Large model-based training experience content acquisition method, device, equipment and medium

Abstract

This application relates to the fields of artificial intelligence and knowledge management technology, and discloses a method, apparatus, device, and medium for acquiring training experience content based on a large model. The method includes: performing semantic-level detection on experience text data using a large language model to extract candidate text fragments with experience description features; performing structured parsing on the candidate text fragments to generate structured experience vectors containing scene attributes, action logic, and result judgment fields; evaluating the structured experience vectors according to a preset utility evaluation strategy to obtain implicit experience levels, and selecting target experience vectors; adding the target experience vectors to an initial knowledge graph to obtain an experience knowledge graph; and extracting and outputting the corresponding training experience content from the experience knowledge graph in response to training information input. The beneficial technical effects of this application are: it can effectively solve the problems of difficult accurate identification of implicit experience, low degree of structure, and lack of dynamic evaluation of teaching value, thereby improving the quality and efficiency of knowledge transfer.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge management technology, and in particular to a method, apparatus, equipment and medium for acquiring training experience content based on a large model. Background Technology

[0002] With the development of artificial intelligence, large language models are widely used in natural language processing. Currently, they are mainly used to directly process technical problems. However, for some refined problems (such as the equipment failure handling skills of veteran employees, customer negotiation strategies, etc.), existing technologies usually treat them as general text and directly perform keyword retrieval, classification or summary generation, without a dedicated modeling and utilization mechanism for implicit experience features. Summary of the Invention

[0003] Based on this, it is necessary to address the existing problems of specialized modeling and utilization mechanisms for implicit experience features, and propose a method, device, equipment and medium for acquiring training experience content based on a large model.

[0004] A method for acquiring training experience content based on a large model, the method comprising: Using a large language model, semantic-level detection is performed on experiential text data to extract multiple candidate text fragments with experiential descriptive features; The candidate text fragments are subjected to structured parsing to generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; The structured experience vectors are evaluated according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each structured experience vector. The structured experience vector whose implicit experience level is greater than the preset experience level is denoted as the target experience vector; Each of the target experience vectors is added to a preset initial knowledge graph to obtain an experience knowledge graph; In response to the input of training information, the corresponding training experience content in the experience knowledge graph is extracted and output.

[0005] Furthermore, the step of using a large language model to perform semantic-level detection on experiential text data and extracting multiple candidate text fragments with experiential descriptive features includes: Identify the identity information of the dialogue subjects in the experiential text data; Analyze the density of causal logic words in the statements of the main participants in the dialogue; The identity information of the dialogue subject and the causal logic word density are used as joint discriminative features and input into the large language model to determine whether each text paragraph contains the intention to convey experience. Text segments deemed to contain the intention to convey experience are extracted as candidate text segments.

[0006] Furthermore, after the step of extracting text paragraphs determined to contain the intention to convey experience as candidate text fragments, the method further includes: Set minimum semantic unit constraints, which include: the text length is not less than a preset number of characters, and it contains at least one action verb and at least one conditional adverb; Based on the minimum semantic unit constraint, each candidate text segment is filtered, and candidate text segments that satisfy the minimum semantic unit constraint are retained.

[0007] Further, the step of performing structured parsing on the candidate text fragments to generate corresponding structured empirical vectors includes: Construct a standardized ontology framework; wherein, the standardized ontology framework defines four types of entities and their relationships: scene entities, condition entities, action entities, and result entities. The information of each field in the candidate text fragment is mapped to the standardized ontology framework, and symbolic vectors are generated by One-Hot encoding combined with TF-IDF weighting. The candidate text fragments are input into a large language model to extract and generate semantically dense vectors. The symbolic vector and the semantically dense vector are weighted and fused according to a preset weight ratio to generate the structured experience vector.

[0008] Further, the step of evaluating each of the structured experience vectors according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each of the structured experience vectors includes: The structured experience vectors are input into the large language model, and cognitive level constraint parameters are set. Under the cognitive level constraint parameters, the large language model is driven to simulate the processing corresponding to the structured experience vector, and generate a simulation result vector. Obtain the pre-stored ideal structured vector; Calculate the overall difference between the simulation result vector and the ideal structured vector; The implicit experience level corresponding to the structured experience vector is determined based on the comprehensive difference degree, wherein the comprehensive difference degree is positively correlated with the implicit experience level.

[0009] Further, the step of adding each of the target experience vectors to a preset initial knowledge graph to obtain an experience knowledge graph includes: Match the scene attribute fields in each of the target experience vectors with the existing graph nodes in the initial knowledge graph; When a match is successful, the target experience vector is stored as newly added attribute information under the matched graph node; When a match fails, a new graph node is created in the initial knowledge graph based on the scene attribute field of the target experience vector, and the target experience vector is stored under the new graph node. Establish a direct connection between the graph node corresponding to the target experience vector and the root node in the initial knowledge graph, and make the path weight of the direct connection higher than the path weight between the graph node corresponding to the non-high implicit experience level and the root node; wherein the non-high implicit experience level is an implicit experience level that is less than or equal to a preset experience level.

[0010] Furthermore, the step of extracting and outputting the corresponding training experience content in the experience knowledge graph in response to the input of training information includes: Receive training information including fault codes, equipment type, or geographical location; Using the fields in the training information as search keys, the scene attribute fields in the experience knowledge graph are matched and searched. Extract the structured empirical vectors stored under the matched graph nodes; According to the field order of the scenario attribute field, action logic field, and result judgment field, the extracted structured experience vectors are concatenated to generate training experience content in natural language format and output.

[0011] A training experience content acquisition device based on a large model, the device comprising: The detection module is used to perform semantic-level detection on experiential text data using a large language model, and extract multiple candidate text fragments with experiential descriptive features. The parsing module is used to perform structured parsing on the candidate text fragments and generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; The evaluation module is used to evaluate each of the structured experience vectors according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each of the structured experience vectors. A tagging module is used to mark structured experience vectors whose implicit experience level is greater than a preset experience level as target experience vectors; An add module is used to add each of the target experience vectors to a preset initial knowledge graph to obtain an experience knowledge graph; An output module is used to extract and output the corresponding training experience content from the experience knowledge graph in response to the input of training information. An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Using a large language model, semantic-level detection is performed on experiential text data to extract multiple candidate text fragments with experiential descriptive features; The candidate text fragments are subjected to structured parsing to generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; The structured experience vectors are evaluated according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each structured experience vector. The structured experience vector whose implicit experience level is greater than the preset experience level is denoted as the target experience vector; Each of the target experience vectors is added to a preset initial knowledge graph to obtain an experience knowledge graph; In response to the input of training information, the corresponding training experience content in the experience knowledge graph is extracted and output.

[0012] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: Using a large language model, semantic-level detection is performed on experiential text data to extract multiple candidate text fragments with experiential descriptive features; The candidate text fragments are subjected to structured parsing to generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; The structured experience vectors are evaluated according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each structured experience vector. The structured experience vector whose implicit experience level is greater than the preset experience level is denoted as the target experience vector; Each of the target experience vectors is added to a preset initial knowledge graph to obtain an experience knowledge graph; In response to the input of training information, the corresponding training experience content in the experience knowledge graph is extracted and output.

[0013] The beneficial effects of this invention are as follows: By utilizing a large language model to perform semantic-level detection on experiential text data, multiple candidate text fragments with experiential descriptive features are accurately extracted. These fragments are then subjected to structured parsing to generate structured experiential vectors containing scene attribute fields, action logic fields, and result judgment fields, thus realizing the transformation of implicit experience from unstructured text to standardized data form. Based on this, each structured experiential vector is evaluated according to a preset utility evaluation strategy to obtain the corresponding implicit experiential level. Vectors with levels higher than the preset experiential level are recorded as target experiential vectors, thereby selecting core experiential content with high teaching value. Subsequently, each target experiential vector is added to a preset initial knowledge graph to construct an experiential knowledge graph with a hierarchical structure. In response to the input of training information, the corresponding training experiential content in the graph is extracted and output. This enables the system to automatically identify and quantify the teaching value of experiences, avoid interference from low-quality or redundant information, improve the relevance and educational effectiveness of knowledge graph construction, enhance the readability and logical consistency of experience entries, ensure the traceability of the knowledge evolution process, and improve the coverage completeness and organizational participation of key experiences. Ultimately, it achieves the automated generation of scenario-based training content from unstructured logs. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] in: Figure 1 This is an application environment diagram of a training experience content acquisition method based on a large model in one embodiment; Figure 2 This is a flowchart of a method for acquiring training experience content based on a large model in one embodiment; Figure 3 This is a structural block diagram of a training experience content acquisition device based on a large model in one embodiment; Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0016] 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.

[0017] Figure 1 This is a diagram illustrating the application environment for acquiring training experience content based on a large model in one embodiment. (Refer to...) Figure 1 This method for acquiring training experience content based on a large model is applied to a system for acquiring training experience content based on a large model. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal, specifically a mobile phone, tablet, laptop, or other similar devices. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to extract and output the corresponding training experience content from the experience knowledge graph, while the server 120 is used to perform semantic-level detection on the experience text data.

[0018] like Figure 2 As shown, in one embodiment, a method for acquiring training experience content based on a large model is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to a terminal. The method for acquiring training experience content based on a large model specifically includes the following steps: S1: Use a large language model to perform semantic-level detection on experiential text data and extract multiple candidate text fragments with experiential descriptive features; S2: Perform structured parsing on the candidate text fragments to generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; S3: Evaluate each of the structured experience vectors according to the preset utility evaluation strategy to obtain the implicit experience level corresponding to each of the structured experience vectors; S4: The structured experience vector whose implicit experience level is greater than the preset experience level is denoted as the target experience vector; S5: Add each of the target experience vectors to the preset initial knowledge graph to obtain the experience knowledge graph; S6: In response to the input of training information, extract and output the corresponding training experience content in the experience knowledge graph.

[0019] As described in step S1 above, experiential descriptive features refer to the set of linguistic features contained in the text that indicate that the text is describing or imparting experience, including but not limited to: (1) causal logic relationship markers (such as 'because', 'therefore', 'leading to', 'at that time I judged'); (2) descriptions of the relationship between the subject of the action and the object of the operation; (3) time series or decision priority ranking words; (4) result status or effect evaluation words. When a text fragment contains at least two of the above features, it is determined to have experiential descriptive features. Experiential text data can refer to original records that exist in the form of unstructured narrative text accumulated within an organization, including but not limited to project review records, expert interview minutes, fault handling logs, work logs, work order records, transcripts of meeting recordings, and instant messaging records. The above data usually carries metadata such as timestamps and source identifiers, serving as the basis for subsequent knowledge tracing. Large Language Model (LLM) can refer to a pre-trained model with natural language understanding and generation capabilities, such as a generative model based on the Transformer architecture. The execution entity can be a data processing system configured with a large language model interface, or a terminal. This data is then uploaded to a server for processing, and the processed data is retrieved. The system inputs experiential text data into the large language model, using the model to deeply understand the semantic content and logical structure of the text to distinguish between experiential statements and general transactional descriptions. Specifically, the model identifies whether there is an intention to convey experience in the text by analyzing causal logical relationships, role behavior patterns, and the use of domain-specific terminology; alternatively, the model guides itself to identify similar text patterns using a few-shot prompting method based on typical examples of input experiential descriptions. In one implementation, a few-shot prompting method is used, specifically: a prompt template containing 3-5 typical examples of experiential descriptions is constructed, and the text paragraph to be detected is combined with the prompt template and input into the large language model, guiding the model to output classification labels indicating whether each paragraph contains an intention to convey experience. In another implementation, the distinction is made by analyzing the density of causal logical words and role identity information in the text. Both methods can be used individually or in combination depending on the actual application scenario.

[0020] As described in step S2 above, structured parsing aims to convert text fragments in natural language form into machine-readable structured data. A structured experience vector is a mathematical vector representation of the core elements of an experience. Its internal dimensions include at least: scenario attribute fields describing the specific background environment in which the experience occurred (such as device type, fault code, geographical location, or network topology); action logic fields recording the specific operational steps, decision-making basis, or processing flow taken in a specific scenario; and result judgment fields defining the output status after the operation (such as business recovery time, success flag, or failure behavior). During execution, the system uses a step-by-step prompting engineering strategy to guide the large language model to complete the parsing: first, identifying key entities in the scenario attribute fields (such as device status, environmental parameters, or error codes); then deriving the priority ranking logic and alternative selection mechanism in the action logic fields; and finally confirming whether the result judgment fields contain quantitative indicators or subjective evaluations.

[0021] As described in step S3 above, the preset utility evaluation strategy is a set of rules or algorithmic models used to quantitatively measure the teaching value and practical guidance significance of experience fragments. Implicit experience level is a quantitative evaluation index for the value of experience, used to characterize the importance of the experience in improving personnel capabilities and solving practical problems. By simulating the cognitive process, experiences that are easily overlooked in routine processing but have key corrective significance are automatically identified. In specific execution, the system inputs structured experience vectors into the large language model and sets cognitive level constraint parameters (e.g., simulating a low initial cognitive level, such as ≤60% from the perspective of new employees). Under this constraint, the large language model is driven to simulate the processing corresponding to the structured experience vector, generating a simulated result vector representing the decision-making or operational results that ordinary employees might produce when facing the same scenario. The simulated result vector has the same field structure as the structured experience vector, including scenario attribute fields, action logic fields, and result judgment fields. Specifically, under the cognitive level constraint parameters, the large language model autonomously derives and generates simulated action logic fields and result judgment fields based on the scenario attribute fields and conditional entities in the input structured experience vector. The simulated action logic and simulation result judgment text output by the model are transformed into the simulation result vector using the same structured parsing method as in step S2 (i.e., mapping text fragments to the standardized ontology framework and performing vectorization encoding). Simultaneously, the system obtains an ideal structured vector representing the optimal processing path, defined by expert experience or standard operating procedures. To reduce the influence of expert subjectivity on the ideal structured vector, a multi-expert consensus mechanism is adopted when constructing the ideal structured vector: at least three domain experts are invited to independently annotate and vectorize the same best practice case, the average similarity between the annotation results of each expert is calculated, outlier annotations with similarity below a preset threshold (e.g., 0.7) are removed, and the average of the remaining annotation results is used as the ideal structured vector. The cognitive level constraint parameters are set based on statistical analysis of the actual operation logs and behavioral data of new employees in the target position, rather than subjective assumptions, to ensure the objectivity of the cognitive simulation. Next, the comprehensive difference between the simulation result vector and the ideal structured vector is calculated. This calculation process uses cosine similarity to measure the overall semantic deviation, Euclidean distance to reflect the matching deviation of key fields, and the Jaccard index to evaluate the scene coverage overlap rate. The weights of each metric are dynamically adjusted based on the experience type. Finally, the implicit experience level corresponding to the structured experience vector is determined based on the comprehensive difference degree. The comprehensive difference degree is positively correlated with the implicit experience level; that is, the greater the difference, the more effectively the experience can correct common errors, and the higher its implicit experience level.In specific implementation, the cognitive level constraint parameters are transformed into input constraints for the large language model in the following way: First, a cognitive level description text is constructed, including but not limited to 'You are a new employee who has been with the company for 3 months and has not yet mastered the use of advanced diagnostic tools, only familiar with standard operating procedures'; the cognitive level description text is then input into the large language model as a system prompt along with the structured experience vector to be processed, to constrain the model's reasoning behavior. Optionally, the cognitive level constraint parameters also include a temperature parameter setting for the large language model. When simulating novice behavior, the temperature parameter is set to 0.7-0.9 to increase the randomness of decision-making; when simulating expert behavior, the temperature parameter is set to 0.1-0.3 to improve the certainty of decision-making.

[0022] As described in step S4 above, the preset experience level is a threshold for selecting high-value experiences. It can be determined based on statistical analysis of historical training effectiveness or by domain experts setting it in advance according to business risk levels. The preset experience level is a threshold used to determine whether an experience has high teaching value. Its value is preset by domain experts based on the knowledge transfer needs of the business scenario or determined based on statistical analysis of historical training effectiveness data. The value range is typically 0.6-0.8 (with normalized implicit experience levels as a reference). When the implicit experience level of a structured experience vector exceeds this threshold, it is considered to have high teaching value and is retained. The system compares the calculated implicit experience level with the preset experience level. If the implicit experience level of a structured experience vector is greater than the preset experience level, it indicates that the experience contains key non-explicit knowledge or can effectively avoid common errors, possessing extremely high transfer value. The system marks this structured experience vector as a target experience vector. Conversely, if the implicit experience level does not exceed the threshold, it is determined to be routine experience or low-value information and is not included in the target set. This screening mechanism ensures that the experience content entering the knowledge graph is characterized by high density and high relevance, thereby improving the quality of subsequent training content.

[0023] As described in step S5 above, the preset initial knowledge graph is a pre-constructed skeleton network containing basic domain concepts and entity relationships. The experience knowledge graph is a knowledge network with scenario-based reasoning capabilities formed by incorporating selected high-value target experience vectors into the initial knowledge graph. This step realizes the structured storage and association of experience knowledge. During execution, the system matches the scenario attribute fields in each target experience vector with existing graph nodes in the initial knowledge graph. When a match is successful, it indicates that a corresponding scenario entity already exists in the graph, and the system associates and stores the target experience vector as new attribute information under the matched graph node, enriching the attribute dimensions of that node. When a match fails, it indicates that the scenario is a newly discovered scenario, and the system creates a new graph node in the initial knowledge graph based on the scenario attribute fields of the target experience vector, and stores the target experience vector under the new graph node. Furthermore, to highlight the importance of high-value experiences, the system establishes direct connections between the graph nodes corresponding to target experience vectors with higher implicit experience levels (e.g., belonging to the high implicit experience level) and the root node in the initial knowledge graph. The path weight of this direct connection is higher than the path weight between graph nodes corresponding to non-high implicit experience levels and the root node. The root node is a concept node representing the highest level of the knowledge domain in the initial knowledge graph, such as 'equipment operation and maintenance experience base', 'fault handling knowledge system', or 'customer service experience set'. The root node serves as the logical starting point for all graph nodes and has hierarchical relationships with each level of child nodes. Through the above addition and association operations, an experience knowledge graph containing rich experience associations is ultimately formed. To control the expansion of the number of nodes in the knowledge graph, a node merging determination step is performed before creating a new graph node: the semantic similarity between the scene attribute field of the new node to be created and the scene attribute field of an existing graph node is calculated; if there is an existing node with a similarity higher than a preset threshold (e.g., 0.85), no new node is created, and the current target experience vector is stored under the existing graph node with the highest similarity. The semantic similarity is measured by the cosine similarity of the semantically dense vectors of the scene attribute fields extracted by the large language model.

[0024] As described in step S6 above, the training information is a query request initiated by the user, typically containing specific business scenario descriptions or problem characteristics, such as fault codes, equipment types, or geographical locations. This step is a crucial link in knowledge application, aiming to transform the implicit experience stored in the knowledge graph into training materials that can be directly learned. When the system receives training information containing fault codes, equipment types, or geographical locations, it uses the fields in the training information as search keys to match and search the scenario attribute fields in the experience knowledge graph. This search process can employ exact matching or semantic similarity retrieval methods. After the search is completed, the system extracts the structured experience vectors stored under the matched graph nodes. To restore the structured data to human-readable natural language, the system concatenates the extracted structured experience vectors in the order of scenario attribute fields, action logic fields, and result judgment fields, generating and outputting training experience content in natural language format. For example, the system generates a sentence structure of executing [action logic] under [scenario attribute], ultimately reaching [result judgment], thus generating fluent training text. The output content can be directly used to generate training manuals, troubleshooting guides, or pushed to the learning platform to achieve efficient reuse and inheritance of knowledge. The training experience content refers to the readable text content that is finally output to trainees or users. The difference between it and the original experience text data is: (1) The training experience content has undergone structured screening and standardization processing, eliminating redundant information and noise in the original text; (2) When the training experience content is output, it is spliced ​​together in natural language according to the three fields of scene attributes, action logic and result judgment, with a unified narrative structure, which is convenient for trainees to understand and learn.

[0025] In one implementation, to reduce the computational overhead of multi-step calls to large language models, at least one of the following optimization strategies is adopted: (1) a lightweight large language model (such as a model with 7 billion parameters) is used in steps S1 and S2, and a high-precision model with a larger number of parameters (such as a model with more than 13 billion parameters) is used only in step S3; (2) the semantic level detection in step S1 and the structured parsing in step S2 are merged into a single large language model call, and prompt words containing both detection and parsing tasks are designed so that the model can output detection labels and structured fields simultaneously in one inference process; (3) a caching mechanism is established for the processed experience text data, and the cached results are directly reused when the same or highly similar texts are processed again.

[0026] In one embodiment, step S1, which uses a large language model to perform semantic-level detection on experiential text data and extracts multiple candidate text fragments with experiential descriptive features, includes: S101: Identify the identity information of the dialogue subjects in the empirical text data; S102: Calculate the density of causal logic words in the speech of the main dialogue subject; S103: The identity information of the dialogue subject and the causal logic word density are used as joint discrimination features and input into the large language model to determine whether each text paragraph contains the intention to convey experience; S104: Extract the text paragraphs that are determined to contain the intention to convey experience as the candidate text segments.

[0027] As described in step S101 above, the identity information of the dialogue subjects, representing the speaker's role attributes and authority level, is extracted from the acquired experience text data. In experience transmission scenarios, the statements of different subjects have differentiated reference value; the statements of senior engineers, technical experts, or on-site supervisors usually contain a higher density of implicit experience. By parsing text metadata or using named entity recognition technology in natural language processing, the speaker's identification information is located and extracted. Specific extraction methods include: scanning speaker tags embedded in the text (e.g., [senior engineer]), retrieving the role definition table associated with the system login account, or reading the job title field of the document signer. The identified dialogue subject identity information is converted into quantifiable or categorizable feature inputs, such as mapping roles to discrete levels like expert, senior, ordinary, and junior, or mapping them to numerical vectors representing experience weights. The converted features serve as a priori basis for subsequently determining the intention of experience transmission, defining the authority boundary for semantic detection, thereby effectively filtering out general chatter or low-value descriptions. In one implementation, when the speaker's role tag or position field is not directly embedded in the experience text data, the large language model is used to analyze the density of technical terms, frequency of imperative statements, and decision-making tone in the speech content to infer and determine the speaker's experience level. Specifically, a role inference prompt is constructed: "Based on the following speech content, determine whether the speaker's experience level is expert, advanced, ordinary, or beginner." The speech text is then input into the large language model to obtain the role classification result, which serves as an alternative source of the dialogue subject's identity information.

[0028] As described in step S102 above, based on the identified dialogue subject, the density of causal logic words is statistically calculated in its spoken text. Causal logic word density is a quantitative measure of the logical rigor and depth of empirical description in text, reflecting the complexity of the causal chain in the statement. This indicator is not a simple keyword count, but rather constructs a structured sequence representing the strength of text logic by statistically analyzing the frequency of specific logical connectors within a unit of text length. Within the text segment spoken by the dialogue subject, a set of words representing causal relationships, such as "because," "therefore," "due to," "thus," "leading to," "at that time I judged," and "based on this," is scanned; the ratio of the total number of these words to the total number of characters (or words) in the text is calculated to obtain the density value. The formula for calculating this density value is expressed as: ,in, Indicates the density of causal logic words. Indicates the number of causal logical terms, The total text length (unit: characters) includes Chinese characters, English letters, and numbers, but excludes spaces and punctuation. The causal logic words are a predefined vocabulary set, including but not limited to 'because', 'therefore', 'due to', 'thus', 'leading to', 'based on this', 'at that time I judged', 'the reason', and 'its reason'. This vocabulary set is expanded and updated based on business domain knowledge. In specific implementation, a density calculation window is set, and independent statistics are performed for each semantic paragraph or consecutive speaking round to ensure that the values ​​accurately reflect the logical characteristics of the current segment. Compared to the traditional approach that only focuses on whether keywords appear, introducing the density of causal logic words can effectively distinguish between simple chronological records and experiential descriptions containing in-depth reflection, providing a key semantic feature dimension for determining the intent of experience transmission.

[0029] As described in step S103 above, the aforementioned dialogue subject identity information and causal logic word density are fused to construct joint discriminative features, which are then input into the Large Language Model (LLM) for intent determination. This joint discriminative mechanism aims to address the problem of high misclassification rates of single features in complex contexts, improving decision robustness through multi-dimensional feature complementarity. The Large Language Model, as the core decision engine, utilizes its contextual understanding capabilities to receive joint input features that integrate identity authority and logical strength. The specific execution process includes: constructing a prompt word prefix containing role labels (e.g., senior engineer) and causal logic word density values ​​(e.g., 1.5 words / 100 characters), combining it with the text paragraph to be detected, and inputting it into the model; guiding the model to output a binary classification result, i.e., whether it contains an intent to convey experience or not. At the model architecture level, a finely tuned Large Language Model, such as LLaMA-2 or the GPT series, is used. Through supervised fine-tuning on the experience dataset, the model learns the inherent rule that high-weight roles combined with high logic density often correspond to high-value experience patterns. When the model receives text such as a senior engineer's speech containing a "because...therefore..." logical chain, its attention mechanism focuses on the causal reasoning part, thereby determining that it contains an intention to convey experience. Through this joint discrimination mechanism, it achieves a leap from shallow semantic matching to deep intent perception, significantly improving the accuracy of accurately locating effective experience fragments in lengthy meeting minutes or mixed data streams.

[0030] As described in step S104 above, the original experience text data is sliced ​​based on the judgment labels output by the large language model, and text segments judged to contain experience transmission intentions are extracted as candidate text segments. This step transforms the semantic judgment results into executable physical data operations, filtering high-value targets from massive unstructured data. The candidate text segments are pure data slices after intent filtering, retaining the core context description and solution logic in the original text. When the model determines that a certain paragraph or several consecutive rounds of dialogue contain experience transmission intentions, the system extracts and encapsulates this part of the text along with its metadata (such as timestamps, source document IDs, and context offsets) into independent text objects. To ensure the usability of the extracted content, minimum semantic unit verification is performed after extraction, such as removing segments that are too short (less than the preset number of characters) or missing core verbs. The final generated candidate text segments serve as the input data source for the downstream structured parsing module, and their quality directly determines the accuracy and completeness of the subsequent knowledge graph construction. The above technical features work together to construct a complete filtering link from identity authority recognition to logical strength quantification to accurate intent judgment. The identity information of the dialogue participants provides a priori reference for the credibility of the content, while the causal logic word density characterizes the objective indicators of the content's logicality. These two elements, as joint discriminative features input into the large language model, effectively compensate for the lack of dedicated intent modeling capabilities in vertical domains by general-purpose models. This multi-dimensional feature coupling judgment method enables the system to accurately eliminate low-value transactional descriptions and noisy data, efficiently transforming text paragraphs containing real implicit experiences into candidate text fragments. This lays a high-quality data foundation for subsequent structured extraction of experience and construction of knowledge graphs.

[0031] In one embodiment, after step S104 of extracting text paragraphs determined to contain an intention to convey experience as candidate text fragments, the method further includes: S1051: Set minimum semantic unit constraints, the minimum semantic unit constraints include: the text length is not less than a preset number of characters, and it contains at least one action verb and at least one conditional adverb; S1052: Based on the minimum semantic unit constraint, filter each candidate text segment and retain the candidate text segments that satisfy the minimum semantic unit constraint.

[0032] As described in step S1051 above, a minimum semantic unit constraint is set. This constraint includes: the text length is not less than a preset number of characters, and it contains at least one action verb and at least one conditional adverb. This minimum semantic unit constraint is used to select fragments with complete experiential expression potential from the initially extracted text paragraphs. Its construction is based on the fact that implicit experience usually exists in a compact logical structure of context-action-result. This condition is not a single-dimensional length limit, but a composite rule that integrates textual quantitative features and grammatical logic features, aiming to ensure that the input source of the subsequent structured parsing module has minimum semantic consistency and information completeness. During system initialization or rule configuration, this constraint is constructed by combining the linguistic distribution characteristics of implicit experience with business scenario requirements. Specifically, the input information comes from statistical analysis of a historical high-value experience database, covering a large number of manually annotated high-quality experience samples. The construction logic is based on syntactic analysis and keyword matching techniques in Natural Language Processing (NLP). By statistically analyzing the character length distribution range and core part-of-speech combination patterns of experiential statements in high-quality samples, a threshold standard that can cover the vast majority of effective experience while eliminating noise is extracted. The resulting constraints contain multiple sets of discrimination parameters, stored in a structured configuration file, supporting dynamic parameter adjustment and loading based on different business scenarios (such as fault handling, sales techniques, and customer communication). The minimum semantic unit constraint specifically includes two dimensions of limiting indicators. First, the text length dimension requires the text length to be no less than a preset number of characters. This preset number of characters aims to filter out overly short phrases or fragmented sentences lacking sufficient information, such as fragmented expressions lacking contextual support like "okay," "received," or "restart." The character threshold can be set to, for example, 50 or 30 characters based on the experience and complexity of the target domain; this is merely an example, and the specific value can be adjusted based on the statistical quantile of the actual data distribution. Second, the semantic composition dimension requires that the fragment must contain at least one action verb and at least one conditional adverb. Action verbs are used to characterize specific intervention behaviors taken by the subject of experience, such as restarting, switching, isolating, and analyzing. Their identification process can be achieved by loading a predefined domain verb dictionary or using a part-of-speech tagging model. Conditional adverbs are used to define the prerequisites, preconditions, or environmental constraints that trigger the behavior, such as when CPU utilization exceeds 90%, when voltage fluctuation is greater than 15%, or when a connection timeout occurs. Their identification can be accomplished by retrieving specific syntactic structure templates (such as when, under, or specific conjunctions). By simultaneously setting the above length and part-of-speech constraints, this condition can accurately identify complete expression units that possess both a certain information density and implicitly contain a condition-action causal chain in their logical structure.

[0033] As described in step S1052 above, each candidate text fragment is filtered according to the minimum semantic unit constraint, retaining only those that meet the constraint. A rigorous multi-dimensional filtering process further refines the candidate text set obtained in the preceding steps, eliminating low-quality text fragments caused by dialogue turn switching, interruptions, or incomplete recording. This filtering process is a crucial preliminary step in improving the quality of subsequent knowledge graph construction, directly determining the purity of the data input to the structured parsing module. Each candidate text fragment extracted in the above embodiments is traversed, and dependency parsing or part-of-speech tagging is performed on each fragment to accurately capture verbs and adverbial components. Subsequently, the total number of characters in each fragment is calculated and compared with a preset number of characters; simultaneously, the set of verb and adverbial keywords in the fragment is scanned to confirm the existence of words that meet the definitions of action verbs and conditional adverbials. For any candidate text fragment, it is only considered valid if its character length meets the threshold requirement and it simultaneously contains both action verbs and conditional adverbials. Fragments that do not meet this constraint, such as short instructions containing verbs but insufficient in length, or paragraphs that are long enough but merely descriptive statements lacking action guidance, are filtered out. Through this hard filtering based on grammatical structure, the system effectively removes noisy data that, although judged by the large language model to contain experiential intent, is fragmented in its surface structure and lacks key logical elements. This ensures that every piece of data entering subsequent processes is a high-quality candidate that meets the minimum semantic unit standard. These technical features work closely together to construct the first line of defense for ensuring the quality of experience extraction. The minimum semantic unit constraint serves as the benchmark for filtering logic, defining the minimum quality standard for experiential text; while the detection mechanism for action verbs and conditional adverbs is the concrete means of implementing this constraint. Together, they transform the abstract requirement for experiential completeness into computable and automated engineering judgment rules. This screening results in more regular morphology and tighter internal logic in the final retained candidate text fragments, providing a high-quality data foundation for the subsequent accurate conversion of unstructured text into structured scenario-action-result triplets, significantly reducing the risk of parsing distortion caused by incomplete input data.

[0034] In one embodiment, step S2, which involves performing structured parsing on the candidate text fragments to generate corresponding structured empirical vectors, includes: S201: Construct a standardized ontology framework; wherein, the standardized ontology framework defines four types of entities and their relationships: scene entities, condition entities, action entities, and result entities; S202: Map the information of each field in the candidate text fragment to the standardized ontology framework, and generate symbolic vectors by using One-Hot encoding combined with TF-IDF weighting; S203: Input the candidate text fragments into a large language model to extract and generate semantically dense vectors; S204: The symbolic vector and the semantically dense vector are weighted and fused according to a preset weight ratio to generate the structured experience vector.

[0035] As described in step S201 above, a standardized ontology framework is constructed. This framework defines four types of entities and their relationships: scenario entities, condition entities, action entities, and result entities. It establishes the logical skeleton for structured experience parsing, aiming to provide a foundation for extracting and standardizing experience elements from unstructured text. This standardized ontology framework is not simply a collection of classification labels, but rather an abstract model of the inherent causal logic and contextual constraints of implicit experience. It adopts an entity-relationship architecture pattern, defining four core entities: scenario entities, condition entities, action entities, and result entities. Scenario entities represent the macro-level business background or equipment environment at the time the experience occurs, such as core switch failure or database connection pool exhaustion, limiting the spatiotemporal boundaries of the applicable experience. Condition entities define the preconditions that trigger specific experience operations, typically including equipment status parameters, environmental indicators, error codes, or operational constraints. Action entities record the specific intervention behaviors or decision logic executed in response to the triggering conditions, involving operational steps, priority ranking, and resource allocation. Result entities describe the objective state changes or subjective evaluations after the action is executed, such as business recovery, shortened response time, or no customer complaints. The four types of entities are closely linked through predefined relationships, forming a complete semantic link of scenario-condition-action-result. When constructing this framework, the hierarchical structure and attribute constraints of each type of entity are established based on high-frequency co-occurrence patterns in the domain knowledge graph or historical data. For example, condition entities must contain timestamp attributes and logically occur before action entities, thereby ensuring that the subsequently extracted experience vectors have logical consistency and interpretability.

[0036] As described in step S202 above, the field information in the candidate text fragments is mapped to a standardized ontology framework, and symbolic vectors are generated using One-Hot encoding combined with TF-IDF weighting. By performing structured mapping and statistical feature extraction on the text fragments, natural language is transformed into machine-readable symbolic representations. First, the candidate text fragments extracted in the aforementioned steps are parsed at the field level. Based on predefined rules or regular expressions, key information units in the text are accurately identified and categorized into corresponding entity slots under the standardized ontology framework. For example, "when CPU utilization exceeds 90%" is mapped to a condition entity slot, and "restart management service" is mapped to an action entity slot. After mapping, for the vocabulary content in each entity slot, One-Hot encoding is used to convert it into a binary vector form. The dimension of this vector is equal to the size of the preset vocabulary, and each dimension uniquely corresponds to a vocabulary word. If the current entity contains the vocabulary word, it is set to 1; otherwise, it is set to 0. To address the limitations of One-Hot encoding in representing word importance and text sparsity, a TF-IDF (Term Frequency-Inverse Document Frequency) weighting strategy is further employed to process the encoding results. Specifically, the term frequency (TF) of each word in each field is calculated to reflect its local importance within the current text segment, while the inverse document frequency (IDF) is calculated to measure the rarity of the word across the entire corpus. Finally, the product of TF and IDF is used as the weight coefficient for the corresponding One-Hot vector dimension. Through this statistical learning approach, diagnostic keywords or technical terms that frequently appear in specific experiential segments but are rare in other scenarios are assigned higher weights. The resulting symbolic vectors not only retain the structured features of the experience but also highlight discriminative information, providing a sparse yet accurate feature foundation for subsequent vector space measurements.

[0037] As described in step S203 above, candidate text fragments are input into a large language model to extract and generate semantically dense vectors. Leveraging the powerful semantic understanding capabilities of deep learning models, deep features of text fragments are captured from a continuous semantic space. Candidate text fragments are directly input into a pre-trained large language model, which can employ architectures such as BERT (Bidirectional Encoder Representations from Transformers), RoBERTa (Robustly Optimized BERT Approach), or GPT (Generative Pre-trained Transformer). During processing, the large language model uses its internal Transformer network layers to perform self-attention mechanisms on the contextual information of the input text, mapping the word sequence into a dense representation in a low-dimensional real-valued vector space. Semantically dense vectors are characterized by low dimensionality (e.g., 768 or 1024 dimensions) and high information density, implicitly expressing semantic similarity, syntactic structure, and implicit intent between words. For example, different representations of restarting services and resetting processes might be considered completely different features in symbolic vectors, but in semantically dense vectors, their distance in the vector space will be very close, thus demonstrating the generalization ability of semantics. Extracting the last hidden state of the large language model or using the corresponding output vector labeled [CLS] as a semantically dense vector comprehensively represents the integrated information of candidate text fragments at the semantic level, effectively compensating for the shortcomings of purely symbolic methods in understanding complex linguistic logic.

[0038] As described in step S204 above, the symbolic vector and the semantically dense vector are weighted and fused according to a preset weight ratio to generate a structured experience vector. By complementaryly fusing symbolic features and semantic features, a composite vector representation with both structural regularity and semantic robustness is constructed. The weighted fusion process aims to balance the advantages of symbolic vectors in precise matching with the strengths of semantically dense vectors in generalization and association. The preset weight ratio between the symbolic vector and the semantically dense vector is set according to the type characteristics of the experience data or the accuracy requirements of the application scenario. For example, for operational experience with clear processes and high standardization, the weight of the symbolic vector can be appropriately increased to ensure the rigor of the logical structure; for fault diagnosis experience that relies on expert intuition and has varied expressions, the weight of the semantically dense vector can be increased to enhance the inclusiveness of different expression methods. In the specific fusion calculation, the two vectors are first normalized to eliminate the influence of dimensions, and then the corresponding dimensions of the two vectors or the overall vector after concatenation are weighted and summed according to the preset weights, or the vectors are directly concatenated. For example, if the symbolic vector weight is set to 0.4 and the semantically dense vector weight to 0.6, the final generated structured experience vector is a linear combination of the two. This structured experience vector, used as input data for subsequent utility evaluation and discrepancy ratios, contains both structured entity information based on the ontology framework and deep semantic relationships based on context, thus significantly improving the completeness and computability of the experience representation and laying a data foundation for accurately identifying implicit experience levels.

[0039] By constructing a standardized ontology framework for four types of entities, a structured parsing specification for implicit experience was established. Symbolic vectors generated using One-Hot encoding combined with TF-IDF weighting accurately characterize the distribution of key feature words in experience entries. Semantically dense vectors extracted from a large language model preserve the deep semantic logic of the text in a continuous space. Finally, the two are fused according to a preset weight ratio to generate a composite vector that simultaneously reflects structural and semantic information. This dual-channel embedding architecture effectively overcomes the limitations of single feature representation in information carrying capacity, enabling the generated structured experience vectors to possess higher discriminative power and accuracy in subsequent utility evaluations.

[0040] In one embodiment, step S3, which evaluates each of the structured experience vectors according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each of the structured experience vectors, includes: S301: Input the structured experience vector into the large language model and set the cognitive level constraint parameters; S302: Under the cognitive level constraint parameters, drive the large language model to simulate the processing corresponding to the structured experience vector and generate a simulation result vector; S303: Obtain the pre-stored ideal structured vector; S304: Calculate the overall difference between the simulation result vector and the ideal structured vector; S305: Determine the implicit experience level corresponding to the structured experience vector based on the comprehensive difference degree, wherein the comprehensive difference degree is positively correlated with the implicit experience level.

[0041] As described in step S301 above, the structured experience vector is input into the large language model, and cognitive level constraint parameters are set. The structured experience vector generated in the preceding steps is fed into the large language model as input data. This structured experience vector is a formal representation of implicit experience in specific business scenarios, and it encapsulates a complete causal chain of scenario-condition-action-result. At the same time, to ensure the relevance and effectiveness of the subsequent simulation process, cognitive level constraint parameters are set for the reasoning state of the large language model. This parameter is not a simple on / off command, but a deep limitation on the model's thinking boundaries and knowledge access scope. Its construction is based on the cognitive ability profile of new employees or beginners in the target position in a specific business domain, covering dimensions such as knowledge mastery, operational proficiency, and decision-making experience thresholds. In practice, the process begins by collecting historical operation logs and training assessment data from new employees in the target positions. This data is then statistically analyzed to identify the distribution of decision-making biases and common error types when handling routine tasks and unexpected malfunctions. Subsequently, based on the statistical results, feature vectors describing the cognitive state of new employees are constructed. For example, setting a knowledge mastery level below a certain percentage (e.g., 60%) or restricting access to complex tools are considered. Finally, these features are transformed into system prompts or instruction constraints that can be understood and followed by the large language model. This approach creates a cognitive firewall at the model's inference level, ensuring that the simulated behavioral logic closely matches the actual skill level of a new employee. For instance, in a fault diagnosis scenario, constrained by cognitive level parameters, the model will be guided to prioritize basic troubleshooting instructions rather than directly activating advanced diagnostic expert systems. This naturally results in decision-making biases due to a lack of experience, as reflected in the simulation results.

[0042] As described in step S301 above, under the cognitive level constraint parameters, the large language model is driven to simulate the processing corresponding to the structured experience vector, generating a simulation result vector. After establishing the cognitive level constraint parameters, the large language model is driven to perform a simulation inference task. This process aims to reproduce the complete thought and behavior path of an agent with the aforementioned limited cognitive level when facing a specific business scenario described by the structured experience vector, from receiving the triggering conditions to taking the final action. During the processing, the large language model does not simply repeat the input text, but rather performs dynamic contextual reasoning and decision generation based on its internally pre-trained world knowledge and logical reasoning ability, combined with the instruction constraints of the cognitive level constraint parameters. Specifically, the conditional entities and scene entities in the structured experience vector are parsed into the initial state description of the simulation task, and the model is required to autonomously deduce the action entities to be taken without relying on external additional prompts. Due to the constraints of the cognitive level constraint parameters, the model will simulate behavioral characteristics that a real novice might exhibit, such as hesitation, incorrect path selection, or neglect of key details. For example, when faced with a device error code, due to a limited cognitive level, the model might prioritize a generic reboot operation while ignoring fine-grained configuration fixes specific to the error code. The model's final output, the simulated action sequence and its corresponding expected results, is encoded as a vector, i.e., a simulation result vector. This vector accurately records the decision-making output under a limited cognitive perspective, and its data dimensions are aligned with the original structured experience vector for subsequent quantitative comparison.

[0043] As described in step S303 above, obtain the pre-stored ideal structured vector. Obtain the ideal structured vector from a pre-set knowledge base or expert experience base to establish a benchmark for evaluating the value of experience. An ideal structured vector can be a vectorized representation of a standard operation or decision-making path that, under specific business scenarios, has been verified by experts or numerous successful cases and can achieve optimal processing results (such as shortest recovery time, lowest resource consumption, zero side effects, etc.). The construction of this vector is a rigorous knowledge accumulation process: during the system's offline preprocessing stage, collect recognized best practice cases, standard operating procedures, and typical cases approved by experts; using the same ontology framework and coding rules as those used to generate structured experience vectors, map these high-quality experience samples into points in the vector space; through multiple rounds of expert verification and consistency testing, select the most representative vectors in various scenarios as ideal structured vectors for storage. The ideal structured vector represents the optimal solution for the scenario under current cognitive conditions; it not only contains the correct action sequence but also implicitly includes complete conditional judgments and accurate result expectations. The ideal structured vector is generated in the same way as the structured experience vector in step S2: First, the best practice text fragments verified by experts are mapped according to the standardized ontology framework to extract scene entities, condition entities, action entities, and result entities; then, symbolic vectors are generated by One-Hot encoding combined with TF-IDF weighting, and input into a large language model to extract semantically dense vectors; finally, the symbolic vectors and the semantically dense vectors are weighted and fused according to a preset weight ratio to generate the ideal structured vector.

[0044] As described in step S304 above, the comprehensive difference degree between the simulation result vector and the ideal structured vector is calculated. After obtaining the simulation result vector and the ideal structured vector, the comprehensive difference degree between the two is calculated to quantify the degree of deviation between the simulated behavior and the ideal standard. This comprehensive difference degree is a composite score calculated based on the fusion of multi-dimensional metrics, aiming to comprehensively capture various omissions and biases that may exist in the experience transmission process. The calculation process specifically includes: First, using cosine similarity to measure the directional difference between the simulation result vector and the ideal structured vector in the semantic space to reflect the degree of deviation of the overall semantic intent; second, using Euclidean distance to calculate the straight-line distance between the two in the vector space to sensitively capture the matching deviation of key field values ​​or specific operation steps; third, applying the Jaccard index to evaluate the overlap rate of the entity sets (such as involving equipment, operation instructions, etc.) corresponding to the two vectors to determine whether there are omissions of key decision branches in the simulation results. The above three metrics characterize the differences from three levels: semantic, numerical, and set coverage. A difference scoring function is defined, which dynamically assigns weights to different metrics (e.g., assigning higher weights to semantic direction differences for fault scenarios with strong logical dependencies), and aggregates them into a normalized comprehensive difference score. This evaluation process can be formalized as the following decision function: In the formula, Represents the simulation result vector. Represents an ideal structured vector. and These represent the entity sets corresponding to the two; , , The preset weighting coefficients are used, and they satisfy the following conditions: ; This represents the function for calculating cosine similarity. express and The standardized Euclidean distance, Representing vectors and The overlap rate of the corresponding entity sets (such as those involving equipment, operation instructions, etc.). The higher the comprehensive difference value, the more significant the gap between the automatically derived processing method and the ideal best practice under cognitively limited conditions. This means that the experience contained in the structured experience vector is more difficult for ordinary novices to learn naturally, and its teaching value and necessity for explicitness are higher. In one implementation, the weight coefficients are dynamically adjusted according to the type of the experience text data: when the experience text data is a fault handling experience with strong logical dependence, w1 (semantic direction difference weight) is increased to 0.5; when the experience text data is an experience with precise operation steps, w2 (numerical matching deviation weight) is increased to 0.5. The dynamic adjustment is achieved through a preset weight configuration table.

[0045] As described in step S305 above, the implicit experience level corresponding to the structured experience vector is determined based on the comprehensive difference degree, where the comprehensive difference degree and the implicit experience level are positively correlated. Based on the calculated comprehensive difference degree, the implicit experience level corresponding to the structured experience vector is finally determined. The implicit experience level is a quantitative classification of the importance and scarcity of the experience item in the training system. Since the comprehensive difference degree directly reflects the deviation of decision-making behavior from the ideal standard when the experience is lacking, the two are positively correlated: the greater the comprehensive difference degree, the higher the contribution of the experience to avoiding erroneous decisions and improving processing effectiveness, and the higher its corresponding implicit experience level. The specific determination process includes: setting implicit experience grading standards, such as dividing implicit experience levels into three levels: low, medium, and high, each corresponding to different comprehensive difference intervals; comparing the currently calculated comprehensive difference value with preset grading boundary thresholds; if the comprehensive difference is lower than the first threshold, it is judged as ordinary experience, with a low level; if the comprehensive difference is between the first and second thresholds, it is judged as advanced experience, with a medium level; if the comprehensive difference is higher than the second threshold, it is judged as core implicit experience, with a high level. For example, in a power module replacement scenario, if the simulation result has an extremely high comprehensive difference (e.g., 0.82) because it does not include electrostatic protection steps, then this experience is marked as a high-level implicit experience, meaning that this step is a key experience point that beginners are very likely to overlook and that has serious consequences. Through this rating mechanism based on differentiated behavior simulation, it is possible to automatically filter out experience fragments that truly have high teaching value, rather than relying solely on text frequency or keyword matching, thereby ensuring that the content ultimately included in the knowledge graph has strong practical guidance significance and corrective value. In the aforementioned technical process, the cognitive level constraint parameter injects the cognitive biases of real novices into the simulation process, enabling the simulation result vector to expose the actual risks caused by a lack of experience. Pre-stored ideal structured vectors serve as standard anchor points, providing an objective benchmark for measuring bias. The multi-dimensional calculation mechanism of comprehensive difference transforms abstract experiential value into calculable vector space distance. The process of determining implicit experience levels then uses this quantitative indicator to accurately calibrate the experiential teaching value. By mapping simulated bias behaviors to experience levels, this scheme solves the problem of traditional methods struggling to identify high-value experiential nodes, achieving in-depth mining and quantitative evaluation of implicit experiential value, making the selected experiential content more targeted and practical. The first and second thresholds are determined statistically based on the comprehensive difference distribution of historical experience data. Specifically, a preset number of historical experience samples (e.g., 1000) are collected, the comprehensive difference of each sample is calculated, and the distribution curve of the comprehensive difference is plotted. The 25th percentile of the distribution curve is set as the first threshold, and the 75th percentile is set as the second threshold. Alternatively, based on the experience of domain experts, the first threshold is set to a value range of 0.3-0.5, and the second threshold is set to a value range of 0.6-0.8.When the overall difference is higher than the second threshold, it is judged as a high level of implicit experience.

[0046] In one embodiment, step S5, which involves adding each of the target experience vectors to a preset initial knowledge graph to obtain an experience knowledge graph, includes: S501: Match the scene attribute fields in each of the target experience vectors with the existing graph nodes in the initial knowledge graph; S502: When a match is successful, the target experience vector is stored as newly added attribute information under the matched graph node; S503: When matching fails, a new graph node is created in the initial knowledge graph based on the scene attribute field of the target experience vector, and the target experience vector is stored under the new graph node; S504: Establish a direct connection between the graph node corresponding to the target experience vector and the root node in the initial knowledge graph, and make the path weight of the direct connection higher than the path weight between the graph node corresponding to the non-high implicit experience level and the root node; wherein the non-high implicit experience level is an implicit experience level that is less than or equal to a preset experience level.

[0047] As described in step S501 above, the scene attribute fields in each target experience vector are matched with the existing graph nodes in the initial knowledge graph. The scene attribute fields contained in the [target experience vector] represent the specific business environment or operating condition type to which the experience fragment is applicable, such as high load on the core switch, database connection pool exhaustion, or significant voltage fluctuations. The scene attribute field is extracted from the [target experience vector] as a retrieval key and compared by traversing the existing node set in the [initial knowledge graph]. This matching process aims to identify whether the current experience belongs to an existing knowledge classification system so that the newly acquired experience can be attached to the correct knowledge branch. If there is a [graph node] in the [initial knowledge graph] whose name or semantic label is completely consistent with the current scene attribute field or whose similarity is higher than a preset threshold, the match is considered successful, laying the topological foundation for subsequent attribute fusion.

[0048] As described in step S502 above, when a match is successful, the target experience vector is stored as a new attribute information under the matched graph node. After confirming the existence of a compatible graph node, the target experience vector is considered an enrichment and refinement of the existing knowledge node. The target experience vector is then attached as a new attribute information or sub-entry under the successfully matched graph node. This associated storage method preserves the hierarchical structure of the graph, avoids the disorderly expansion of the number of nodes, and enhances the completeness and coverage of the knowledge node by aggregating experience content from different dimensions or sources under the same node. For example, under the graph node of network operation and maintenance - high availability assurance, multiple target experience vectors related to device restart, link switching, and traffic cleaning may be stored simultaneously, thus forming a complete set of problem-solving solutions.

[0049] As described in step S503 above, when a match fails, a new graph node is created in the initial knowledge graph based on the scene attribute field of the target experience vector, and the target experience vector is stored under the new graph node. If no existing graph node matching the current scene attribute field is found during the matching process, it indicates that the current experience represents a completely new scene or domain not yet covered in the knowledge graph. A new graph node is dynamically created in the initial knowledge graph based on the scene attribute field in the target experience vector. This creation operation involves assigning a unique identifier to the new node, setting the node name, and initializing the node's attribute list, where the node name is usually directly taken from the scene attribute field content. After creation, the target experience vector is stored as the first piece of content under the new graph node, establishing the starting anchor point of the new knowledge branch. This mechanism gives the knowledge graph good scalability, enabling it to grow continuously with the accumulation of business data and automatically adapt to new technical scenarios or business forms.

[0050] As described in step S504 above, the graph node corresponding to the target experience vector with a high implicit experience level is directly connected to the root node in the initial knowledge graph, and the path weight of the direct connection is higher than the path weight between the graph node with a non-high implicit experience level and the root node. After the node matching and creation storage are completed, the graph topology is optimized based on the implicit experience level carried by the [target experience vector]. For target experiences marked as [high implicit experience level], the [graph node] where it is located is identified as a core teaching node, and a direct connection is established with the [root node] in the [initial knowledge graph]. When constructing this connection, the [path weight] is set differently, so that the connection weight value between the high implicit experience level node and the [root node] is significantly higher than the connection weight value between the non-high implicit experience level node and the [root node]. This high-weight direct connection path will be traversed first during graph retrieval or knowledge push, thereby ensuring that key experiences with high teaching value and the ability to correct typical errors and deviations are presented to learners by the retrieval system first. By quantifying the value of experience into weight parameters in the graph topology, this method constructs a knowledge navigation network oriented towards teaching effectiveness. It effectively solves the problem of insufficient guidance caused by the lack of dynamic evaluation mechanism in conventional association graphs, and improves the responsiveness and pertinence of knowledge transfer in actual training scenarios.

[0051] In one embodiment, step S6, which extracts and outputs the corresponding training experience content in the experience knowledge graph in response to the input of training information, includes: S601: Receive training information including fault codes, device type, or geographical location; S602: Using the fields in the training information as search keys, match and search the scene attribute fields in the experience knowledge graph; S603: Extract the structured empirical vectors stored under the matched graph nodes; S604: According to the field order of the scenario attribute field, action logic field, and result judgment field, the extracted structured experience vector is concatenated to generate training experience content in natural language format and output.

[0052] As described in step S601 above, training information containing fault codes, device types, or geographical locations is received. During the knowledge application phase, user-inputted training requirements are obtained through a pre-defined interactive interface. This training information is limited to structured or semi-structured query instructions containing specific business entities, specifically covering key fields such as fault codes, device types, or geographical locations. The fault code uniquely identifies a specific system anomaly or error type; for example, ERR-5001 represents high CPU usage in the core switch. The device type indicates the category of the faulty hardware or software component, such as optical transmission equipment or a database server. The geographical location specifies the physical area or data center location where the fault occurred, such as the Asia Pacific-Singapore data center. The receiving process supports multiple input methods, including receiving query requests sent from the front-end page via an application programming interface (API), or parsing the entity elements extracted from natural language query instructions, thereby ensuring that subsequent retrieval operations have a clear indexing basis.

[0053] As described in step S602 above, the fields in the training information are used as retrieval keys to match and retrieve the scene attribute fields in the experience knowledge graph. The obtained training information is parsed into independent field units, and these field units are directly used as retrieval keys. The retrieval operation aims to locate graph nodes in the experience knowledge graph that highly match the current query scenario. Since each graph node in the experience knowledge graph is associated with a scene attribute field, this field is a key label extracted from the original experience text during the knowledge construction phase and used to characterize the applicable background of the experience, such as a combination of fault code + equipment type + geographical location information. During the matching retrieval process, the similarity between the retrieval key and the scene attribute field of the graph node is calculated, or an exact matching logic is used to filter out nodes that meet the conditions. This retrieval mechanism utilizes the structured indexing capabilities of the knowledge graph. Compared with keyword-based full-text retrieval, it can more accurately utilize the semantic association of the scenario, filter out redundant experience that does not have a direct relevance, and thus improve the accuracy of the retrieval results.

[0054] As described in step S603 above, the structured experience vectors stored under the matched graph nodes are extracted. Once a match is successful and the target graph node is locked, the specific knowledge content stored within that node is further read. The knowledge stored under the graph nodes exists in the form of structured experience vectors, which are data generated in the preprocessing stage and contain core elements of the experience. The extraction process involves calling the query interface of the storage engine or graph database to load the data records pointed to by the graph nodes into memory. The extracted structured experience vectors organize data according to a preset pattern, preserving at least the complete information of the three dimensions: scene attributes, action logic, and result judgment. This step ensures that the data source upon which the subsequent content is generated is high-quality experience that has undergone structured processing and has clear semantic boundaries, rather than raw unstructured text fragments, laying a data foundation for generating logically coherent training content.

[0055] As described in step S604 above, the extracted structured experience vectors are concatenated to generate training experience content in natural language format and output, following the field order of scenario attribute field, action logic field, and result judgment field. After obtaining the structured experience vectors, a conversion operation to natural language text is performed. This process concatenates the data according to a strict field order, specifically scenario attribute field, action logic field, and result judgment field. First, the scenario attribute field is converted into a declarative sentence describing the problem background, such as "when the core switch CPU utilization rate continuously exceeds 90%"; then, the action logic field is connected to convert it into an imperative or declarative sentence describing the operation steps or decision basis, such as "prioritize restarting the management service rather than resetting the entire machine"; finally, the result judgment field is connected to convert it into a declarative sentence describing the processing effectiveness, such as "the business returned to normal within 2 minutes". Through this pipeline-style concatenation, the originally discrete vector data is reorganized into complete paragraphs that conform to human narrative logic and causal chains. The resulting natural language format training experience content can be directly output to a display terminal or document generator to form an easy-to-understand troubleshooting guide, thereby achieving an efficient mapping from machine-stored structured knowledge to human-usable training text.

[0056] Reference Figure 3 The present invention also provides a training experience content acquisition device based on a large model, the device comprising: The detection module 902 is used to perform semantic-level detection on experiential text data using a large language model, and extract multiple candidate text fragments with experiential descriptive features. The parsing module 904 is used to perform structured parsing on the candidate text fragments and generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; The evaluation module 906 is used to evaluate each of the structured experience vectors according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each of the structured experience vectors. The marking module 908 is used to mark the structured experience vector whose implicit experience level is greater than the preset experience level as the target experience vector; Adding module 910 is used to add each of the target experience vectors to a preset initial knowledge graph to obtain an experience knowledge graph; The output module 912 is used to extract and output the corresponding training experience content in the experience knowledge graph in response to the input of training information.

[0057] In one embodiment, the detection module 902 includes: The subject identity information recognition submodule is used to identify the identity information of the dialogue subjects in the experience text data; The causal logic word density statistics submodule is used to count the density of causal logic words in the speech of the main dialogue subject; The joint discriminative feature input submodule is used to input the identity information of the dialogue subject and the causal logic word density as joint discriminative features into the large language model to determine whether each text paragraph contains the intention to convey experience. The text paragraph extraction submodule is used to extract text paragraphs that are determined to contain the intention of conveying experience as candidate text segments.

[0058] In one embodiment, the training experience content acquisition device based on a large model further includes: The constraint setting module is used to set the minimum semantic unit constraint conditions, which include: the text length is not less than a preset number of characters, and it contains at least one action verb and at least one conditional adverb. The text fragment filtering module is used to filter each candidate text fragment according to the minimum semantic unit constraint condition, and retain the candidate text fragments that satisfy the minimum semantic unit constraint condition.

[0059] In one embodiment, the parsing module 904 includes: The framework construction submodule is used to build a standardized ontology framework; wherein, the standardized ontology framework defines four types of entities, namely scene entities, condition entities, action entities and result entities, and their relationships. The field information mapping submodule is used to map the field information of each field in the candidate text fragment to the standardized ontology framework, and uses One-Hot encoding combined with TF-IDF weighting to generate symbolic vectors. The candidate text fragment input submodule is used to input the candidate text fragments into the large language model and extract and generate semantically dense vectors; The weighted fusion submodule is used to weight and fuse the symbolic vector and the semantically dense vector according to a preset weight ratio to generate the structured experience vector.

[0060] In one embodiment, the evaluation module 906 includes: The structured experience vector input submodule is used to input the structured experience vector into the large language model and set the cognitive level constraint parameters; The processing execution submodule is used to drive the large language model to simulate the processing corresponding to the structured experience vector under the cognitive level constraint parameters, and generate a simulation result vector; The ideal structured vector acquisition submodule is used to acquire pre-stored ideal structured vectors; The comprehensive difference calculation submodule is used to calculate the comprehensive difference between the simulation result vector and the ideal structured vector; The implicit experience level determination submodule is used to determine the implicit experience level corresponding to the structured experience vector based on the comprehensive difference degree, wherein the comprehensive difference degree is positively correlated with the implicit experience level.

[0061] In one embodiment, module 910 is added, including: The matching submodule is used to match the scene attribute fields in each of the target experience vectors with the existing graph nodes in the initial knowledge graph; The first processing submodule is used to store the target experience vector as newly added attribute information under the matched graph node when the match is successful. The second processing submodule is used to create a new graph node in the initial knowledge graph based on the scene attribute field of the target experience vector when the matching fails, and to store the target experience vector under the new graph node. The direct connection establishment submodule is used to establish a direct connection between the graph node corresponding to the target experience vector and the root node in the initial knowledge graph, and to make the path weight of the direct connection higher than the path weight between the graph node corresponding to the non-high implicit experience level and the root node; wherein the non-high implicit experience level is an implicit experience level that is less than or equal to a preset experience level.

[0062] In one embodiment, the output module 912 includes: The training information receiving submodule is used to receive training information including fault codes, equipment types, or geographical locations. The matching and retrieval submodule is used to match and retrieve the scene attribute fields in the experience knowledge graph by using the fields in the training information as retrieval keys. The structured empirical vector extraction submodule is used to extract the structured empirical vectors stored under the retrieved and matched graph nodes; The training experience content output submodule is used to concatenate the extracted structured experience vectors according to the field order of the scenario attribute field, action logic field, and result judgment field to generate training experience content in natural language format and output it.

[0063] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a method for acquiring training experience content based on a large model. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to execute a method for acquiring training experience content based on a large model. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0064] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Using a large language model, semantic-level detection is performed on experiential text data to extract multiple candidate text fragments with experiential descriptive features; The candidate text fragments are subjected to structured parsing to generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; The structured experience vectors are evaluated according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each structured experience vector. The structured experience vector whose implicit experience level is greater than the preset experience level is denoted as the target experience vector; Each of the target experience vectors is added to a preset initial knowledge graph to obtain an experience knowledge graph; In response to the input of training information, the corresponding training experience content in the experience knowledge graph is extracted and output.

[0065] This enables the system to automatically identify and quantify the teaching value of experiences, avoid interference from low-quality or redundant information, improve the relevance and educational effectiveness of knowledge graph construction, enhance the readability and logical consistency of experience entries, ensure the traceability of the knowledge evolution process, and improve the coverage completeness and organizational participation of key experiences. Ultimately, it achieves the automated generation of scenario-based training content from unstructured logs.

[0066] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps: Using a large language model, semantic-level detection is performed on experiential text data to extract multiple candidate text fragments with experiential descriptive features; The candidate text fragments are subjected to structured parsing to generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; The structured experience vectors are evaluated according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each structured experience vector. The structured experience vector whose implicit experience level is greater than the preset experience level is denoted as the target experience vector; Each of the target experience vectors is added to a preset initial knowledge graph to obtain an experience knowledge graph; In response to the input of training information, the corresponding training experience content in the experience knowledge graph is extracted and output.

[0067] This enables the system to automatically identify and quantify the teaching value of experiences, avoid interference from low-quality or redundant information, improve the relevance and educational effectiveness of knowledge graph construction, enhance the readability and logical consistency of experience entries, ensure the traceability of the knowledge evolution process, and improve the coverage completeness and organizational participation of key experiences. Ultimately, it achieves the automated generation of scenario-based training content from unstructured logs.

[0068] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for acquiring training experience content based on a large model, characterized in that, The method includes: Using a large language model, semantic-level detection is performed on experiential text data to extract multiple candidate text fragments with experiential descriptive features; The candidate text fragments are subjected to structured parsing to generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; The structured experience vectors are evaluated according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each structured experience vector. Specifically, the structured experience vectors are input into a large language model, and cognitive level constraint parameters are set. Under the cognitive level constraint parameters, the large language model is driven to simulate the processing corresponding to the structured experience vector to generate a simulation result vector. A pre-stored ideal structured vector is obtained. The comprehensive difference between the simulation result vector and the ideal structured vector is calculated. The implicit experience level corresponding to the structured experience vector is determined according to the comprehensive difference, wherein the comprehensive difference is positively correlated with the implicit experience level. The structured experience vector whose implicit experience level is greater than the preset experience level is denoted as the target experience vector; Each of the target experience vectors is added to a preset initial knowledge graph to obtain an experience knowledge graph; In response to the input of training information, the corresponding training experience content in the experience knowledge graph is extracted and output.

2. The method for acquiring training experience content based on a large model according to claim 1, characterized in that, The step of using a large language model to perform semantic-level detection on experiential text data and extracting multiple candidate text fragments with experiential descriptive features includes: Identify the identity information of the dialogue subjects in the experiential text data; Analyze the density of causal logic words in the statements of the main participants in the dialogue; The identity information of the dialogue subject and the causal logic word density are used as joint discriminative features and input into the large language model to determine whether each text paragraph contains the intention to convey experience. Text segments deemed to contain the intention to convey experience are extracted as candidate text segments.

3. The method for acquiring training experience content based on a large model according to claim 2, characterized in that, After the step of extracting text paragraphs determined to contain the intention to convey experience as candidate text fragments, the method further includes: Set minimum semantic unit constraints, which include: the text length is not less than a preset number of characters, and it contains at least one action verb and at least one conditional adverb; Based on the minimum semantic unit constraint, each candidate text segment is filtered, and candidate text segments that satisfy the minimum semantic unit constraint are retained.

4. The method for acquiring training experience content based on a large model according to claim 1, characterized in that, The step of performing structured parsing on the candidate text fragments to generate corresponding structured empirical vectors includes: A standardized ontology framework is constructed; wherein, the standardized ontology framework defines four types of entities and their relationships: scene entities, condition entities, action entities, and result entities. The information of each field in the candidate text fragment is mapped to the standardized ontology framework, and symbolic vectors are generated by One-Hot encoding combined with TF-IDF weighting. The candidate text fragments are input into a large language model to extract and generate semantically dense vectors. The symbolic vector and the semantically dense vector are weighted and fused according to a preset weight ratio to generate the structured experience vector.

5. The method for acquiring training experience content based on a large model according to claim 1, characterized in that, The step of adding each of the target experience vectors to a preset initial knowledge graph to obtain an experience knowledge graph includes: Match the scene attribute fields in each of the target experience vectors with the existing graph nodes in the initial knowledge graph; When a match is successful, the target experience vector is stored as newly added attribute information under the matched graph node; When a match fails, a new graph node is created in the initial knowledge graph based on the scene attribute field of the target experience vector, and the target experience vector is stored under the new graph node. Establish a direct connection between the graph node corresponding to the target experience vector and the root node in the initial knowledge graph, and make the path weight of the direct connection higher than the path weight between the graph node corresponding to the non-high implicit experience level and the root node; wherein the non-high implicit experience level is an implicit experience level that is less than or equal to a preset experience level.

6. The method for acquiring training experience content based on a large model according to claim 1, characterized in that, The step of extracting and outputting the corresponding training experience content from the experience knowledge graph in response to the input of training information includes: Receive training information including fault codes, equipment type, or geographical location; Using the fields in the training information as search keys, the scene attribute fields in the experience knowledge graph are matched and searched. Extract the structured empirical vectors stored under the matched graph nodes; According to the field order of the scenario attribute field, action logic field, and result judgment field, the extracted structured experience vectors are concatenated to generate training experience content in natural language format and output.

7. A device for acquiring training experience content based on a large model, characterized in that, The device includes: The detection module is used to perform semantic-level detection on experiential text data using a large language model, and extract multiple candidate text fragments with experiential descriptive features. The parsing module is used to perform structured parsing on the candidate text fragments and generate corresponding structured experience vectors; wherein, the structured experience vectors include at least scene attribute fields, action logic fields, and result judgment fields; An evaluation module is used to evaluate each of the structured experience vectors according to a preset utility evaluation strategy to obtain the implicit experience level corresponding to each structured experience vector. Specifically, the module involves: inputting the structured experience vectors into a large language model and setting cognitive level constraint parameters; driving the large language model to simulate the processing corresponding to the structured experience vector under the cognitive level constraint parameters, generating a simulation result vector; obtaining a pre-stored ideal structured vector; calculating the comprehensive difference between the simulation result vector and the ideal structured vector; and determining the implicit experience level corresponding to the structured experience vector based on the comprehensive difference, wherein the comprehensive difference is positively correlated with the implicit experience level. A tagging module is used to mark structured experience vectors whose implicit experience level is greater than a preset experience level as target experience vectors; An add module is used to add each of the target experience vectors to a preset initial knowledge graph to obtain an experience knowledge graph; The output module is used to extract and output the corresponding training experience content in the experience knowledge graph in response to the input of training information.

8. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the method for acquiring training experience content based on a large model as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method for acquiring training experience content based on a large model as described in any one of claims 1 to 6.

Images