assessment as described in claim 4, characterized in that: The seed entity generation strategy in step S2 includes: S201: Single Scale Entity Extraction: For each scale's descriptive text, accurately extract the scale's explicitly defined measurement factors and assessment dimensions; S202: Concept Fusion and Standardization: The extracted entities are merged with synonyms, processed hierarchically, and normalized for variants. The terminology is standardized according to standards such as DSM-5 to form a seed entity set Q.

6. The method for constructing a knowledge graph in the field of mental health assessment as described in claim 1, characterized in that: Step S3 includes extracting structural and semantic relationships within the scale under the constraints of a single scale context; using a large language model to perform semantic discrimination based on logical reasoning; and combining this with a manual verification mechanism to eliminate semantic redundancy and retain independent relationships, thereby forming a high-quality set of triples.

7. The method for constructing a knowledge graph in the field of mental health assessment as described in claim 6, characterized in that: In step S3, when using the large language model for semantic discrimination, a logical reasoning discrimination method based on Prompt constraints is adopted. The input is the triple pair to be compared, and the large language model is required to analyze whether their semantics are semantically equivalent, hierarchically contained, or mutually independent, and perform deduplication operation based on the judgment result.

8. The method for constructing a knowledge graph in the field of mental health assessment as described in claim 1, characterized in that: Step S4 includes constructing a global three-level semantic skeleton containing a domain layer, a category layer, and a standard concept layer using the seed entity set Q; normalizing and mapping the underlying entities extracted from each single scale to the standard concepts of the global skeleton through a hierarchical mapping mechanism to resolve local semantic conflicts; and at the same time, integrating the questionnaire questions into the knowledge graph to construct a full-link mapping from questions to global concepts, thereby achieving global fusion and dynamic completion of the knowledge graph.

9. The method for constructing a knowledge graph in the field of mental health assessment as described in claim 8, characterized in that: The dynamic access and hybrid completion strategy for questionnaire questions in step S4 is as follows: if the concept measured by the question already exists in the three-level skeleton, it is directly connected to the skeleton node; if the concept belongs to the fine-grained constructs specific to the scale and is not included in the skeleton, a new concept node is automatically created, and the large language model is used to dynamically map it to the upper-level node with the closest semantics.

10. A computer storage medium, characterized in that: The computer storage medium stores a computer program, which, when executed by a processor, implements the method as described in any one of claims 1-9.

Images