A large model-based marketing knowledge base construction system and method

Abstract

The application provides a marketing knowledge base construction system and method based on a large model, first determines the knowledge construction demand of the automobile marketing field and generates a marketing knowledge demand description, inputs the marketing knowledge demand description into a pre-trained large model to generate candidate knowledge content, then evaluates the candidate knowledge content according to a preset knowledge quality evaluation standard to generate an evaluation result, inputs the marketing knowledge demand description, the candidate knowledge content and the evaluation result into the large model for instruction fine-tuning to obtain optimized knowledge content, finally structures and organizes the optimized knowledge content, generates an automobile marketing knowledge unit and adds the automobile marketing knowledge unit to the knowledge base, so that the construction efficiency and quality of the automobile marketing knowledge base can be effectively improved, and a systematic, accurate and practical knowledge base is provided for automobile marketing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large models, in particular to a marketing knowledge base construction system and method based on a large model. BACKGROUND

[0002] In the increasingly competitive automobile industry, automobile marketing plays a crucial role in improving product sales and brand influence. Traditional automobile marketing knowledge acquisition and management methods have many limitations. On the one hand, marketing personnel often rely on their own experience or scattered materials to acquire marketing knowledge, which lacks systematicness and comprehensiveness and is difficult to cover all aspects of automobile marketing, such as precise positioning of different vehicle models, development of diversified marketing strategies, and communication skills for different customer groups. On the other hand, existing knowledge management systems are mostly based on fixed rules and templates, lacking flexibility and intelligence, and unable to update and optimize knowledge content in a timely manner according to changing marketing needs and market dynamics. In addition, there is a lack of effective evaluation and optimization mechanism for the quality of generated marketing knowledge, which may result in inaccurate, incomplete or unsuitable knowledge content, thereby affecting the effectiveness and efficiency of automobile marketing. SUMMARY

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a marketing knowledge base construction method based on a large model, which comprises:

[0004] determining the knowledge construction needs in the field of automobile marketing, generating a marketing knowledge demand description, the marketing knowledge demand description containing marketing theme range, knowledge content type and knowledge application scene limitation;

[0005] inputting the marketing knowledge demand description into a pre-trained large model, calling the text generation interface of the large model to perform knowledge content preliminary generation processing, and obtaining candidate knowledge content corresponding to the marketing knowledge demand description;

[0006] acquiring a preset knowledge quality evaluation standard, performing knowledge quality evaluation processing on the candidate knowledge content according to the knowledge quality evaluation standard, and generating a knowledge quality evaluation result;

[0007] inputting the marketing knowledge demand description, the candidate knowledge content and the knowledge quality evaluation result into the large model together, calling the instruction fine-tuning interface of the large model to perform knowledge content optimization processing, and obtaining optimized knowledge content meeting the knowledge quality evaluation standard;

[0008] performing knowledge structured organization processing on the optimized knowledge content, generating automobile marketing knowledge units containing knowledge theme identification, content hierarchical relationship and associated index information, and adding the automobile marketing knowledge units to the automobile marketing knowledge base.

[0009] In still another aspect, the embodiment of the present application also provides a marketing knowledge base construction system based on a large model, comprising a processor, a machine readable storage medium, the machine readable storage medium being connected with the processor, the machine readable storage medium being used for storing programs, instructions or codes, and the processor being used for executing the programs, instructions or codes in the machine readable storage medium to realize the above-mentioned method.

[0010] Based on the above aspects, the embodiment of the present application determines the knowledge construction requirements in the field of automobile marketing and generates a detailed description, uses a text generation interface of a pre-trained large model to preliminarily generate candidate knowledge content, fully utilizes the powerful language understanding and generation capability of the large model, can quickly obtain rich and diverse knowledge information, and preset knowledge quality evaluation standards are used to evaluate the candidate knowledge content, so that the quality problems of the knowledge content can be accurately identified. The marketing knowledge requirement description, the candidate knowledge content and the knowledge quality evaluation result are jointly input into the large model for instruction fine-tuning, the targeted optimization of the knowledge content is realized, the optimized knowledge content meeting the quality standards is obtained, the accuracy and practicality of the knowledge are improved, and finally the optimized knowledge content is structured and organized to generate automobile marketing knowledge units and add them to the knowledge base, so that the knowledge management is more systematic and orderly, the marketing personnel can quickly search and use, and the construction efficiency and quality of the automobile marketing knowledge base are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is an execution flow diagram of the marketing knowledge base construction method based on a large model provided by the embodiment of the present application.

[0012] Figure 2 is a schematic diagram of exemplary hardware and software components of the marketing knowledge base construction system based on a large model provided by the embodiment of the present application. DETAILED DESCRIPTION

[0013] The present application will be specifically described below in combination with the drawings of the specification, Figure 1 is a flow diagram of the marketing knowledge base construction method based on a large model provided by an embodiment of the present application, and the marketing knowledge base construction method based on a large model will be described in detail below.

[0014] Step S110: Determine the knowledge construction requirements in the field of automobile marketing, generate a marketing knowledge requirement description, and the marketing knowledge requirement description contains a marketing theme range, a knowledge content type and a knowledge application scene limitation.

[0015] The embodiment focuses on the construction demand determination of "battery endurance and charging facility correlation knowledge" in the new energy vehicle marketing field. The vehicle marketing business department finds in daily marketing activities that customers pay great attention to the battery endurance of new energy vehicles and the distribution of surrounding charging facilities, and the existing knowledge reserve cannot fully answer customer questions, so the demand for constructing related knowledge base is put forward.

[0016] Step S111: Obtain the preliminary knowledge demand document provided by the vehicle marketing business department, perform text analysis processing on the preliminary knowledge demand document, extract the content type indication information therein, and the content type indication information includes demand keywords, theme phrases and scene description paragraphs.

[0017] The preliminary knowledge demand document provided by the business department records the general demand for "battery endurance and charging facility correlation knowledge" in natural language form. A rule-based text analysis method is used to process the document. First, the text is divided into sentence units by punctuation marks, and then each sentence unit is tagged with part of speech.

[0018] For the extraction of demand keywords, a pre-set domain word table is used for matching, which contains words such as "new energy vehicle", "battery endurance", "charging facility", "endurance mileage", "charging pile", "fast charging", "slow charging", etc. These demand keywords are identified and extracted from the sentence units.

[0019] For the extraction of theme phrases, the subject-predicate-object structure of the sentence is analyzed to identify phrases formed by multiple words that express complete theme meanings, such as "battery endurance influencing factors", "charging facility coverage", "charging strategies corresponding to different endurance mileages", etc.

[0020] The extraction of scene description paragraphs is through the identification of continuous sentence sets containing time, place, behavior and other elements, such as "At the new energy vehicle promotion meeting for family users, the daily use scenarios corresponding to different battery endurance mileages and the accessibility of surrounding charging facilities need to be explained in detail" "For customers who often travel long distances, relevant knowledge of battery endurance planning based on their common routes and the distribution of charging facilities along the way is needed".

[0021] Step S112: Perform word frequency statistics and semantic clustering processing on the extracted demand keywords to identify core demand themes and secondary demand themes, and generate a demand theme hierarchical structure.

[0022] Word frequency statistics are performed on the extracted demand keywords. The statistical process does not involve specific numerical values, but only compares the frequency of each keyword appearing in the document to determine its importance. It is found through statistics that "battery endurance" and "charging facility" have relatively high frequencies.

[0023] Subsequently, semantic clustering processing is performed, and a clustering algorithm based on semantic similarity is adopted. First, each requirement keyword is converted into a vector form, and the dimension of the vector is determined by the semantic features of the vocabulary. Different vocabularies correspond to different combinations of semantic features. The semantic similarity between any two keyword vectors is calculated, and the semantic similarity is embodied by the correlation degree between the vectors. The higher the correlation degree, the more similar the semantics.

[0024] According to the semantic similarity, the keywords are divided into different clustering groups. For example, "battery endurance", "endurance mileage" and "endurance capability" are clustered into one group, and "charging facilities", "charging piles", "fast charging stations" and "slow charging stations" are clustered into another group.

[0025] Based on the clustering results, the core requirement theme is identified as "the correlation between battery endurance and charging facilities and applications", and the secondary requirement themes include "factors affecting battery endurance", "types and characteristics of charging facilities", and "differences in endurance and charging needs of different user groups".

[0026] When generating the requirement theme hierarchical structure, the core requirement theme is taken as the top-level node, and each secondary requirement theme is taken as a child node of the core requirement theme. Each secondary requirement theme can be further divided into a lower-level node according to its specific content, forming a hierarchical tree structure.

[0027] Step S113: performing scene feature extraction processing on the extracted scene description paragraph to determine the scene feature parameters of knowledge application, wherein the scene feature parameters include specific business scenarios, target user groups and expected application methods.

[0028] The scene feature extraction is performed on the extracted scene description paragraph. First, the paragraph is segmented and semantically understood to identify the scene-related feature information therein.

[0029] The determination of the specific business scenario is performed by extracting the activity type and occasion described in the paragraph, such as "family user promotion meeting", "long-distance travel customer consultation" and "4S store daily reception".

[0030] The identification of the target user group is performed according to the description of the user features in the paragraph, such as "family users", "customers who often travel long distances" and "city commuters". These user groups have different requirements for battery endurance and charging facilities.

[0031] The determination of the expected application method is performed by analyzing how the knowledge will be used, such as "sales personnel explaining to customers orally", "textual description in marketing materials", and "automatic reply on online consultation platform".

[0032] These extracted information is integrated into scene feature parameters, and each parameter contains multiple specific description items, which together constitute the complete features of the knowledge application scene.

[0033] Step S114: integrating the demand topic hierarchy, the content type indication information, and the scene characteristic parameters to generate a preliminary marketing knowledge demand description.

[0034] The demand topic hierarchy, the content type indication information, and the scene characteristic parameters are integrated. During the integration process, the logical consistency between the information is ensured.

[0035] Taking the demand topic hierarchy as a framework, the demand keywords and topic phrases in the content type indication information are corresponded to the corresponding topic nodes, and then the requirements of different scenes in the scene characteristic parameters are combined to determine the knowledge content that needs to be covered in each scene.

[0036] For example, in the "family user promotion meeting" scene, in combination with the core demand topic and the secondary demand topic, it is determined that the knowledge content such as the battery endurance demand of family users in daily travel and the distribution of charging facilities in the community and the surrounding area needs to be included.

[0037] Through the above integration process, a preliminary marketing knowledge demand description is formed. The marketing knowledge demand description is in the form of natural language and comprehensively covers the topic range, content type, and application scene limitation of knowledge construction.

[0038] Step S115: feeding back the preliminary marketing knowledge demand description to the automobile marketing business department for demand confirmation, receiving the demand modification opinions returned by the business department, iteratively adjusting the preliminary marketing knowledge demand description according to the demand modification opinions, and generating a final marketing knowledge demand description.

[0039] The preliminary marketing knowledge demand description is fed back to the automobile marketing business department, and the business department organizes relevant personnel to audit it. During the auditing process, the business personnel may propose modification opinions such as "in the long-distance travel scene, the influence of different weather conditions on battery endurance and the corresponding charging facility coping strategies need to be supplemented" and "the association analysis between charging facility use cost and battery endurance needs to be added".

[0040] After receiving these modification opinions, the preliminary marketing knowledge demand description is adjusted point by point. For the newly added content requirements, they are integrated into the corresponding topics and scenes; for the parts that are not accurate or complete, they are modified and supplemented.

[0041] After multiple iterative adjustments, until the business department confirms that the description accurately reflects the knowledge construction demand, the final marketing knowledge demand description is generated.

[0042] Step S120: input the marketing knowledge requirement description into a pre-trained large model, call the text generation interface of the large model to perform knowledge content preliminary generation processing, and obtain candidate knowledge content corresponding to the marketing knowledge requirement description.

[0043] The finally generated marketing knowledge requirement description is input into a pre-trained large model, which is trained on a large amount of text data and has strong text understanding and generation capabilities. The text generation interface of the large model is called, which receives input in the form of text and outputs corresponding text content according to the preset generation logic.

[0044] When performing knowledge content preliminary generation processing, the large model will deeply understand the marketing knowledge requirement description, combine the relevant knowledge stored in it, and generate knowledge content corresponding to the requirement description, i.e. candidate knowledge content. The candidate knowledge content covers various aspects related to battery endurance and charging facilities and is presented in the form of natural language text.

[0045] Step S121: requirement feature vectorization processing is performed on the marketing knowledge requirement description, and the text form requirement description is converted into a requirement feature vector that meets the input format of the large model.

[0046] The requirement feature vectorization processing is performed on the marketing knowledge requirement description, first, the text is preprocessed, including removing stop words, performing morphological reduction, etc. Stop words are those that frequently appear in text but have little effect on semantic expression, such as "of", "in", "and", etc.

[0047] Then, the word embedding technology is used to convert each word in the processed text into a vector form, and the dimension of each word vector is determined by the parameters of the word embedding model. Different words correspond to different vector representations.

[0048] The vectors of all words in the text are arranged in the order of their appearance in the text to form a matrix, and then the matrix is processed through pooling operation to obtain a fixed dimension vector, i.e. the requirement feature vector. The requirement feature vector can reflect the semantic information of the marketing knowledge requirement description and meet the requirements of the large model for input format.

[0049] Step S122: call the domain knowledge awakening module of the large model, associate and match the requirement feature vector with the pre-trained knowledge graph in the field of automobile marketing, and activate the domain knowledge parameters in the large model related to the marketing knowledge requirement description.

[0050] The domain knowledge awakening module of the large model is called, and the domain knowledge awakening module stores a pre-training knowledge graph in the field of automobile marketing. The knowledge graph is composed of entity nodes and relationship edges, and the entity nodes include "new energy vehicles", "batteries", "charging piles", etc., and the relationship edges represent the association between entities, such as "batteries contain the attribute of endurance mileage" and "charging piles provide charging services for new energy vehicles".

[0051] The demand feature vector is associated and matched with the entity node vector in the knowledge graph, and the association degree between the demand feature vector and each entity node vector is calculated, which is embodied by the semantic association degree between the vectors.

[0052] According to the degree of association, the domain knowledge parameters related to the marketing knowledge demand description in the large model are activated, and these parameters correspond to the related entity and relationship information in the knowledge graph, so that the large model can call more knowledge related to the field when generating knowledge content.

[0053] Step S123: Configure the generation parameter set of the text generation interface, which includes the theme relevance weight, the content depth coefficient, the professional term density threshold, and the output length range.

[0054] The generation parameter set of the text generation interface is configured, and the theme relevance weight is used to control the association between the generated content and the theme of the marketing knowledge demand description. The higher the weight setting, the stronger the relevance between the generated content and the theme.

[0055] The content depth coefficient determines the detail and depth of the generated knowledge content. The larger the coefficient, the more detailed and in-depth the content, which can cover more details and the underlying principles.

[0056] The professional term density threshold is used to limit the proportion of professional terms in the generated text, which is set according to different knowledge application scenarios. For example, in scenarios targeting ordinary customers, the threshold is set low and fewer professional terms are used; in scenarios targeting professionals, the threshold can be set high.

[0057] The output length range specifies the approximate length interval of the generated text, ensuring that the generated knowledge content is neither too short to be incomplete nor too long to be cumbersome.

[0058] These parameters are set through the configuration interface of the text generation interface, forming a generation parameter set and passing it to the interface.

[0059] Step S124: The demand feature vector and the generation parameter set are input into the text generation interface, triggering the large model to perform knowledge content generation operations, and obtaining the initial knowledge text.

[0060] The demand feature vector and the generated parameter set are input into a text generation interface, which parses and processes the input information and converts it into an internal format that can be understood by the large model.

[0061] After receiving this information, the large model performs knowledge content generation operations according to the requirements of the generated parameter set, combined with the activated domain knowledge parameters. During the generation process, the large model organizes language generation and topic-related content based on the semantic information of the demand feature vector, controls the level of detail of the content by referring to the content depth coefficient, uses an appropriate number of professional terms based on the professional term density threshold, and completes text generation within the output length range, ultimately obtaining the initial knowledge text.

[0062] Step S125: Perform redundant information filtering processing on the initial knowledge text, delete repetitive expression content, irrelevant expansion content, and contradictory content with expression conflicts, and obtain the purified intermediate knowledge text.

[0063] During the redundant information filtering processing of the initial knowledge text, first, a text similarity calculation method is used to identify repetitive expression content, and sentences or paragraphs with high semantic similarity are determined as repetitive content, and only one of them is retained.

[0064] Then, by comparing with the demand feature vector, expansion content unrelated to the marketing knowledge demand description is identified. These contents may be related to the relevant field, but they are beyond the scope of this knowledge construction, and will be deleted.

[0065] For contradictory content with expression conflicts, logical analysis is used to identify, for example, for the same question, there are two completely opposite statements, at this time, the domain knowledge and demand scenario need to be combined for judgment, and the content that does not conform to the logic or does not match the demand is deleted.

[0066] After the above filtering processing, the purified intermediate knowledge text is obtained, which retains the core information while removing redundant and conflicting content.

[0067] Step S126: Perform content integrity check on the intermediate knowledge text, if there are obvious missing paragraphs, mark the missing paragraphs and the original demand feature vector, and re-input them into the text generation interface for supplementary generation, and splice the supplementary generated content with the intermediate knowledge text to obtain the candidate knowledge content.

[0068] The content integrity check of the intermediate knowledge text checks whether all necessary content is covered in the intermediate knowledge text according to the theme range and content type specified in the marketing knowledge demand description.

[0069] If it is found that there is a significant content missing paragraph, such as not involving the topic of "the impact of charging speed of different charging facilities on battery endurance use", the missing paragraph is identified.

[0070] The missing paragraph and the original requirement feature vector are re-input into the text generation interface, which triggers the large model to perform supplementary generation according to the previous generation parameter set requirements, and generates content related to the missing paragraph.

[0071] The supplementary generated content is spliced and fused with the intermediate knowledge text. The arrangement is arranged according to the logical order of the content to ensure that the fused text is smooth and coherent, and finally the candidate knowledge content is obtained.

[0072] Step S130: Obtain a preset knowledge quality evaluation standard, and perform knowledge quality evaluation processing on the candidate knowledge content according to the knowledge quality evaluation standard to generate a knowledge quality evaluation result.

[0073] The preset knowledge quality evaluation standard is obtained, which is formulated according to the characteristics and application requirements of automobile marketing knowledge, and contains multiple evaluation dimensions and corresponding evaluation indexes.

[0074] According to the knowledge quality evaluation standard, the candidate knowledge content is comprehensively evaluated. In the evaluation process, the candidate knowledge content is checked and scored according to the indexes of each dimension, and finally the evaluation results of each dimension are summarized to generate a knowledge quality evaluation result.

[0075] Step S131: Retrieve the preset knowledge quality evaluation standard from the standard database of the knowledge management system, analyze the knowledge quality evaluation standard, and determine the specific evaluation indexes and index weights contained in the content completeness evaluation dimension, the expression accuracy evaluation dimension and the application adaptability evaluation dimension.

[0076] The preset knowledge quality evaluation standard is retrieved from the standard database of the knowledge management system, which is stored in the form of structured data. After analyzing it, the content completeness evaluation dimension, the expression accuracy evaluation dimension and the application adaptability evaluation dimension are extracted.

[0077] The specific evaluation indexes contained in the content completeness evaluation dimension include knowledge point coverage and the degree of elaboration of each knowledge point; the expression accuracy evaluation dimension includes the index of professional term use standardization, logical coherence and data accuracy; the application adaptability evaluation dimension includes the index of scene matching degree, user applicability and operability.

[0078] Each specific evaluation index has a corresponding index weight, which reflects the importance of the index in the evaluation dimension it belongs to, and the sum of all index weights is 1.

[0079] Step S132: For the content completeness evaluation dimension, knowledge point coverage detection is performed on the candidate knowledge content to identify whether the candidate knowledge content contains all the core knowledge points required in the marketing knowledge demand description, and the number of missing knowledge points and the elaboration sufficiency score of each knowledge point are counted.

[0080] For the content completeness evaluation dimension, first, all the core knowledge points required in the marketing knowledge demand description are determined, which are the key contents determined according to the demand theme hierarchical structure.

[0081] The knowledge point coverage detection is performed on the candidate knowledge content, and through text matching and semantic understanding, the knowledge points contained in the candidate knowledge content are identified and compared with the core knowledge points to determine the contained core knowledge points and the uncontained core knowledge points.

[0082] The number of missing knowledge points, i.e., the number of uncontained core knowledge points, is counted. At the same time, for each contained core knowledge point, the breadth and depth of its elaboration are evaluated to give an elaboration sufficiency score, and the score reflects the perfection degree of the knowledge point elaboration.

[0083] Step S1321: Core knowledge point extraction processing is performed on the marketing knowledge demand description, and combined with the automobile marketing field knowledge system, a core knowledge point set corresponding to the marketing knowledge demand description is determined, each core knowledge point containing a theme name, a necessary elaboration aspect and a detailed degree requirement.

[0084] Core knowledge point extraction processing is performed on the marketing knowledge demand description, combined with the automobile marketing field knowledge system, which contains various knowledge frameworks and classifications related to new energy vehicle marketing.

[0085] Through semantic analysis of the marketing knowledge demand description, the key concepts and core contents are identified, and these contents are matched with the knowledge units in the field knowledge system to determine the corresponding core knowledge points.

[0086] Each core knowledge point contains a theme name, such as "battery endurance mileage calculation method", a necessary elaboration aspect, such as "factors affecting endurance mileage" "endurance performance under different road conditions", and a detailed degree requirement, such as explaining the basic principle and providing actual cases. These constitute a core knowledge point set.

[0087] Step S1322: A knowledge point recognition model is constructed, which uses a bidirectional long short-term memory network combined with a conditional random field algorithm to perform segment-by-segment scanning processing on the candidate knowledge content to identify the contained knowledge points and the actual elaboration aspects of each knowledge point.

[0088] The knowledge point recognition model is constructed, and the knowledge point recognition model adopts a bidirectional long short-term memory network combined with a conditional random field algorithm. The bidirectional long short-term memory network is composed of long short-term memory layers in two directions of forward and backward, can process text sequences in two directions, and capture context information.

[0089] The conditional random field algorithm is used for sequence labeling of the output result of the bidirectional long short-term memory network, and determines which parts of the text belong to knowledge points and the boundaries of the knowledge points.

[0090] When training the model, the text data in the automobile marketing field annotated with knowledge points is used as a training set, and the model parameters are adjusted through a back propagation algorithm, so that the model can accurately recognize knowledge points.

[0091] When scanning the candidate knowledge content section by section, the text is divided into multiple paragraphs, and each paragraph is used as the input of the model. The model first performs word embedding processing on the paragraph, converts the words into vectors, and then inputs them into the bidirectional long short-term memory network. The network outputs the hidden state vector of each word.

[0092] The conditional random field algorithm calculates the probability of the label sequence according to these hidden state vectors, selects the label sequence with the highest probability as the output, and thus identifies the knowledge points contained in the candidate knowledge content and the actual aspects of each knowledge point.

[0093] Step S1323: Compare and match the identified contained knowledge points with the core knowledge point set to determine the uncontained core knowledge points and count the number of missing knowledge points.

[0094] Compare and match the contained knowledge points identified by the knowledge point recognition model with the core knowledge point set. The comparison is based on the topic name and the aspect of the knowledge point.

[0095] For each core knowledge point, check whether there is a knowledge point with the same topic name and related aspect in the contained knowledge points. If not, the core knowledge point is determined as an uncontained core knowledge point.

[0096] Count the number of all uncontained core knowledge points, that is, the number of missing knowledge points, which reflects the missing situation of the candidate knowledge content in the core knowledge point coverage.

[0097] Step S1324: For each contained core knowledge point, match its actual aspect with the necessary aspect and calculate the aspect coverage rate.

[0098] For each core knowledge point that is already included, match its actual explanations with its necessary explanations. First, list all the necessary explanations for that core knowledge point, then list the actual explanations, and finally count the number of actual explanations that match the necessary explanations.

[0099] The coverage rate of explanatory aspects is calculated by dividing the number of successfully matched actual explanatory aspects by the total number of necessary explanatory aspects. For example, if a core knowledge point has 5 necessary explanatory aspects, and 3 of the actual explanatory aspects match it, then the coverage rate is 3 divided by 5.

[0100] Step S1325: Based on the coverage of the described aspects and the level of detail of each described aspect, calculate the sufficiency score of each knowledge point according to the preset scoring rules. The sufficiency score is positively correlated with the coverage and level of detail of the described aspects.

[0101] The sufficiency score is calculated based on the coverage of the aspects explained and the level of detail in each aspect. In the pre-defined scoring rules, the coverage of the aspects explained accounts for a certain percentage of the score, while the level of detail in each aspect accounts for another percentage.

[0102] The level of detail is assessed based on the depth and breadth of information contained in each aspect of the explanation, such as whether it includes explanations of principles, case studies, and data support. The higher the level of detail, the higher the score.

[0103] The score for the coverage of the explanation is added to the score for the level of detail to obtain the sufficiency score for each knowledge point. The higher the sufficiency score, the more fully and comprehensively the knowledge point has been explained.

[0104] Step S1326: Record the number of missing knowledge points and the sufficiency score of each knowledge point in the content completeness assessment sub-result, as the basis for calculating the comprehensive content completeness score.

[0105] The number of missing knowledge points and the adequacy of explanation for each knowledge point are compiled and recorded in the content completeness assessment sub-result. The content completeness assessment sub-result is presented in a structured format, clearly showing the coverage and adequacy of explanation for each core knowledge point. By comprehensively considering the number of missing knowledge points and the adequacy of explanation for each knowledge point, the performance of candidate knowledge content in terms of content completeness is comprehensively evaluated.

[0106] Step S133: For the accuracy of expression assessment dimension, perform terminology standardization check, logical coherence check, and data accuracy check on the candidate knowledge content, and mark the use of incorrect professional terms, sentence paragraphs with logical contradictions, and content fragments with inaccurate data expression.

[0107] For the expression accuracy evaluation dimension, the candidate knowledge content is checked from three aspects of term use, logical structure and data information. Through professional checking methods and tools, the parts that do not meet the requirements are identified and marked.

[0108] Step S1331: The standard term set is called from the automobile marketing professional term library, and the standard term set includes term name, standard definition, correct use scene and common error usage example.

[0109] The standard term set is called from the automobile marketing professional term library, which is accumulated and professionally audited for a long time, and contains various types of standard professional terms in the field of automobile marketing.

[0110] Each term in the standard term set includes a term name such as "battery energy density", "charging power", etc.; a standard definition that clearly defines the accurate meaning of the term; a correct use scene that explains under what circumstances the use of the term is appropriate; and a common error usage example that shows the situation where the term is easily misused for comparison and checking.

[0111] Step S1332: The candidate knowledge content is subjected to term extraction, all professional terms appearing in the text are identified, the extracted professional terms are matched and compared with the standard term set, the spelling correctness, definition consistency and use scene adaptability of the terms are checked, and the error terms that do not meet the standard are marked.

[0112] The candidate knowledge content is subjected to term extraction, and the method based on dictionary matching and part-of-speech tagging is adopted to identify all professional terms in the text. The extracted professional terms are matched and compared with the standard term set one by one.

[0113] The spelling correctness of the term is checked to see if the extracted term is completely consistent with the standard term name, and if there is a wrong character or omission, multiple writing. The definition consistency is checked to analyze whether the meaning of the term in the candidate knowledge content is consistent with the standard definition. The use scene adaptability is checked to judge whether the use of the term in the current context is consistent with the correct use scene specified in the standard term set.

[0114] For the terms with spelling errors, inconsistent definitions or inappropriate use scenes, mark them as error terms and record the error type and specific location.

[0115] Step S1333: Perform sentence-level logical relationship analysis on the candidate knowledge content, use dependency syntax analysis method to identify the logical connection relationship between adjacent sentences, including cause and effect relationship, progressive relationship, transition relationship and parallel relationship, check whether there is improper use of logical connection words or logical relationship contradiction, and mark the sentence paragraphs with logical contradiction.

[0116] The candidate knowledge content is analyzed at the sentence level, and the dependency syntax analysis method is used to analyze the syntax structure of each sentence, identify the subject-predicate-object, state-supplement and other components in the sentence, and the dependency relationship between words.

[0117] Based on this, the logical connection relationship between adjacent sentences is analyzed to determine whether it is a cause and effect relationship (such as "because…so…"), a progressive relationship (such as "not only…but also…"), a transition relationship (such as "although…but…") or a parallel relationship (such as "at the same time…in addition…") and the like.

[0118] Check if the use of logical connection words is appropriate, if there is a mismatch between the connection words and the actual logical relationship, and if there is a logical contradiction between sentences, such as conflicting content, inverted cause and effect, etc. Mark the sentence paragraphs with logical contradiction.

[0119] Step S1334: Extract data expressions in the candidate knowledge content, compare the extracted data expressions with the preset authoritative data source of automobile marketing, check if the data values are accurate, the units are consistent, and the descriptions are objective, mark the content segments with inaccurate data expressions, the data expressions include numerical values, percentages, times, places and verifiable information of event descriptions.

[0120] Extract data expressions in the candidate knowledge content, which include numerical values, percentages, times, places and verifiable information of event descriptions. Compare the extracted data expressions with the preset authoritative data source of automobile marketing, which contains verified accurate data information.

[0121] Check if the data values are consistent with the corresponding data in the authoritative data source, if the data units are unified, if the descriptions are objective and true, if there is exaggeration, reduction or false description. Mark the content segments with inaccurate data expressions.

[0122] Step S1335: Classify and count the error term marking, logical contradiction paragraph marking and data error segment marking, calculate the term error rate, logical contradiction rate and data error proportion, which are the calculation basis of the comprehensive score of expression accuracy.

[0123] The error terms, logical contradiction paragraphs, and data error segments are classified and counted according to different error types. The number of error terms, the number of paragraphs with logical contradictions, and the number of data error segments are counted.

[0124] The term error rate, i.e., the ratio of the number of error terms to the total number of professional terms in the candidate knowledge content, is calculated. The logical contradiction occurrence rate, i.e., the ratio of the number of paragraphs with logical contradictions to the total number of paragraphs, is calculated. The data error rate, i.e., the ratio of the number of data error segments to the total number of data expression segments, is calculated.

[0125] These ratio data will be used as the basis for calculating the comprehensive score of expression accuracy, and will comprehensively reflect the problems of the candidate knowledge content in terms of expression accuracy.

[0126] Step S134: For the application adaptability evaluation dimension, the candidate knowledge content is matched with the knowledge application scenarios defined in the marketing knowledge demand description for matching degree analysis, to evaluate the support degree of the knowledge content for specific business scenarios, the applicability to target user groups, and the operability in actual marketing activities.

[0127] For the application adaptability evaluation dimension, the candidate knowledge content is matched with the knowledge application scenarios defined in the marketing knowledge demand description for multi-aspect matching degree analysis. From the perspectives of support of knowledge content for business scenarios, fit for user groups, and feasibility in actual marketing, etc., the candidate knowledge content is evaluated to determine whether it can meet the needs of the application scenarios.

[0128] Step S1341: The marketing knowledge demand description is parsed to extract scene characteristic parameters of the knowledge application scenarios, including business scenario type, target user group characteristics, marketing activity purpose, knowledge application method, and expected effect indicators.

[0129] The marketing knowledge demand description is parsed to extract scene characteristic parameters of the knowledge application scenarios through text analysis and semantic understanding. The business scenario type is as described above, such as "family user introduction meeting" and "long-distance travel customer consultation". The target user group characteristics include user travel habits, consumer ability, and cognitive level of new energy vehicles. The marketing activity purpose includes improving customer purchase willingness, answering customer questions, and enhancing brand image. The knowledge application method includes oral explanation, written material display, and online interaction. The expected effect indicators include customer satisfaction improvement and consultation conversion rate improvement.

[0130] Step S1342: Application scenario related information is extracted from the candidate knowledge content to identify the implied application scenario description, recommended use object, suggested application method, and expected achieved target in the knowledge content.

[0131] The application scenario association information of the candidate knowledge content is extracted, and information related to the application scenario is identified through semantic analysis and keyword matching.

[0132] The application scenario description refers to the specific occasion where the knowledge content is suitable for application; the recommended use object refers to the user group to which the knowledge content is more suitable for delivery; the suggested application mode refers to the propagation or display mode recommended by the knowledge content; and the expected target refers to the effect expected to be achieved by using the knowledge content.

[0133] The extracted information is sorted to form a set of application scenario association information of the candidate knowledge content.

[0134] Step S1343: Similarity calculation is performed between the application scenario description of the candidate knowledge content and the business scenario type in the scenario characteristic parameter, and a support degree score of the knowledge content for the specific business scenario is evaluated.

[0135] The similarity calculation is performed between the application scenario description of the candidate knowledge content and the business scenario type in the scenario characteristic parameter. Both are first converted into vector form, and the dimension of the vector is determined based on the characteristic elements of the scenario, such as activity type, participants, and environment atmosphere.

[0136] The similarity between the two vectors is calculated. The higher the similarity, the more matched the application scenario of the candidate knowledge content and the business scenario type, and the higher the support degree of the knowledge content for the specific business scenario. According to the size of the similarity, a corresponding support degree score is given.

[0137] Step S1344: Matching analysis is performed between the recommended use object of the candidate knowledge content and the target user group characteristics in the scenario characteristic parameter, and an applicability score of the knowledge content for the target user group is evaluated.

[0138] The matching analysis is performed between the recommended use object of the candidate knowledge content and the target user group characteristics in the scenario characteristic parameter. The matching degree of the characteristics of the recommended use object and the characteristics of the target user group, such as age, occupation, and travel demand, is compared.

[0139] The higher the matching degree, the more suitable the knowledge content for the target user group, and the higher the applicability score. By comprehensively considering the matching of multiple characteristics, an applicability score of the knowledge content for the target user group is given.

[0140] Step S1345: According to the matching degree of the suggested application mode and the expected target of the candidate knowledge content and the marketing activity purpose, the knowledge application mode and the expected effect index in the scenario characteristic parameter, an operability score of the knowledge content in the actual marketing activity is evaluated.

[0141] The analysis is whether the recommended application mode of the candidate knowledge content is consistent with the knowledge application mode in the scene characteristic parameter, and whether the expected target and the marketing activity purpose and the expected effect index are consistent.

[0142] If the recommended application mode is easy to implement in the actual marketing activity, and the expected target is highly consistent with the marketing activity purpose and the expected effect index, it means that the operability of the knowledge content is strong, and the operability score is high. Otherwise, the score is low. By comprehensively considering these factors, the operability score of the knowledge content in the actual marketing activity is evaluated.

[0143] Step S1346: The scene support degree score, the user applicability score and the operability score are weighted and calculated according to preset weights to obtain an application adaptability comprehensive score, which reflects the overall matching level of the candidate knowledge content and the application scene.

[0144] The scene support degree score, the user applicability score and the operability score are weighted and calculated according to preset weights. The preset weights are determined according to the importance of each score in the application adaptability evaluation, such as the scene support degree score weight is 0.4, the user applicability score weight is 0.3, and the operability score weight is 0.3.

[0145] The calculation method is to multiply each score by its corresponding weight, and then add the products to obtain the application adaptability comprehensive score. The application adaptability comprehensive score comprehensively reflects the overall matching level of the candidate knowledge content and the application scene.

[0146] Step S135: According to the specific evaluation index score and index weight of each evaluation dimension, the content completeness comprehensive score, the expression accuracy comprehensive score and the application adaptability comprehensive score are calculated by weighted summation.

[0147] For the content completeness evaluation dimension, the scores of the knowledge point coverage, the elaboration sufficiency of each knowledge point and other specific evaluation indexes are weighted and summed according to the corresponding index weights to obtain the content completeness comprehensive score.

[0148] Similarly, for the expression accuracy evaluation dimension, the scores of the professional term use standardization, the logical coherence, the data accuracy and other indexes are weighted and summed with the corresponding weights to obtain the expression accuracy comprehensive score.

[0149] For the application adaptability evaluation dimension, the application adaptability comprehensive score is calculated according to the method of step S1346.

[0150] The specific way of weighted summation is that each index score is multiplied by its weight and then added, and the sum of the weights of each index is 1, which ensures that the comprehensive score can reasonably reflect the overall performance of each dimension.

[0151] Step S136: Summarize and organize the comprehensive scores of each dimension, the scores of specific evaluation indicators, the list of missing knowledge points, the error marking information and the scenario matching analysis results to generate a knowledge quality assessment result that includes the scoring results, problem descriptions and improvement suggestions.

[0152] The scores for content completeness, accuracy of expression, and application suitability, along with the scores for specific evaluation indicators under each dimension, are summarized. Additionally, a list of missing knowledge points, markings of incorrect terminology, logically contradictory paragraphs, and data errors are compiled, along with issues discovered during the scenario matching analysis.

[0153] Based on this aggregated information, a problem description is provided, clearly identifying the deficiencies in the candidate knowledge content, and corresponding improvement suggestions are proposed, such as supplementing missing knowledge points, correcting erroneous terminology, and adjusting content to improve its relevance to the scenario.

[0154] These contents are integrated to generate a knowledge quality assessment result, which comprehensively reflects the quality status of the candidate knowledge content and the areas that need improvement.

[0155] Step S140: Input the marketing knowledge demand description, the candidate knowledge content and the knowledge quality assessment result into the large model, call the instruction fine-tuning interface of the large model to perform knowledge content optimization processing, and obtain optimized knowledge content that meets the knowledge quality assessment criteria.

[0156] The marketing knowledge demand description, candidate knowledge content, and knowledge quality assessment results are input into the large model. The large model's instruction fine-tuning interface is then invoked. This instruction fine-tuning interface can receive specific instructions and data to make targeted adjustments and optimizations to the model's output.

[0157] By performing knowledge content optimization processing, the large model modifies and improves candidate knowledge content based on the problems and suggestions in the knowledge quality assessment results, and finally obtains optimized knowledge content that meets the knowledge quality assessment standards.

[0158] Step S141: Perform instruction conversion processing on the knowledge quality assessment results, converting the problem descriptions and improvement suggestions into optimization instructions that conform to the large model instruction format. The optimization instructions include supplementary instructions for content completeness, correction instructions for expression accuracy, and adjustment instructions for application adaptability.

[0159] The problem descriptions and improvement suggestions in the knowledge quality assessment results are converted into instructions. Following an instruction format that the large model can understand, the problem descriptions are transformed into tasks that need to be solved, and the improvement suggestions are transformed into specific operational requirements.

[0160] The supplementary instructions for content completeness clearly indicate the missing knowledge points that need to be supplemented and the improvement requirements for the elaboration of each knowledge point; the correction instructions for expression accuracy indicate the error terms, logical contradiction paragraphs, and data error segments that need to be corrected; and the adjustment instructions for application adaptability propose how to adjust the content to improve the matching degree with the application scenario.

[0161] Step S142: Construct an optimization prompt word template, fill the marketing knowledge demand description, the candidate knowledge content, and the optimization instruction into the optimization prompt word template according to a preset text structure, and generate a comprehensive optimization prompt word containing demand background, current content, existing problems, and optimization direction.

[0162] An optimization prompt word template is constructed, which contains a fixed text structure, such as a demand background part, a current content part, an existing problem part, and an optimization direction part.

[0163] The marketing knowledge demand description is filled into the demand background part to enable the large model to understand the original demand for knowledge construction; the candidate knowledge content is filled into the current content part to show the basic content that needs to be optimized; and the optimization instruction is filled into the existing problem part and the optimization direction part to clearly indicate the problems in the content and the specific direction of optimization.

[0164] Through the above filling, a comprehensive optimization prompt word is generated, which provides the large model with comprehensive optimization context information.

[0165] Step S143: Configure the optimization parameters of the large model instruction fine-tuning interface, including the learning rate, the number of training rounds, the context window size, and the output content length limit.

[0166] The optimization parameters of the large model instruction fine-tuning interface are configured. The learning rate is used to control the magnitude of model parameter updates, and the size of the learning rate will affect the speed and effect of model optimization; the number of training rounds refers to the number of times the model trains data during the optimization process, and a suitable number of training rounds can ensure that the model learns the optimization instruction sufficiently; the context window size determines the range of context that the model can consider when processing text, and a larger window size helps the model understand longer text logic; and the output content length limit ensures that the optimized knowledge content is within a reasonable length range.

[0167] According to the characteristics and optimization requirements of the candidate knowledge content, appropriate optimization parameter values are set to achieve the best optimization effect.

[0168] Step S144: Input the comprehensive optimization prompt word and the optimization parameter into the instruction fine-tuning interface of the large model, trigger the large model to perform a knowledge content optimization operation, and obtain a preliminary optimized knowledge content.

[0169] The instruction fine-tuning interface of the large model analyzes and processes these input information and transmits them to the optimization module inside the large model.

[0170] The large model performs knowledge content optimization operation according to the information in the comprehensive optimization prompt word and in combination with the setting of the optimization parameter. In the optimization process, the model modifies, supplements and adjusts the candidate knowledge content according to the optimization direction in view of the existing problems to generate preliminary optimized knowledge content.

[0171] Step S145: Perform secondary quality evaluation on the preliminary optimized knowledge content to generate secondary knowledge quality evaluation result of the preliminary optimized knowledge content.

[0172] The same knowledge quality evaluation standard and evaluation method as in step S130 are adopted to perform secondary quality evaluation on the preliminary optimized knowledge content.

[0173] The evaluation process includes checking content completeness, expression accuracy and application adaptability, etc. to generate secondary knowledge quality evaluation result which is used to judge whether the preliminary optimized knowledge content meets the expected quality requirement.

[0174] Step S146: Compare the secondary knowledge quality evaluation result with the knowledge quality evaluation standard. If the secondary knowledge quality evaluation result meets the knowledge quality evaluation standard, the preliminary optimized knowledge content is determined as the optimized knowledge content. If the secondary knowledge quality evaluation result does not meet the knowledge quality evaluation standard, the secondary knowledge quality evaluation result is taken as new knowledge quality evaluation result and the optimization processing step is repeated until the optimized knowledge content meeting the knowledge quality evaluation standard is obtained.

[0175] Compare the secondary knowledge quality evaluation result with the knowledge quality evaluation standard to check whether the comprehensive score of each dimension reaches the threshold value specified in the standard.

[0176] If the standard is met, it means that the preliminary optimized knowledge content has met the requirement and is determined as the optimized knowledge content. If the standard is not met, the secondary knowledge quality evaluation result is taken as new knowledge quality evaluation result and the instruction conversion processing is performed again in step S141 to generate new optimization instruction. Subsequently, the comprehensive optimization prompt word is reconstructed, the optimization parameter is configured, the instruction fine-tuning interface is inputted to perform optimization operation and the quality evaluation is performed again according to the procedures of steps S142 to S145.

[0177] This cycle is repeated, and each cycle is targeted at the problems found in the previous evaluation, until the secondary knowledge quality evaluation result meets the knowledge quality evaluation standard, at which point the preliminary optimized knowledge content obtained is the final optimized knowledge content. For example, in the optimization process of the "battery endurance and charging facility associated knowledge", if it is found after the first optimization that the knowledge point "the influence of extreme weather on battery endurance and charging facility efficiency" is still missing, then this result is taken as the new evaluation result, the optimization step is re-executed, the relevant content is supplemented, and then evaluated again until all problems are solved.

[0178] Step S150: Perform knowledge structured organization processing on the optimized knowledge content to generate an automobile marketing knowledge unit containing knowledge topic identification, content hierarchical relationship, and association index information, and add the automobile marketing knowledge unit to the automobile marketing knowledge base.

[0179] The optimized knowledge content is subjected to knowledge structured organization processing, which converts the knowledge content originally in natural language form into a knowledge unit with fixed structure and association relationship through a series of processing steps. The above structured processing method facilitates knowledge storage, retrieval, and application, and finally the generated automobile marketing knowledge unit is added to the automobile marketing knowledge base to enrich the content of the knowledge base.

[0180] Step S151: Perform theme division processing on the optimized knowledge content, and divide the optimized knowledge content into multiple knowledge sub-modules according to the content theme, each knowledge sub-module focusing on a core theme.

[0181] The optimized knowledge content is subjected to theme division processing, which first reads through the entire optimized knowledge content and identifies different core themes involved therein. Taking the "battery endurance and charging facility associated knowledge" as an example, the following core themes can be divided: "basic principles of battery endurance", "factors affecting battery endurance", "types and characteristics of charging facilities", "matching strategies for battery endurance and charging facilities", etc.

[0182] According to these core themes, the optimized knowledge content is divided into multiple knowledge sub-modules, each focusing on a core theme and containing all knowledge content related to the theme, ensuring that the content of each sub-module has strong relevance and independence.

[0183] Step S152: Assign a unique knowledge topic identification to each knowledge sub-module, which uses an alphanumeric combination code containing a domain classification code, a theme category code, and a unique serial number.

[0184] A unique knowledge topic identifier is assigned to each knowledge sub-module, which uses an alphanumeric combination coding method. Among them, the domain classification code is used to identify the large field to which the knowledge belongs, such as "new energy vehicle marketing" which can be represented by "XNYQC"; the theme category code is used to distinguish different theme categories, such as "battery endurance related" represented by "DX" and "charging facility related" represented by "CD"; and the unique serial number is a unique digital number assigned to each sub-module to ensure the uniqueness of the identifier.

[0185] For example, the knowledge topic identifier of the "battery endurance basic principle" sub-module can be "XNYQC-DX-001", and the identifier of the "charging facility type and characteristics" sub-module can be "XNYQC-CD-001", etc.

[0186] Step S153: Analyze the logical relationship between each knowledge sub-module, determine the parent-child relationship, parallel relationship, reference relationship and supplementary relationship between the knowledge sub-modules, and construct the knowledge content hierarchical structure tree.

[0187] By analyzing the logical relationship between each knowledge sub-module, the relationship type between them is determined by comparing and analyzing the core content and theme of each sub-module.

[0188] The parent-child relationship refers to the content of one sub-module containing the content of another sub-module, such as the parent-child relationship between the "factors affecting battery endurance" sub-module and the "temperature effect on battery endurance" sub-module, the latter being a part of the former.

[0189] The parallel relationship refers to multiple sub-modules at the same level, which are related in content but do not have a containing relationship, such as the parallel relationship between the "fast charging facility characteristics" and "slow charging facility characteristics" sub-modules.

[0190] The reference relationship refers to the content of one sub-module referencing the content of another sub-module, such as the "battery endurance and charging facility matching strategy" sub-module referencing the content in the "charging facility type and characteristics" sub-module.

[0191] The supplementary relationship refers to the content of one sub-module supplementing the content of another sub-module, such as the "battery maintenance effect on endurance" sub-module supplementing the "factors affecting battery endurance" sub-module.

[0192] According to these relationships, the knowledge content hierarchical structure tree is constructed to clearly show the hierarchy and association between each knowledge sub-module.

[0193] For example, step S1531: Extract the content summary of each knowledge sub-module to generate a module summary that can reflect the core content of the knowledge sub-module.

[0194] The content abstract of each knowledge sub-module is extracted, and a text abstract generation algorithm is used to analyze and extract the content of each sub-module.

[0195] First, the text is segmented and key words are extracted to identify the core vocabulary and key information in the sub-module, and then a module abstract that can summarize the core content of the sub-module is constructed according to the information. The module abstract should be concise and clear, and can accurately reflect the main content and theme of the sub-module.

[0196] For example, the module abstract of the "temperature effect on battery endurance" sub-module can be extracted as "elaborates the specific influence of different temperature conditions on the endurance mileage of new energy vehicle batteries and the related principles".

[0197] Step S1532: Calculate the semantic similarity between the module abstracts of any two knowledge sub-modules to generate a knowledge sub-module similarity matrix.

[0198] To calculate the semantic similarity between the module abstracts of any two knowledge sub-modules, first convert each module abstract into a vector form, and the dimension of the vector is determined based on the semantic features of the words in the abstract.

[0199] Then, the semantic similarity is determined by calculating the degree of association between the two vectors. The higher the degree of association, the more similar the semantics of the two module abstracts. The results of calculating the semantic similarity between all pairs of sub-modules are arranged in matrix form, i.e. the knowledge sub-module similarity matrix, and each element in the matrix represents the semantic similarity of the corresponding two sub-modules.

[0200] Step S1533: According to the knowledge sub-module similarity matrix, a hierarchical clustering algorithm is used to perform clustering analysis on the knowledge sub-modules, and the knowledge sub-modules with a semantic similarity greater than a set similarity threshold are classified into a class to form a preliminary knowledge module group.

[0201] According to the knowledge sub-module similarity matrix, a hierarchical clustering algorithm is used to perform clustering analysis. The hierarchical clustering algorithm starts with each knowledge sub-module as a separate cluster, and then gradually merges clusters with a similarity greater than a set similarity threshold in order of decreasing similarity to form larger clusters.

[0202] By the above method, knowledge sub-modules with high semantic similarity are classified into a class to form a preliminary knowledge module group. For example, sub-modules related to battery endurance will be clustered into one group, and sub-modules related to charging facilities will be clustered into another group.

[0203] Step S1534: Analyze the semantic relationship between each knowledge sub-module in the knowledge module group, identify the knowledge sub-module pair with the inclusion relationship, determine the contained knowledge sub-module as the child node, the knowledge sub-module containing other knowledge sub-modules as the parent node, and establish the parent-child relationship.

[0204] Analyze the semantic relationship between each knowledge sub-module in the knowledge module group, and identify the knowledge sub-module pair with the inclusion relationship by comparing the module abstract and core content of the sub-module.

[0205] If the content of a sub-module is completely covered by the content of another sub-module, the contained sub-module is the child node, and the sub-module containing it is the parent node, thereby establishing the parent-child relationship. For example, in the group of battery endurance influencing factors, the "low temperature effect on battery endurance" sub-module is contained in the "temperature effect on battery endurance" sub-module, the former is the child node, and the latter is the parent node.

[0206] Step S1535: Identify the knowledge sub-modules in the same level that do not have an inclusion relationship but are semantically related, and determine them as parallel relationship.

[0207] In the same knowledge module group or the same level, identify those knowledge sub-modules that do not have an inclusion relationship but are semantically related, and determine them as parallel relationship.

[0208] These sub-modules complement or associate with each other in content, but each has an independent theme and content, such as the "battery capacity effect on endurance" and "battery aging degree effect on endurance" sub-modules, which belong to the same level of the "factors affecting battery endurance" group and do not have an inclusion relationship, but are semantically related, so they are in parallel relationship.

[0209] Step S1536: Find the case where the content of a knowledge sub-module explicitly refers to the content of another knowledge sub-module, and establish a reference relationship.

[0210] Check the content of each knowledge sub-module one by one to find the case where it explicitly mentions or refers to the content of another knowledge sub-module.

[0211] When a sub-module refers to the concept, data or conclusion of another sub-module in the process of elaboration, a reference relationship is established between the two sub-modules, and the specific content and location of the reference are recorded. For example, the "battery endurance and charging facility matching strategy" sub-module refers to the content of "fast charging facility charging speed", so a reference relationship is established between the two sub-modules.

[0212] Step S1537: Identify the knowledge sub-module that supplements or expands the content of any knowledge sub-module, and establish a supplementary relationship.

[0213] Identify those knowledge sub-modules that supplement or extend the content of other knowledge sub-modules. When the content of one sub-module can provide additional information, details or extended explanation for the content of another sub-module, a supplementary relationship is established between them.

[0214] For example, the "Battery endurance test standard" sub-module provides supplementary explanation for the endurance data mentioned in the "Battery endurance basic principle" sub-module, so a supplementary relationship is established between them.

[0215] Step S1538: Based on the knowledge module groups, according to the determined parent-child relationship, parallel relationship, reference relationship and supplementary relationship, the knowledge content hierarchical structure tree of multi-way tree structure is constructed, in which the parent nodes are located in the upper layer, the child nodes are located in the lower layer, and the parallel nodes are located in the same level, and the reference relationship and the supplementary relationship are represented by additional edges.

[0216] Based on the knowledge module groups, each group is taken as a main branch of the knowledge content hierarchical structure tree. According to the determined parent-child relationship, the parent nodes are placed in the upper layer and the child nodes are placed in the lower layer to form the hierarchical structure of the tree.

[0217] The sub-modules of the parallel relationship are placed in the same level to maintain the parallel position between each other. For the reference relationship and the supplementary relationship, additional edges are added between the corresponding child nodes to represent the non-hierarchical association between the sub-modules, which can clearly show the non-hierarchical association between the sub-modules.

[0218] Finally, the knowledge content hierarchical structure tree of multi-way tree structure is constructed, which comprehensively and intuitively reflects various logical relationships between knowledge sub-modules.

[0219] Step S154: Extract the key concepts, core terms and important data in each knowledge sub-module, establish the concept association network within the knowledge sub-module and the cross-reference index between the knowledge sub-modules, generate the association index information, and the association index information includes the association knowledge theme identifier, the association type and the association strength.

[0220] Extract the key concepts in each knowledge sub-module, such as "battery energy density", "charging efficiency", etc.; core terms, such as various professional terms mentioned above; important data, such as charging time range under different charging methods, etc.

[0221] Establish the concept association network within the knowledge sub-module, analyze the association relationship between the key concepts, core terms and important data within the sub-module, such as the positive correlation between "battery energy density" and "battery endurance mileage".

[0222] Meanwhile, an index of cross-references between knowledge sub-modules is established, and the association between different sub-modules is recorded according to the previously determined reference relationship and supplementary relationship. The generated association index information includes an associated knowledge topic identifier, i.e., the knowledge topic identifier of the associated sub-module; an association type, such as reference, supplement, etc.; and an association strength, which is divided according to the closeness and importance of the association, such as strong association, medium association, and weak association.

[0223] Step S155: The knowledge topic identifier, knowledge sub-module content, knowledge content hierarchical relationship, and association index information are integrated and packaged to generate a structured automobile marketing knowledge unit, which is stored in an extensible markup language format.

[0224] The knowledge topic identifier, knowledge sub-module content, knowledge content hierarchical relationship, and association index information are integrated and packaged. According to the syntax rules of the extensible markup language, appropriate tags are defined for each part, such as <topic identifier>, <sub-module content>, <hierarchical relationship>, and <association index>.

[0225] The contents of each part are filled into the corresponding tags to form an extensible markup language document with clear structure and standardized format, i.e., a structured automobile marketing knowledge unit. The above format facilitates computer recognition and processing, and is conducive to knowledge storage and management.

[0226] Step S156: The management interface of the automobile marketing knowledge base is called to add the automobile marketing knowledge unit to the corresponding classification directory of the knowledge base, and to update the global index and knowledge relationship graph of the automobile marketing knowledge base.

[0227] The management interface of the automobile marketing knowledge base is called, which provides functions such as adding knowledge units and updating indexes. According to the knowledge topic identifier and content topic of the automobile marketing knowledge unit, it is added to the corresponding classification directory of the knowledge base, such as the "battery endurance knowledge" directory or the "charging facility knowledge" directory.

[0228] After the addition is completed, the global index of the automobile marketing knowledge base is updated, which records the storage location and key information of all knowledge units for quick retrieval. At the same time, according to the knowledge content hierarchical relationship and association index information in the knowledge unit, the knowledge relationship graph of the knowledge base is updated, so that the graph can accurately reflect the association relationship between the newly added knowledge unit and the existing knowledge unit, ensuring the integrity and association of the knowledge base.

[0229] Step S210: Model training step, the method further comprising a step of training a large model for knowledge content generation and optimization.

[0230] To ensure that the large model can better adapt to the generation and optimization of knowledge content in the field of automobile marketing, it needs to be trained specifically. The training process will combine professional data in the field of automobile marketing, adjust model parameters, and improve the model's ability to generate and optimize knowledge in this field.

[0231] Step S211: Collect professional corpus data in the field of automobile marketing, including automobile marketing knowledge documents, industry reports, marketing cases, and user consultation records, etc.

[0232] Collect professional corpus data in the field of automobile marketing, which comes from a wide range of sources, including marketing knowledge documents published by automobile manufacturers, automobile marketing industry reports issued by industry research institutions, marketing success cases of various automobile brands, and historical user consultation records, etc.

[0233] During the collection process, the legality and compliance of the data must be ensured. For consultation records and other data involving user privacy, data desensitization technology is used to process and remove personal information such as name, contact information, and ID number, etc., to protect user privacy.

[0234] Step S212: Preprocess the collected professional corpus data, including data cleaning, deduplication, format unification, and word segmentation processing.

[0235] Preprocess the collected professional corpus data. Data cleaning is to remove noise and irrelevant information in the data, such as ad pop-up content in documents, and format errors; deduplication is to delete duplicate data to avoid repeated training; format unification is to convert different formats of documents into a unified text format for subsequent processing; and word segmentation processing is to segment continuous text into independent words to prepare for model training.

[0236] Step S213: Construct training data set and validation data set, divide the preprocessed professional corpus data into training data set and validation data set according to the set proportion.

[0237] Divide the preprocessed professional corpus data according to the set proportion, such as 8:2 to divide into training data set and validation data set. The training data set is used for parameter learning of the model, and the validation data set is used to evaluate the performance of the model during training and adjust the training strategy in time.

[0238] Step S214: Configure the training parameters of the large model, including training batch size, training iteration number, learning rate, and regularization coefficient.

[0239] The training parameters of the large model are configured, the training batch size refers to the number of data samples input into the model each time, the number of training iterations refers to the number of complete training of the model on the training data set, the learning rate is used to control the step size of the model parameter update, and the regularization coefficient is used to prevent the model from overfitting and improve the generalization ability of the model.

[0240] According to the size and characteristics of the professional corpus data, appropriate training parameter values are set.

[0241] Step S215: input the training data set into the large model, train the large model according to the configured training parameters, and evaluate the performance of the model using the validation data set during the training process, and adjust the training parameters according to the evaluation results.

[0242] The training data set is input into the large model, and the model is trained according to the configured training parameters. After each training iteration period ends, the performance of the model is evaluated using the validation data set, and the evaluation indicators include the relevance, accuracy and fluency of the generated text, etc.

[0243] According to the evaluation results, if the performance of the model does not meet the expectations, such as the relevance of the generated text to the topic is low, the training parameters are adjusted, such as reducing the learning rate, increasing the number of training iterations, etc., until the performance of the model on the validation data set reaches a satisfactory level.

[0244] Step S216: after the training is completed, the large model is tested, the test data set is used to evaluate the generation effect and optimization ability of the model, if the test results meet the preset requirements, it is determined that the large model is trained; if not, the training parameters are adjusted or the training data is supplemented, and the training is performed again.

[0245] After the training is completed, the large model is tested using an independent test data set, the test data set does not participate in the training and validation process of the model, and can more objectively evaluate the performance of the model.

[0246] The generation effect of the model on the test data set is evaluated, such as the quality of the generated knowledge content, the matching degree with the demand, etc., and the optimization ability of the model, such as the correction effect on the existing problem knowledge content. If the test results meet the preset requirements, the model is a trained large model; if not, the reasons are analyzed, the training parameters are adjusted or more professional corpus data is supplemented, and the training step is executed again until the model test is passed.

[0247] Figure 2 A schematic diagram of exemplary hardware and software components of a large model-based marketing knowledge base construction system 100 that can implement the idea of the present application is shown. For example, the processor 120 can be used in the large model-based marketing knowledge base construction system 100 and used to execute the functions in the present application.

[0248] The large model-based marketing knowledge base construction system 100 can be a general server or a special-purpose server, both of which can be used to implement the large model-based marketing knowledge base construction method of the present application. The present application illustrates only one server, but for the sake of convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0249] For example, the large model-based marketing knowledge base construction system 100 can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. Illustratively, the large model-based marketing knowledge base construction system 100 can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The large model-based marketing knowledge base construction system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0250] For the sake of illustration, only one processor is described in the large model-based marketing knowledge base construction system 100. However, it should be noted that the large model-based marketing knowledge base construction system 100 in the present application can also include multiple processors, so the steps described in the present application as performed by one processor can also be jointly performed or separately performed by multiple processors. For example, if the processor of the large model-based marketing knowledge base construction system 100 performs steps A and B, it should be understood that steps A and B can also be jointly performed by two different processors or separately performed in one processor. For example, a first processor performs step A and a second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0251] In addition, the present application embodiment also provides a readable storage medium, wherein computer executable instructions are pre-set in the readable storage medium, and when the processor executes the computer executable instructions, the large model-based marketing knowledge base construction method as described above is implemented.

[0252] It should be noted that, in order to simplify the description of the present application and to help understand one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, drawing, or description thereof.

Claims

1. A method for constructing a marketing knowledge base based on a large model, characterized in that, The method includes: Determine the knowledge construction needs in the field of automotive marketing, and generate a marketing knowledge requirement description, which includes the scope of marketing topics, the type of knowledge content, and the limitation of knowledge application scenarios. The marketing knowledge requirement description is input into a pre-trained large model, and the text generation interface of the large model is called to perform preliminary knowledge content generation processing to obtain candidate knowledge content corresponding to the marketing knowledge requirement description. Obtain a preset knowledge quality assessment standard, perform knowledge quality assessment processing on the candidate knowledge content according to the knowledge quality assessment standard, and generate a knowledge quality assessment result; The marketing knowledge requirement description, the candidate knowledge content, and the knowledge quality assessment results are input into the large model. The instruction fine-tuning interface of the large model is called to perform knowledge content optimization processing to obtain optimized knowledge content that meets the knowledge quality assessment criteria. The optimized knowledge content is subjected to knowledge structuring and organization processing to generate automotive marketing knowledge units containing knowledge topic identifiers, content hierarchical relationships and related index information, and the automotive marketing knowledge units are added to the automotive marketing knowledge base. The step of obtaining a preset knowledge quality assessment standard, performing knowledge quality assessment processing on the candidate knowledge content according to the knowledge quality assessment standard, and generating a knowledge quality assessment result includes: The system retrieves the preset knowledge quality assessment standards from the standard database of the knowledge management system, analyzes the knowledge quality assessment standards, and determines the specific assessment indicators and indicator weights included in each of the content completeness assessment dimension, expression accuracy assessment dimension, and application adaptability assessment dimension. For the content completeness assessment dimension, the candidate knowledge content is tested for knowledge point coverage to identify whether the candidate knowledge content contains all the core knowledge points required in the marketing knowledge requirement description, and the number of missing knowledge points and the sufficiency score of each knowledge point are calculated. For the accuracy assessment dimension, the candidate knowledge content is checked for terminology standardization, logical coherence, and data accuracy, and is marked with incorrect professional terms, logically contradictory sentences and paragraphs, and inaccurate data descriptions. For the application adaptability assessment dimension, the candidate knowledge content is matched with the knowledge application scenarios defined in the marketing knowledge requirement description to evaluate the degree of support of the knowledge content for specific business scenarios, its applicability to the target user group, and its operability in actual marketing activities. Based on the specific evaluation indicator scores and indicator weights of each evaluation dimension, a weighted summation method is used to calculate the comprehensive score for content completeness, the comprehensive score for expression accuracy, and the comprehensive score for application suitability. The comprehensive scores of each dimension, the scores of specific evaluation indicators, the list of missing knowledge points, the error marking information, and the results of scenario matching analysis are summarized and organized to generate a knowledge quality assessment result that includes the scoring results, problem descriptions, and improvement suggestions.

2. The marketing knowledge base construction method based on a large model according to claim 1, characterized in that, The process of determining knowledge construction needs in the automotive marketing field and generating marketing knowledge requirement descriptions includes: Obtain the preliminary knowledge requirements document provided by the automotive marketing business department, perform text parsing on the preliminary knowledge requirements document, and extract the content type indication information, which includes requirement keywords, topic phrases, and scenario description paragraphs. The extracted demand keywords are subjected to word frequency statistics and semantic clustering to identify core demand themes and secondary demand themes, and to generate a demand theme hierarchy structure. The extracted scene description paragraphs are processed for scene feature extraction to determine the scene feature parameters of knowledge application. The scene feature parameters include the specific business scenario, target user group and expected application method. The hierarchical structure of the demand topics, the content type indication information, and the scenario feature parameters are integrated and processed to generate a preliminary description of marketing knowledge demand. The preliminary marketing knowledge requirement description is fed back to the automotive marketing business department for requirement confirmation. The department then receives feedback on the requirement modification, and the preliminary marketing knowledge requirement description is iteratively adjusted based on the feedback to generate the final marketing knowledge requirement description.

3. The method for constructing a marketing knowledge base based on a large model according to claim 1, characterized in that, The process involves inputting the marketing knowledge requirement description into a pre-trained large model, calling the large model's text generation interface to perform preliminary knowledge content generation processing, and obtaining candidate knowledge content corresponding to the marketing knowledge requirement description, including: The marketing knowledge demand description is processed into demand feature vectorization, converting the text-based demand description into a demand feature vector that conforms to the input format of the large model. The domain knowledge activation module of the large model is invoked to associate and match the demand feature vector with the pre-trained knowledge graph of the automotive marketing domain, thereby activating the domain knowledge parameters in the large model that are related to the marketing knowledge demand description. Configure the generation parameter set of the text generation interface, which includes topic relevance weight, content depth coefficient, professional terminology density threshold and output length range; The demand feature vector and the generation parameter set are input into the text generation interface to trigger the large model to perform knowledge content generation operations and obtain the initial knowledge text. The initial knowledge text is subjected to redundant information filtering to remove duplicate statements, irrelevant extended content, and contradictory content with conflicting statements, resulting in purified intermediate knowledge text. The intermediate knowledge text is checked for completeness. If there are obvious missing paragraphs, the missing paragraph identifier and the original requirement feature vector are re-inputted into the text generation interface to supplement the text. The supplemented content is then spliced ​​and merged with the intermediate knowledge text to obtain candidate knowledge content.

4. The method for constructing a marketing knowledge base based on a large model according to claim 1, characterized in that, Regarding the content completeness assessment dimension, the candidate knowledge content undergoes a knowledge point coverage test to identify whether it contains all the core knowledge points required in the marketing knowledge requirement description. The test also calculates the number of missing knowledge points and scores the sufficiency of explanation for each knowledge point, including: The core knowledge points of the marketing knowledge requirement description are extracted and processed. Combined with the knowledge system of the automotive marketing field, the core knowledge point set corresponding to the marketing knowledge requirement description is determined. Each core knowledge point includes the topic name, necessary exposition aspects and level of detail requirements. A knowledge point identification model is constructed. The knowledge point identification model uses a bidirectional long short-term memory network combined with a conditional random field algorithm to scan the candidate knowledge content segment by segment and identify the knowledge points already contained in the candidate knowledge content and the actual explanatory aspects of each knowledge point. The identified included knowledge points are compared and matched with the core knowledge point set to determine the core knowledge points that are not included, and the number of missing knowledge points is counted. For each core knowledge point that has been included, match its actual exposition with the necessary exposition, and calculate the coverage of the exposition. Based on the coverage of the aspects described and the level of detail of each aspect, a score for the sufficiency of explanation for each knowledge point is calculated according to a preset scoring rule. The score for the sufficiency of explanation is positively correlated with the coverage and level of detail of the aspects described. The number of missing knowledge points and the sufficiency score of each knowledge point will be recorded in the content completeness assessment sub-result, which will serve as the basis for calculating the comprehensive content completeness score.

5. The method for constructing a marketing knowledge base based on a large model according to claim 1, characterized in that, The assessment of accuracy involves checking the candidate knowledge content for terminology standardization, logical coherence, and data accuracy. Incorrect terminology, logically contradictory statements, and inaccurate data representations are flagged. Retrieve a standard terminology set from the automotive marketing terminology database. The standard terminology set includes term names, standard definitions, correct usage scenarios, and examples of common incorrect usage. The candidate knowledge content is subjected to terminology extraction, all professional terms appearing in the text are identified, the extracted professional terms are matched and compared with the standard terminology set, the spelling correctness, definition consistency and usage scenario adaptability of the terms are checked, and incorrect terms that do not conform to the standard are marked. Sentence-level logical relationship analysis is performed on the candidate knowledge content. Dependency parsing is used to identify the logical connection relationship between adjacent sentences, including causal relationship, progressive relationship, adversative relationship and parallel relationship. It is checked for improper use of logical connectors or logical contradictions, and sentence paragraphs with logical contradictions are marked. Extract data descriptions from the candidate knowledge content, compare the extracted data descriptions with preset authoritative automotive marketing data sources, check whether the data values ​​are accurate, whether the units are consistent, and whether the descriptions are objective, and mark content segments with inaccurate data descriptions. The data descriptions include verifiable information such as values, percentages, time, location, and event descriptions. The errors in terminology, logical contradictions, and data errors are categorized and statistically analyzed. The error rate of terminology, the occurrence rate of logical contradictions, and the proportion of data errors are calculated and used as the basis for calculating the comprehensive score of expression accuracy.

6. The method for constructing a marketing knowledge base based on a large model according to claim 1, characterized in that, The application adaptability assessment dimension involves performing a matching degree analysis between the candidate knowledge content and the knowledge application scenarios defined in the marketing knowledge requirement description. This assesses the degree to which the knowledge content supports specific business scenarios, its applicability to the target user group, and its operability in actual marketing activities. This includes: The marketing knowledge demand description is analyzed, and the scenario feature parameters of the knowledge application scenario are extracted. The scenario feature parameters include business scenario type, target user group characteristics, marketing activity purpose, knowledge application method and expected effect indicators. The candidate knowledge content is subjected to application scenario association information extraction to identify the applicable scenario description, recommended users, suggested application methods and expected goals implicit in the knowledge content; The similarity between the applicable scenario description of the candidate knowledge content and the business scenario type in the scenario feature parameters is calculated to evaluate the degree of support of the knowledge content for the specific business scenario. The recommended users of the candidate knowledge content are matched and analyzed with the target user group characteristics in the scene feature parameters to evaluate the applicability score of the knowledge content to the target user group. The operability score of the knowledge content in actual marketing activities is evaluated based on the degree of fit between the suggested application methods and expected goals of the candidate knowledge content and the marketing activity objectives, knowledge application methods and expected effect indicators in the scenario feature parameters. The scenario support score, user applicability score, and operability score are weighted according to preset weights to obtain the application adaptability comprehensive score, which reflects the overall matching level between candidate knowledge content and application scenario.

7. The method for constructing a marketing knowledge base based on a large model according to claim 1, characterized in that, The process involves inputting the marketing knowledge requirement description, the candidate knowledge content, and the knowledge quality assessment results into the large model, then calling the large model's instruction fine-tuning interface to perform knowledge content optimization processing to obtain optimized knowledge content that meets the knowledge quality assessment criteria, including: The knowledge quality assessment results are processed by instruction conversion, which converts the problem descriptions and improvement suggestions into optimization instructions that conform to the instruction format of the large model. The optimization instructions include supplementary instructions for content completeness, correction instructions for expression accuracy, and adjustment instructions for application adaptability. Construct an optimization prompt word template, and fill the marketing knowledge requirement description, the candidate knowledge content and the optimization instructions into the optimization prompt word template according to the preset text structure to generate a comprehensive optimization prompt word that includes the requirement background, current content, existing problems and optimization direction; Configure the optimization parameters of the large model instruction fine-tuning interface, including the learning rate, number of training epochs, context window size, and output content length limit; The comprehensive optimization prompts and optimization parameters are input into the instruction fine-tuning interface of the large model to trigger the large model to perform knowledge content optimization operations and obtain preliminary optimized knowledge content. A secondary quality assessment is performed on the initially optimized knowledge content to generate the secondary knowledge quality assessment results of the initially optimized knowledge content; The secondary knowledge quality assessment result is compared with the knowledge quality assessment standard. If the secondary knowledge quality assessment result meets the knowledge quality assessment standard, the preliminary optimized knowledge content is determined as optimized knowledge content. If the secondary knowledge quality assessment result does not meet the knowledge quality assessment standard, the secondary knowledge quality assessment result is used as a new knowledge quality assessment result, and the optimization processing steps are repeated until optimized knowledge content that meets the knowledge quality assessment standard is obtained.

8. The method for constructing a marketing knowledge base based on a large model according to claim 1, characterized in that, The step of performing knowledge structuring organization processing on the optimized knowledge content to generate automotive marketing knowledge units containing knowledge topic identifiers, content hierarchical relationships, and related index information, and adding the automotive marketing knowledge units to the automotive marketing knowledge base, includes: The optimized knowledge content is subject to topic segmentation, and the optimized knowledge content is decomposed into multiple knowledge sub-modules according to the content topics. Each knowledge sub-module revolves around a core topic. Each knowledge submodule is assigned a unique knowledge topic identifier, which is encoded using an alphanumeric combination and includes a domain classification code, a topic category code, and a unique serial number. Analyze the logical relationships between each knowledge sub-module, determine the parent-child relationship, parallel relationship, referencing relationship and supplementary relationship between the knowledge sub-modules, and construct a hierarchical structure tree of knowledge content; Extract key concepts, core terms and important data from each knowledge submodule, establish a concept association network within the knowledge submodule and a cross-reference index between knowledge submodules, and generate association index information, which includes the associated knowledge topic identifier, association type and association strength. The knowledge topic identifier, knowledge sub-module content, knowledge content hierarchy and related index information are integrated and encapsulated to generate a structured automotive marketing knowledge unit, which is stored in Extensible Markup Language format. The management interface of the automotive marketing knowledge base is invoked to add the automotive marketing knowledge units to the corresponding category directories of the knowledge base, and to update the global index and knowledge relationship graph of the automotive marketing knowledge base.

9. A marketing knowledge base construction system based on a large model, characterized in that, The system includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the marketing knowledge base construction method based on a large model as described in any one of claims 1-8.

Images