Intelligent marketing document generation device and method based on multi-source data fusion

Abstract

The invention discloses an intelligent marketing document generation device and method based on multi-source data fusion. The method comprises four processes of user instruction analysis, regular cleaning, model parameter extraction, Neo4j knowledge graph verification and structured parameter output. External data acquisition: calling a Baidu large model to acquire data according to parameters, and screening high-quality data through duplicate removal, semantic enhancement and weighting; internal cases are processed, timed slicing cases are stored in a library, filtering, propagation calculation and multi-dimensional scoring are combined, and a CoT inference chain report is generated; and performing multi-modal output, adapting formats, integrating data by means of a BART model and outputting a document with metadata. The device and the method cooperate with each other, through multi-source fusion, knowledge graph association and multi-mode conversion, high-quality marketing documents adaptive to multiple industries are efficiently generated, support is provided for decision making, and enterprises are assisted to improve market response and output efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of computer application technology, and specifically relates to an intelligent marketing document generation device and method based on multi-source data fusion. Background Art

[0002] Current mainstream marketing document generation tools are primarily based on discrete data call models and fixed template engines. These tools face significant technical bottlenecks in data integration efficiency, content intelligence, and output formats, as exemplified by the following:

[0003] 1. Overview of existing technical solutions

[0004] Data integration method: A decentralized data storage architecture is used. Brand data (such as CRM systems), industry information (third-party API platforms), and marketing cases (local databases) are stored in different formats such as JSON / XML / CSV. Communication protocols (RESTfulAPI / SOAP) are not unified, and there is a lack of standardized access interfaces, resulting in serious data silos.

[0005] Generation engine architecture: Relying on fixed template engines (such as Word macros) or simple rule engines, it can only generate basic content through preset marketing scenario templates (such as competitive product analysis report templates) and cannot achieve dynamic semantic association and knowledge reasoning.

[0006] Application of intelligent technology: Only the primary application of NLP technology is used to achieve keyword extraction, and no industry knowledge graph is constructed; case matching relies on manually preset rules (such as "promotional activities" keyword matching) and lacks semantic understanding capabilities.

[0007] Output format: Mainly supports plain text output. A few tools can convert to PDF format, but lack multimodal interaction capabilities such as chart embedding and dynamic data visualization.

[0008] 2. Analysis of shortcomings of existing technologies

[0009] (1) Inefficient data integration

[0010] Heterogeneous data barriers: Brand CRM systems (such as Salesforce), social media monitoring tools (Brandwatch), and industry databases (iResearch) require customized interfaces to connect due to differences in data formats and protocols. This results in data synchronization cycles generally exceeding T+1 days, making it impossible to meet the real-time data requirements of market analysis documents (such as sudden public opinion events).

[0011] Lack of real-time performance: Traditional systems are unable to capture and analyze data in real time. For example, when sudden marketing events (such as hot topic associations) occur, document generation still relies on overnight synchronized data, causing the content to lag behind market trends.

[0012] (2) Insufficient intelligent content generation

[0013] Lack of knowledge reasoning capabilities: Lacking industry knowledge graph support, the system cannot automatically link historical brand activity data with current market trends. For example, when a food brand launches a new product, the system cannot automatically link the "healthy snacks" market trend with the brand's "low sugar" selling point, resulting in a lack of data-driven logic in content planning.

[0014] Weak semantic understanding capabilities: The traditional cosine similarity algorithm only calculates the angle between term vectors and cannot identify the hyponym relationship between "mobile phone" and "smartphone". The case matching accuracy rate is less than 62%, resulting in poor relevance of recommended content.

[0015] (3) Simplified output format

[0016] Existing tools only support basic text or static PDF output, failing to meet the multimodal content demands of marketing scenarios. For example, they cannot automatically generate documents containing visual elements such as trend charts and user profile heat maps. They also lack the dynamic interactivity of HTML formats (such as clicking on a chart to drill down to details), making them difficult to adapt to the display needs of online marketing channels.

[0017] 3. Necessity of Technological Improvement

[0018] As market competition intensifies, existing technologies can no longer meet the three core requirements of marketing document generation:

[0019] Data-driven decision-making: It is necessary to integrate multi-source and multi-dimensional data on brands, markets, and users to achieve intelligent content generation based on real-time insights;

[0020] Agile response to the market: Document generation cycles need to be shortened from the traditional four hours to within 10 minutes to cope with rapidly changing marketing scenarios;

[0021] Multimodal intelligent output: It must support multiple formats such as PDF / HTML / Word, and integrate rich media content such as dynamic charts and interactive elements.

[0022] Therefore, there is an urgent need for an intelligent marketing document generation solution that can solve heterogeneous data fusion, improve semantic understanding capabilities and support multimodal output. Summary of the Invention

[0023] In view of this, the purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art, such as low data integration efficiency, insufficient content intelligence, and single output format, and to provide an intelligent marketing document generation device and method based on multi-source data fusion. This solution achieves efficient integration of brand data, industry information, and internal cases through standardized data interfaces, enhances semantic understanding and case matching capabilities with the help of knowledge graphs and natural language processing technologies, and realizes diversified document format generation based on multimodal output adapters, thereby solving the data island problem in the prior art, shortening the document generation cycle, improving content relevance and display effects, and meeting the modern marketing needs of data-driven decision-making, agile market response, and multimodal output.

[0024] To achieve the above objectives, the present invention provides, in a first aspect, an intelligent marketing document generation device based on multi-source data fusion, comprising:

[0025] User instruction input module: used to receive marketing document generation demand instructions submitted by users;

[0026] NLP parsing engine module: connected to the user instruction input module, configured to perform natural language processing on the user instruction, extract structured parameters and verify the legitimacy of the parameters;

[0027] Parameter decomposition module: connected to the NLP parsing engine module, used to split the parsed instructions into task parameters corresponding to external data acquisition and internal case calls;

[0028] External data engine module: connected to the parameter disassembly module, configured to collect and process external industry data according to task parameters to generate semantically enhanced industry information;

[0029] Internal case engine module: connected to the parameter decomposition module, configured to initiate case library search based on task parameters, screen matching marketing cases and output them;

[0030] Multimodal output module: connected to the external data engine module and the internal case engine module respectively, used to integrate industry information and marketing cases, generate and output marketing documents in multiple formats;

[0031] Data layer interface module: connected with the external data engine module and the internal case engine module, including a standardized API gateway, supporting RESTful / SOAP dual protocols to access heterogeneous data sources and realize data interaction;

[0032] Knowledge graph storage unit: connected to the internal case engine module and NLP parsing engine module, using Neo4j graph database to store entity relationships, covering the "brand-industry-case portrait" association form, for semantic association and case retrieval support;

[0033] Core algorithm engine: including

[0034] Hybrid generation engine: Connected with the multimodal output module and parameter decomposition module, it integrates the Wenxin Yiyan API call module and rule engine, supports dynamic adjustment of generation parameters, and is used for intelligent construction of marketing content;

[0035] Case recommendation and evaluation engine: This engine connects to the internal case engine module, deploys a TF-IDF vector index library, supports millisecond-level similarity calculation for tens of millions of case libraries, and provides similarity evaluation for case library retrieval;

[0036] Case library retrieval module: connected to the internal case engine module and the case recommendation and evaluation engine, relying on the semantic association of the knowledge graph storage unit and combining the similarity calculation results of the case recommendation and evaluation engine, to screen suitable cases from the case library.

[0037] Furthermore, the NLP parsing engine module uses a natural language processing model to perform word segmentation, entity recognition and intent understanding on user instructions, and the natural language processing model includes a pre-trained language model;

[0038] The external data engine module includes a data acquisition unit and a data processing unit. The data acquisition unit is configured to call a third-party data source to obtain industry trends and market dynamics data; the data processing unit is configured to clean, semantically enhance and importance score the collected data.

[0039] Furthermore, when the internal case engine module initiates a case library search, it calls the TF-IDF vector index library capability of the case recommendation and evaluation engine, and combines the entity relationship association of the knowledge graph storage unit to achieve case screening.

[0040] Furthermore, the semantic matching of the case library retrieval module adopts a vector similarity calculation method to compare the user instruction parameters with the case feature vector, and screen out cases whose similarity meets the threshold. The vector similarity calculation is implemented based on the TF-IDF vector index library of the case recommendation and evaluation engine.

[0041] A second aspect of the present invention provides a method for generating intelligent marketing documents based on multi-source data fusion, the production method comprising:

[0042] S1. User instruction parsing process, processing the original instruction to obtain structured parameters;

[0043] S2. External data acquisition process, collecting and optimizing industry data based on structured parameters;

[0044] S3. Internal case processing process, searching, scoring and generating evaluations for the enterprise case library;

[0045] S4. Multimodal output process, converting the fused data into the target format document.

[0046] Furthermore, the S1 user instruction parsing process includes: S11. using regular expressions to clean the original instruction text, remove redundant symbols and unify punctuation formats;

[0047] S12. Extract parameters such as brand, industry level, and output format through the model and generate a structured JSON object;

[0048] S13. Verify the legitimacy of the brand-industry level based on the Neo4j knowledge graph and output the verification mark.

[0049] Furthermore, the S2 external data acquisition process includes:

[0050] S21. Construct a search query based on the structured parameters, call Baidu's large model to obtain industry data, and use the similarity hash algorithm to remove duplicates;

[0051] S22. Use the Wenxin model to perform TF-IDF cleaning, BERT-CLS semantic enhancement, and summary generation on the data;

[0052] S23. Sort the data using the weighted formula: score = 0.6 × semantic_sim + 0.3 × time_decay + 0.1 × authority, where:

[0053] semantic_sim is the semantic similarity (range: 0 to 1), which is used to calculate the relevance between the text and user needs based on the model;

[0054] time_decay is the time decay factor (range: 0 to 1), which is determined by the difference between the data release time and the current time;

[0055] Authority is the authority score of the data source (ranging from 0 to 1), which is based on the source level settings such as industry white papers (0.7) and internal corporate reports (0.5). The top N best data are filtered by the score.

[0056] Furthermore, the S3 internal case processing process includes:

[0057] Steps for building a knowledge base and case library: Slice the original cases using the Wenxin model through scheduled tasks, extract entity relationships, and store them in a structured manner in the Neo4j graph database and Milvus vector library.

[0058] S32. Semantic retrieval step: Industry path filtering is performed in Neo4j based on user instruction parameters, and candidate cases are screened using the propagation data score propagation_score = ∑(propagation amount × e^(-0.1 × Δt)), where:

[0059] The amount of dissemination refers to the dissemination data of the case on Weibo and WeChat platforms;

[0060] Δt is the time difference between the data release time and the current time, in months;

[0061] S33. Dynamic scoring step: Comprehensively score the candidate cases using the formula final_score = 0.6 × semantic_sim + 0.3 × industry_sim + 0.1 × propagation_score, where:

[0062] industry_sim is the industry matching degree (range 0 to 1), which is calculated based on the Jaccard similarity of industry paths;

[0063] semantic_sim is the semantic similarity between the case summary and the user instruction (range 0 to 1), calculated by the Sentence-BERT model;

[0064] S34. Steps for generating explainable evaluation: Generate a CoT and reasoning chain evaluation report based on the scoring dimensions, summarizing the highlights of the case.

[0065] Furthermore, the S4 multimodal output process includes:

[0066] S41. Format Adaptation Step: Based on the output format, including PDF / HTML / Word, use Apache POI and pdfkit components to automatically convert Markdown to the target format;

[0067] S42. Content integration step: Use the BART model to integrate external data and case content, ensuring a logical coherence score > 0.7;

[0068] S43. Metadata generation step: output document metadata including data source identification and generation time.

[0069] The present invention adopts the above technical solution, which has at least the following beneficial effects:

[0070] This invention efficiently solves the problem of integrating heterogeneous data sources through standardized data interfaces and multi-source data fusion technology: it uses a unified API gateway to achieve access to data sources of different protocols and formats, breaks down data silos, significantly improves data synchronization efficiency, and can obtain market dynamic data in real time to meet the data timeliness requirements of marketing scenarios.

[0071] Based on knowledge graphs and natural language processing technology, this invention greatly enhances the intelligence level of marketing document generation: by constructing a relationship network of brands, industries, and cases, combined with the semantic understanding ability of large models, it realizes cross-domain semantic associations, accurately matches related cases and market trends, and changes the low accuracy problem of traditional keyword matching, making document content more logical and data-supported.

[0072] The present invention adopts a modular architecture and parallel processing mechanism to achieve the rapid generation of marketing documents: the functional modules work together to efficiently process the entire process from user instruction parsing to document output, greatly shortening the document generation time, enabling enterprises to quickly respond to market hotspots and changes, and improve marketing efficiency.

[0073] The present invention uses multimodal output technology to meet diverse marketing display needs: it supports conversion of multiple document formats, and can embed dynamic charts, interactive elements, etc., breaking through the limitations of traditional plain text or static documents, adapting to multiple online and offline scenarios, and enhancing the display effect and attractiveness of marketing content.

[0074] The present invention has dynamic expansion capabilities and flexible adaptability: standardized interfaces and extensible knowledge graph architecture facilitate the addition of new data sources and adaptation to the needs of different industries, without the need for large-scale system reconstruction, reducing the cost of enterprise system upgrades and improving the versatility of technical solutions.

[0075] This invention provides reliable support for marketing decisions through data credibility assessment and explainable reasoning: it comprehensively scores the data to ensure data quality, and generates clear case recommendation logic to make the content of marketing documents more credible, helping companies make accurate decisions based on reliable information. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0077] Figure 1 A schematic diagram of a structure of an intelligent marketing document generation device according to the present invention;

[0078] Figure 2 This is a flow chart of the intelligent marketing document generation method of the present invention;

[0079] Figure 3 This is a schematic diagram of the third structure of the intelligent marketing document generation device of the present invention;

[0080] Figure 4 This is a schematic diagram of constructing the knowledge base case library of the present invention. DETAILED DESCRIPTION

[0081] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0082] In the description of the embodiments of the present invention, unless otherwise specified or limited, the terms "connected" and "connection" should be understood broadly. For example, they can refer to electrical connection, direct connection, indirect connection through an intermediate medium, or internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention.

[0083] Example 1

[0084] like Figure 1 As shown, this embodiment provides an intelligent marketing document generation device based on multi-source data fusion, including:

[0085] User instruction input module: used to receive marketing document generation demand instructions submitted by users;

[0086] NLP parsing engine module: connected to the user instruction input module, configured to perform natural language processing on the user instruction, extract structured parameters and verify the legitimacy of the parameters;

[0087] Parameter decomposition module: connected to the NLP parsing engine module, used to split the parsed instructions into task parameters corresponding to external data acquisition and internal case calls;

[0088] External data engine module: connected to the parameter disassembly module, configured to collect and process external industry data according to task parameters to generate semantically enhanced industry information;

[0089] Internal case engine module: connected to the parameter decomposition module, configured to initiate case library search based on task parameters, screen matching marketing cases and output them;

[0090] Multimodal output module: connected to the external data engine module and the internal case engine module respectively, used to integrate industry information and marketing cases, generate and output marketing documents in multiple formats;

[0091] Data layer interface module: connected with the external data engine module and the internal case engine module, including a standardized API gateway, supporting RESTful / SOAP dual protocols to access heterogeneous data sources and realize data interaction;

[0092] Knowledge graph storage unit: connected to the internal case engine module and NLP parsing engine module, using Neo4j graph database to store entity relationships, covering the "brand-industry-case portrait" association form, for semantic association and case retrieval support;

[0093] Core algorithm engine: including

[0094] Hybrid generation engine: Connected with the multimodal output module and parameter decomposition module, it integrates the Wenxin Yiyan API call module and rule engine, supports dynamic adjustment of generation parameters, and is used for intelligent construction of marketing content;

[0095] Case recommendation and evaluation engine: This engine connects to the internal case engine module, deploys a TF-IDF vector index library, supports millisecond-level similarity calculation for tens of millions of case libraries, and provides similarity evaluation for case library retrieval;

[0096] Big model scoring module: connected to the internal case engine module and the case recommendation and evaluation engine, configured to call the Wenxin big model API to perform semantic matching and credibility scoring on candidate cases. The scoring results are used for case screening and sorting;

[0097] Case library retrieval module: connected to the internal case engine module and the case recommendation and evaluation engine, relying on the semantic association of the knowledge graph storage unit and combining the similarity calculation results of the case recommendation and evaluation engine, to screen suitable cases from the case library.

[0098] As an implementation method, the NLP parsing engine module in this embodiment uses a natural language processing model to perform word segmentation, entity recognition, and intent understanding on user instructions. The natural language processing model includes a pre-trained language model.

[0099] The external data engine module includes a data acquisition unit and a data processing unit. The data acquisition unit is configured to call a third-party data source to obtain industry trends and market dynamics data; the data processing unit is configured to clean, semantically enhance and perform large model-driven importance scoring on the collected data.

[0100] As an implementation method, when the internal case engine module in this embodiment initiates a case library search, it calls the TF-IDF vector index library capability of the case recommendation and evaluation engine, combines the entity relationship association of the knowledge graph storage unit and the dynamic scoring results of the large model scoring module to achieve case screening.

[0101] As an implementation method, the semantic matching of the case library retrieval module described in this embodiment adopts a vector similarity calculation method to compare the user instruction parameters with the case feature vector, and combines the semantic score (value range 0 to 1) output by the large model scoring module to screen cases whose similarity meets the threshold. The vector similarity calculation is implemented based on the TF-IDF vector index library of the case recommendation evaluation engine.

[0102] This invention efficiently solves the problem of integrating heterogeneous data sources through standardized data interfaces and multi-source data fusion technology: it uses a unified API gateway to achieve access to data sources of different protocols and formats, breaks down data silos, significantly improves data synchronization efficiency, and can obtain market dynamic data in real time to meet the data timeliness requirements of marketing scenarios.

[0103] Based on knowledge graphs and natural language processing technology, this invention greatly enhances the intelligence level of marketing document generation: by constructing a relationship network of brands, industries, and cases, combining the semantic understanding ability of the large model and the dynamic scoring mechanism of the large model scoring module, it realizes cross-domain semantic association, accurately matches related cases and market trends, and changes the low accuracy problem of traditional keyword matching, making the document content more logical and data-supported.

[0104] Example 2

[0105] like Figure 2 As shown, this embodiment provides a method for generating intelligent marketing documents based on multi-source data fusion, which adopts any one of the devices in the above embodiment 1. The production method includes:

[0106] S1. User instruction parsing process, processing the original instruction to obtain structured parameters;

[0107] S2. External data acquisition process, collecting and optimizing industry data based on structured parameters;

[0108] S3. Internal case processing process, searching, scoring and generating evaluations for the enterprise case library;

[0109] S4. Multimodal output process, converting the fused data into the target format document.

[0110] As an implementation method, the S1 user instruction parsing process in this embodiment includes: S11. using regular expressions to clean the original instruction text, remove redundant symbols and unify the punctuation format;

[0111] S12. Extract parameters such as brand, industry level, and output format through the model and generate a structured JSON object;

[0112] S13. Verify the legitimacy of the brand-industry level based on the Neo4j knowledge graph and output the verification mark.

[0113] As an implementation method, the S2 external data acquisition process in this embodiment includes:

[0114] S21. Construct a search query based on the structured parameters, call Baidu's large model to obtain industry data, and use the similarity hash algorithm to remove duplicates;

[0115] S22. Use the Wenxin model to perform TF-IDF cleaning, BERT-CLS semantic enhancement, and summary generation on the data;

[0116] S23. Use weighted formula to sort data:

[0117] score = 0.6 × semantic_sim + 0.3 × time_decay + 0.1 × authority, where:

[0118] semantic_sim is the semantic similarity (range: 0 to 1), which is used to calculate the relevance between the text and user needs based on the model;

[0119] time_decay is the time decay factor (range: 0 to 1), which is determined by the difference between the data release time and the current time;

[0120] Authority is the authority score of the data source (ranging from 0 to 1), which is based on the source level settings such as industry white papers (0.7) and internal corporate reports (0.5). The top N best data are filtered by the score.

[0121] like Figure 4 As shown, as an implementation method, the S3 internal case processing process in this embodiment includes:

[0122] Steps for building a knowledge base and case library: Slice the original cases using the Wenxin model through scheduled tasks, extract entity relationships, and store them in a structured manner in the Neo4j graph database and Milvus vector library.

[0123] S32. Semantic retrieval step: Industry path filtering is performed in Neo4j based on user instruction parameters, and candidate cases are screened using the propagation data score propagation_score = ∑(propagation amount × e^(-0.1 × Δt)), where:

[0124] The amount of dissemination refers to the dissemination data of the case on Weibo and WeChat platforms;

[0125] Δt is the time difference between the data release time and the current time, in months;

[0126] S33. Dynamic scoring step: Comprehensively score the candidate cases using the formula final_score = 0.6 × semantic_sim + 0.3 × industry_sim + 0.1 × propagation_score, where:

[0127] industry_sim is the industry matching degree (range 0 to 1), which is calculated based on the Jaccard similarity of industry paths;

[0128] semantic_sim is the semantic similarity between the case summary and the user instruction (range 0 to 1), calculated by the Sentence-BERT model;

[0129] S34. Steps for generating explainable evaluation: Generate a CoT and reasoning chain evaluation report based on the scoring dimensions, summarizing the highlights of the case.

[0130] As an implementation method, the S4 multimodal output process in this embodiment includes:

[0131] S41. Format Adaptation Step: Based on the output format, including PDF / HTML / Word, use Apache POI and pdfkit components to automatically convert Markdown to the target format;

[0132] S42. Content integration step: Use the BART model to integrate external data and case content, ensuring a logical coherence score > 0.7;

[0133] S43. Metadata generation step: output document metadata including data source identification and generation time.

[0134] Implementation Three

[0135] The device of this embodiment adopts a modular layered architecture. Figure 3 As shown in the figure, the process logic diagram of the intelligent marketing document generation system, the module names and core logic are analyzed as follows:

[0136] 1. Module name and function

[0137] User instructions: are the "input starting point" of the system, representing the user's demand for generating marketing documents (such as "generate a beauty brand summer promotion case analysis report").

[0138] NLP parsing engine: Core function: Using natural language processing (NLP) technology to break down the semantics, intent, and parameters of user instructions (such as extracting key information such as "beauty brand", "summer promotion", and "case analysis") is a key link in "understanding user needs".

[0139] Parameter decomposition: Function: Split the requirements after NLP analysis into external data collection parameters (such as required industry data dimensions) and internal case call parameters (such as case screening conditions) to achieve "demand structuring".

[0140] External data engine: Function: Collect data (such as summer promotion trends in the beauty industry and competitive product dynamics) from external data sources (such as industry databases and news platforms) according to parameter decomposition requirements.

[0141] Baidu search + Wenxin optimization: is the "specific implementation sub-process" of the external data engine:

[0142] First, use Baidu search to crawl public data (such as industry information and public opinion); then use the Wenxin big model (such as ERNIE) to perform semantic enhancement, deduplication, and summarization (such as extracting the core strategy of promotional cases).

[0143] Internal case engine: Function: Filtering data from the company's internal case library (such as historical marketing plans and success stories) according to parameter decomposition requirements is the core of "reusing internal experience."

[0144] Case library search: This is the "specific implementation sub-process" of the internal case engine: using search algorithms (such as vector search and keyword matching) to find content that matches user needs (such as past beauty brand summer promotion cases) from the case library.

[0145] Big model scoring: Function: Use big models (such as Wenxin and GPT) to perform quality scoring on retrieved internal cases (such as evaluating the relevance, effectiveness, and reusability of the cases) to screen out high-quality content.

[0146] Multimodal output: It is the "output endpoint" of the system: it integrates external optimization data and internal high-quality cases to generate multi-format documents (such as PDF reports, PPT solutions) or visual content (such as data dashboards, trend charts) to meet the needs of marketing scenarios.

[0147] 2. Core logical link: Demand understanding link: user instructions → NLP parsing engine → parameter decomposition, realizing the transformation of "natural language requirements" to "structured task parameters", allowing the system to "understand what the user wants."

[0148] Data collection link:

[0149] External: Parameter decomposition → External data engine → Baidu search + Wenxin optimization

[0150] Internal: Parameter decomposition → Internal case engine → Case library search → Large model scoring

[0151] Achieve "multi-source data integration": capture real-time industry dynamics (external) and reuse the company's historical experience (internal) to ensure data comprehensiveness.

[0152] Content generation link: external data (after optimization) + internal cases (after scoring) → multimodal output

[0153] Core logic: "Data fusion + intelligent processing" generates documents / content that meet marketing scenarios and solves the problems of "data isolation and rigid generation" in traditional solutions.

[0154] Method steps:

[0155] 1. User command analysis (A→B→C)

[0156] Input information: Original command text: "Please output the brand L'Oréal, the primary industry is consumer goods, the secondary industry is cosmetics, and output in PDF format."

[0157] Processing process:

[0158] Text cleaning: Regular expressions are used to remove redundant symbols and unify punctuation formats.

[0159] Entity Recognition: BERT-CRF model extraction parameters:

[0160] json

[0161] {

[0162] "brand":"L'Oréal",

[0163] "industry_level1":"Consumer goods",

[0164] "industry_level2":"cosmetics",

[0165] "output_format":"pdf"

[0166] }

[0167] Intent Verification: Verify brand-industry level legitimacy through Neo4j knowledge graph.

[0168] Output information:

[0169] Structured parameter object (including brand, industry level, and output format).

[0170] Verify flag (True / False).

[0171] 2. External data information source engine (D→F)

[0172] Input information: structured parameter object

[0173] Processing process:

[0174] Baidu large model search call: Construct query: "L'Oréal cosmetics industry trends 2025" to obtain the first 50 results (SimHash deduplication rate > 92%)

[0175] Wenxin model optimization:

[0176] Data cleaning: TF-IDF filters advertising content (retaining core marketing analysis text).

[0177] Semantic enhancement: Generate summaries (BERT-CLS extracts key sentences). Associate brand tags (calculate similarity through knowledge graph embedding vectors).

[0178] Sorting strategy: pythonscore = 0.6*semantic_sim + 0.3*time_decay + 0.1*authority, taking the top three best data. This formula is used to comprehensively score and rank candidate content in search results or case libraries. By weightedly integrating three core dimensions (semantic relevance semantic_sim, timeliness time_decay, and authority time_decay), it ensures that the output meets user needs while also ensuring data credibility and real-time performance. For example:

[0179] Candidate content A (Baidu search results):

[0180] Semantic similarity: 0.85 (the content is highly relevant to “L’Oréal Marketing Trends”)

[0181] Time Decay: 0.6 (released 6 months ago)

[0182] Authority: 0.7 (Source: industry white paper)

[0183] Total score: 0.6 × 0.85 + 0.3 × 0.6 + 0.1 × 0.7 = 0.78

[0184] Candidate content B (internal case library):

[0185] Semantic similarity: 0.92 (accurate match for "L'Oreal cosmetics case")

[0186] Time Decay: 0.9 (released 1 month ago)

[0187] Authority: 0.5 (Source: internal company report)

[0188] Total score: 0.6 × 0.92 + 0.3 × 0.9 + 0.1 × 0.5 = 0.87

[0189] Sorting results: Candidate content B (score 0.87) takes precedence over content A (score 0.78) because of its better semantic matching and timeliness.

[0190] Output information: Structured external data sets (including brand information, industry trends, and marketing activities). Data credibility score (0-1 points).

[0191] 3. Internal case data source engine (E→G→H)

[0192] Input information: structured parameter object, such as (brand, industry level, output format)

[0193] Processing process:

[0194] The internal case engine is the core module of the intelligent marketing webpage generation system. It is responsible for structured storage, intelligent retrieval, and dynamic scoring of the company's internal case library data to ensure that the output content is highly aligned with user needs. Its core processes include: knowledge base case library construction (scheduled tasks), semantic retrieval (based on industry tags and communication data), dynamic scoring (integrating semantic matching, industry associations, and communication effects), and interpretable evaluation generation (CoT reasoning chain).

[0195] 3.1、As Figure 4 As shown, the knowledge base case library is constructed.

[0196] A[Scheduled Task]-->B{Project System Case}

[0197] B-->C[Slice of Wenxin Yiyan Large Model]

[0198] C-->D[Structured Storage]

[0199] Input: original case text (unstructured data) or structured data in the enterprise project management system.

[0200] Processing process:

[0201] Slice of the Wenxin model:

[0202] Entity recognition: Extract brands, industry levels (primary / secondary), and marketing nodes (such as "KOL cooperation").

[0203] Relation extraction: Construct triple relationships (e.g., “L’Oréal-is-a-consumer-goods→cosmetics”).

[0204] Structured storage:

[0205] Generate case cards in Markdown format (including image links and dissemination data).

[0206] Stored in the Neo4j graph database (supporting industry path queries) and the Milvus vector library (supporting semantic retrieval).

[0207] Example: Original case text: "L'Oréal achieved a 120% sales increase and 500,000 Weibo shares during Singles' Day 2023 through Douyin KOL matrix advertising."

[0208] Processed structured data: json

[0209] {

[0210] "case_id":"CASE_001",

[0211] "brand":"L'Oréal",

[0212] "industry_level1":"Consumer goods",

[0213] "industry_level2":"cosmetics",

[0214] "case_img":"http: / / marketing-case.qiniudn.com / olay-campaign-2023.jpg",

[0215] "case_intro": "Douyin KOL matrix deployment achieved a 120% sales increase",

[0216] "spread_data":{"Weibo":500000,"WeChat":200000},

[0217] "semantic_vector":[0.12,-0.35,...] / / Sentence-BERT encoding vector

[0218] }

[0219] 3.2 Semantic Retrieval (User Command Driven)

[0220] Input: User instruction parameters (brand, industry level).

[0221] Processing process:

[0222] Industry path filtering: Query Neo4j for all cases whose secondary industry is "cosmetics." Use Jaccard similarity to verify brand-industry associations (e.g., "L'Oréal - Cosmetics" path matching ≥ 0.8).

[0223] Propagation data filtering:

[0224] Calculate the communication effect score: python

[0225] propagation_score = ∑(propagation amount × e^(-0.1 × Δt)) where Δt is the data time difference (months)

[0226] Screen candidate cases with scores >50.

[0227] Example:

[0228] User command parameters: {"brand":"L'Oreal","industry_level2":"cosmetics"}

[0229] Search results:

[0230] Case A: L’Oréal’s 2023 Double 11 (Double 11) Douyin campaign (Δt = 1 month, communication score = 85)

[0231] Case B: L’Oréal’s 2022 Xiaohongshu campaign (Δt = 6 months, communication score = 60)

[0232] 3.3 Dynamic Scoring (Large Model Driven)

[0233] Input: List of candidate cases.

[0234] Processing process:

[0235] Semantic matching score: Sentence-BERT is used to calculate the cosine similarity between the case summary and the user instruction.

[0236] Example: A user searches for "cosmetics marketing trends" and the summary similarity score for Case A is 0.92.

[0237] Industry matching score: based on the Jaccard similarity of industry paths (e.g., the path weight of “consumer goods → cosmetics” is 0.8).

[0238] Communication effect score: Time decay factor calculation: e^(-0.1×Δt) (score is 0.90 when Δt = 1 month).

[0239] Comprehensive scoring formula: python

[0240] final_score=0.6*semantic_sim+0.3*industry_sim+0.1*propagation_score

[0241] Example:

[0242]

[0243] 3.4. Interpretable Evaluation Generation (CoT Reasoning Chain)

[0244] Input: Scoring data of the top 5 cases.

[0245] Processing process:

[0246] Template filling: Generate evaluation reasons based on rating dimensions.

[0247] Logical chain verification: Ensure that the evaluation reasons are consistent with the scoring parameters.

[0248] Highlights analysis: The large model summarizes the highlights based on the content of the case itself

[0249] Sample output:

[0250] Case A highlights:

[0251] Highlights: 1. A comprehensive content marketing strategy covering multiple aspects of Douyin, including product selection and traffic generation. 2. Boosting brand exposure through high-quality short videos and attracting followers.

[0252] Case A evaluation reasons: 1. Industry match reached 98% (cosmetics secondary industry) 2. Weibo share volume exceeded 500,000 times in the past month (timeliness weight 0.90) 3. Contains a reusable Douyin KOL matrix strategy (semantic match 0.92)

[0253] Application scenario examples

[0254] User instructions:

[0255] "Please output the brand L'Oréal, the primary industry is consumer goods, and the secondary industry is cosmetics, in PDF format."

[0256] Internal case engine processing flow:

[0257] Case A (TikTok advertising) and Case B (Xiaohongshu promotion) were retrieved.

[0258] After scoring, Case A scored 0.87 and Case B scored 0.65.

[0259] When generating a PDF web page, prioritize inserting the images and strategy descriptions of Case A.

[0260] Output effect:

[0261] Case NamexxxCase Name

[0262] Case picture display

[0263] Industry: Personal Care FMCG - Cosmetics

[0264] Business Type: E-commerce Campaign Marketing

[0265] Background: Early product selection and content planning, regular high-quality short videos to attract fans, content diversion, and continuous fan retention through segmented editing

[0266] Highlights: 1. A comprehensive content marketing strategy, encompassing multiple aspects of Douyin, including product selection and traffic generation. 2. Highlights: Highlights: Highly engaging followers and increased brand exposure through high-quality short videos. Rating: "98% industry fit + 0.90 timeliness."

[0267] 4. Multi-modal output adapter (I)

[0268] Input information: external dataset (F output) + internal Top 5 cases (H output)

[0269] Processing process:

[0270] Format adaptation rule engine:

[0271]

[0272] Content integration: Use the BART model to integrate external data and case content to ensure logical coherence.

[0273] Output information: Markdown converted to multi-modal web page files (PDF / HTML / Word). File metadata (generation time, version number, data source identifier).

[0274] Experimental data support

[0275]

[0276] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. An intelligent marketing document generation device based on multi-source data fusion, characterized in that: include: User instruction input module: used to receive marketing document generation demand instructions submitted by users; NLP parsing engine module: connected to the user instruction input module, configured to perform natural language processing on the user instruction, extract structured parameters and verify the legitimacy of the parameters; Parameter decomposition module: connected to the NLP parsing engine module, used to split the parsed instructions into task parameters corresponding to external data acquisition and internal case calls; External data engine module: connected to the parameter disassembly module, configured to collect and process external industry data according to task parameters to generate semantically enhanced industry information; Internal case engine module: connected to the parameter decomposition module, configured to initiate case library search based on task parameters, screen matching marketing cases and output them; Multimodal output module: connected to the external data engine module and the internal case engine module respectively, used to integrate industry information and marketing cases, generate and output marketing documents in multiple formats; Data layer interface module: connected with the external data engine module and the internal case engine module, including a standardized API gateway, supporting RESTful / SOAP dual protocols to access heterogeneous data sources and realize data interaction; Knowledge graph storage unit: connected to the internal case engine module and NLP parsing engine module, using the Neo4j graph database to store entity relationships, covering the "brand-industry-case portrait" association form, for semantic association and case retrieval support; Core algorithm engine: including Hybrid generation engine: Connected with the multimodal output module and parameter decomposition module, it integrates the Wenxin Yiyan API call module and rule engine, supports dynamic adjustment of generation parameters, and is used for intelligent construction of marketing content; Case recommendation and evaluation engine: This engine connects to the internal case engine module, deploys a TF-IDF vector index library, supports millisecond-level similarity calculation for tens of millions of case libraries, and provides similarity evaluation for case library retrieval; Case library retrieval module: connected to the internal case engine module and the case recommendation and evaluation engine, relying on the semantic association of the knowledge graph storage unit and combining the similarity calculation results of the case recommendation and evaluation engine, to screen suitable cases from the case library.

2. The intelligent marketing document generation device based on multi-source data fusion according to claim 1 is characterized in that: The NLP parsing engine module uses a natural language processing model to perform word segmentation, entity recognition and intent understanding on user instructions. The natural language processing model includes a pre-trained language model. The external data engine module includes a data acquisition unit and a data processing unit, wherein the data acquisition unit is configured to call a third-party data source to obtain industry trends and market dynamics data; The data processing unit is configured to perform cleaning, semantic enhancement and importance scoring on the collected data.

3. The intelligent marketing document generation device based on multi-source data fusion according to claim 1 is characterized in that: When the internal case engine module initiates a case library search, it calls the TF-IDF vector index library capability of the case recommendation and evaluation engine, and combines the entity relationship association of the knowledge graph storage unit to achieve case screening.

4. The intelligent marketing document generation device based on multi-source data fusion according to claim 3 is characterized in that: The semantic matching of the case library retrieval module adopts a vector similarity calculation method to compare user instruction parameters with case feature vectors, and screen cases whose similarity meets the threshold. The vector similarity calculation is implemented based on the TF-IDF vector index library of the case recommendation and evaluation engine.

5. A method for generating intelligent marketing documents based on multi-source data fusion, characterized by: Production methods include: S1. User instruction parsing process, processing the original instruction to obtain structured parameters; S2. External data acquisition process, collecting and optimizing industry data based on structured parameters; S3. Internal case processing process, searching, scoring and generating evaluations for the enterprise case library; S4. Multimodal output process, converting the fused data into the target format document.

6. The method according to claim 5, characterized in that The S1 user instruction parsing process includes: S11. using regular expressions to clean the original instruction text, remove redundant symbols and unify the punctuation format; S12. Extract parameters such as brand, industry level, and output format through the model and generate a structured JSON object; S13. Verify the legitimacy of the brand-industry level based on the Neo4j knowledge graph and output the verification mark.

7. The method according to claim 5, characterized in that The S2 external data acquisition process includes: S21. Construct a search query based on the structured parameters, call Baidu's large model to obtain industry data, and use the similarity hash algorithm to remove duplicates; S22. Use the Wenxin model to perform TF-IDF cleaning, BERT-CLS semantic enhancement, and summary generation on the data; S23. Sort the data using the weighted formula: score = 0.6 × semantic_sim + 0.3 × time_decay + 0.1 × authority, where: semantic_sim is the semantic similarity (range: 0-1), which calculates the relevance between the text and user needs based on the model; time_decay is the time decay factor (range: 0-1), which is determined by the difference between the data release time and the current time; Authority is the authority score of the data source (ranging from 0 to 1), which is based on source levels such as industry white papers (0.7) and internal company reports (0.5). The top N best data are filtered by the score.

8. The method according to claim 5, characterized in that The S3 internal case processing process includes: Steps for building a knowledge base and case library: Slice the original cases using the Wenxin model through scheduled tasks, extract entity relationships, and store them in a structured manner in the Neo4j graph database and Milvus vector library. S32. Semantic retrieval step: Filter industry paths in Neo4j based on user instruction parameters, and select candidate cases based on the propagation data score propagation_score = ∑(propagation amount × e^(-0.1 × Δt)), where: The amount of dissemination refers to the dissemination data of the case on Weibo and WeChat platforms; Δt is the time difference between the data release time and the current time, in months; S33. Dynamic scoring step: using formula Final_score = 0.6 × semantic_sim + 0.3 × industry_sim + 0.1 × propagation_score to comprehensively score the candidate cases, where: industry_sim is the industry matching degree (range: 0~1), which is calculated based on the Jaccard similarity of industry paths; semantic_sim is the semantic similarity between the case summary and the user instruction (range: 0-1), calculated by the Sentence-BERT model; S34. Steps for generating explainable evaluation: Generate a CoT (thinking chain) and reasoning chain evaluation report based on the scoring dimensions, summarizing the highlights of the case.

9. The method according to claim 5, characterized in that The S4 multimodal output process includes: S41. Format Adaptation Step: Based on the output format, including PDF / HTML / Word, use Apache POI and pdfkit components to automatically convert Markdown to the target format; S42. Content integration step: Use the BART model to integrate external data and case content, ensuring a logical coherence score > 0.7; S43. Metadata generation step: output document metadata including data source identification and generation time.

Images