LLM-based industry map multi-modal report generation system

Abstract

The invention relates to the technical field of pre-training models, and discloses an LLM-based industry graph multi-modal report generation system, which comprises a multi-source data fusion module, a triple extraction module, a knowledge graph construction module and a multi-modal output module. According to the method, comprehensive coverage of multi-source heterogeneous news text data is realized in combination with a large language model and a traditional crawler technology; calling a large language model to generate a structured industry report based on the industry relation graph node features and the relation weights; complementing the key data through a networking data complementing tool, and automatically generating a dynamic chart in linkage with a visual tool format; meanwhile, an interactive industry knowledge graph is rendered, user exploration entity association is achieved, and multi-dimensional output of texts, charts and graphs is achieved; high automation from data acquisition and processing to report generation is realized, flexible arrangement of a networking data completion tool workflow is combined, the manual intervention cost is remarkably reduced, and the system response efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pre-training models, in particular to an industry graph multi-modal report generation system based on LLM. BACKGROUND

[0002] With the explosive growth of Internet information, enterprises need to quickly extract key information from massive news and industry reports to support decision-making. Traditional methods rely on manual collection, organization and analysis of data, which has problems such as low efficiency, strong subjectivity and poor real-time performance. Existing automated tools are mostly limited to single data source processing, lack deep fusion capabilities for multi-source data, and the generated reports are mostly static text, which is difficult to intuitively show industry trends and correlation.

[0003] The deficiencies of the prior art mainly lie in the following aspects: 1. Weak data integration capability: existing tools cannot efficiently process multi-source, multi-format news text data, leading to information fragmentation.

[0004] 2. Insufficient semantic understanding: traditional rules or shallow machine learning methods cannot deeply analyze the entity relationships in the text, resulting in low knowledge extraction accuracy.

[0005] 3. Poor interpretability and visualization: the generated report lacks intuitive charts and relationship graphs, making it difficult for users to quickly grasp industry trends and competitive patterns.

[0006] 4. Low degree of process automation: from data collection to report generation, multiple tools and methods are required, which cannot achieve end-to-end integrated processing.

[0007] Therefore, there is an urgent need for an industry analysis system that can automatically integrate multi-source data, deeply understand semantic associations, and implement multi-modal output. SUMMARY

[0008] The purpose of the present application is to provide an industry graph multi-modal report generation system based on LLM to solve the technical problems raised in the background.

[0009] To achieve the above purpose, the following technical solution is disclosed: an industry graph multi-modal report generation system based on LLM, comprising: A multi-source data fusion module configured to obtain multi-source heterogeneous news text data through large language model enhanced retrieval and traditional crawler technology, and perform data cleaning through a hybrid denoising strategy, the hybrid denoising strategy including rule cleaning and large language model self-supervised filtering; A triple extraction module configured to perform triple data dynamic extraction on the cleaned news text data based on the hierarchical architecture of the large language model, and eliminate redundancy and conflicts through semantic alignment technology; The knowledge graph construction module is configured to: import the triple data output by the triple extraction module into a graph database, and dynamically construct an industry relationship graph based on a graph structure analysis algorithm; The multi-modal output module is configured to: based on the industry relationship graph, call a large language model to generate a structured industry report, supplement key data through a networked data completion tool, generate dynamic charts through a visualization tool, and render an interactive industry knowledge graph based on a front-end framework to complete multi-modal collaborative output of the structured industry report, dynamic charts, and industry relationship graph.

[0010] Preferably, the multi-source data fusion module includes: The large model crawling unit is configured to: through multi-channel collection and automation process technology of the Wiseflow open source platform, directionally crawl news text data from multi-source channels including news websites, social media, industry forums, and authoritative reports; The crawler crawling unit is configured to: through XML path language and cascading style sheet selector, configure dynamic crawler rules to structure and parse news text data in hypertext markup language, portable document format, and application programming interface; The data cleaning unit is configured to: through a hybrid denoising strategy, perform data cleaning, which includes: Primary cleaning: through regular expressions and Python parsing libraries, remove web page tags, script codes, and advertisement templates; Fine denoising: based on the self-supervised learning capability of the large language model, guide the model to identify and filter noise segments through a prompt template.

[0011] Preferably, the data cleaning unit is further configured to: In the primary cleaning, generate noise segment classification labels through the large language model, and perform differential processing based on the classification results; In the fine denoising, through the rule engine and the large language model output results, dynamically adjust the denoising threshold to adapt to the text features of different industries.

[0012] Preferably, the triple extraction module includes: The topic extraction unit is configured to: based on a few-shot prompt, guide the large language model to identify core topics through a prompt template, and output a topic label list composed of multiple core topics; The relationship extraction unit is configured to: based on the topic label list, construct a triple structure, and through a prompt template, large language model parameter optimization, and few-shot prompt, extract triple relationship to obtain triple data; The result fusion unit is configured to: eliminate redundant, mismatched or incorrect triple data by performing context semantic matching on the cross-batch news text data through the large language model; and unify the topic labels of the multiple core topics extracted across batches by using the reasoning capability of the large language model, to obtain the final triple data and output the same.

[0013] Preferably, the topic extraction unit is further configured to: automatically generate a customized prompt template based on an industry keyword input by a user.

[0014] Preferably, the relation extraction unit is further configured to: The influence of different parameters of the large language model on the triple data extraction effect is verified through experiments, and the configuration of the parameters of the large language model is dynamically adjusted.

[0015] Preferably, the knowledge graph construction module comprises: The data import unit is configured to store the triple data through a Neo4j graph database, and to interact between Python and the Neo4j graph database through a py2neo software package. The structure modeling unit is configured to define entity nodes, attribute features and relationship edges between nodes, and to construct an industry relationship graph. The dynamic analysis unit is configured to perform correlation analysis on entities and their relationships in the industry relationship graph through a Cypher query language.

[0016] Preferably, the dynamic analysis unit is further configured to: Based on a community discovery algorithm, the clustering relationship of industry entities is identified, the importance of the relationship between entities is quantified through a page ranking algorithm or node degree centrality analysis, and the entity attributes and relationship edges in the industry relationship graph are regularly updated based on a timestamp field.

[0017] Preferably, the multi-modal output module comprises: The report generation unit is configured to intelligently analyze the triple data based on the large language model, and to generate a structured industry report through a prompt template. The data chart generation unit is configured to: perform directional network retrieval through a network data completion tool, and supplement key data; build a collaborative retrieval mechanism through a SearXNG open source search engine, to obtain missing time series data or key fields from external data sources; and generate dynamic charts including proportion pie charts and trend line charts through a visualization tool in a standard format. The front-end display unit is configured to: build a front-end platform based on a front-end framework, obtain an industry knowledge graph by dynamically rendering structured industry reports, dynamic charts and industry relationship graphs through a force-directed layout algorithm, and trigger an asynchronous request through natural language query to update graph data in the industry knowledge graph in real time.

[0018] As preferred, the front-end display unit is further configured to: screen nodes and edges in the industry relationship graph by industry category, time range or relationship strength, analyze a user's input natural language instruction through a pre-trained model and return matching subgraph data in the industry relationship graph in real time, and when the user clicks on any entity in the industry knowledge graph, synchronously highlight related content in the structured industry report and the corresponding dynamic chart.

[0019] Beneficial effects: The LLM-based industry graph multi-modal report generation system of the present application combines the information retrieval capability of a large language model with traditional crawler technology to achieve comprehensive coverage of multi-source heterogeneous news data; through dynamic completion of blind areas, the completeness and timeliness of data acquisition are improved; based on the node features and relationship weights of the industry relationship graph, a structured industry report is generated by calling a large language model; key data is completed through a networked data completion tool, and a visualization tool is linked to automatically generate proportion pie charts and trend line charts; an interactive industry knowledge graph is simultaneously rendered to enable users to explore entity associations and realize multi-dimensional output of text, charts and graphs; high automation is achieved from data acquisition, processing to report generation, combined with flexible arrangement of the workflow of the networked data completion tool, significantly reducing the cost of manual intervention and improving system response efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 The structural block diagram of the LLM-based industry graph multi-modal report generation system provided by the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0023] In this document, the term "comprising" is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0024] This embodiment discloses, as follows: Figure 1 This paper presents an LLM-based industry graph multimodal report generation system. It aims to achieve real-time analysis of industry dynamics, dynamic generation of relationship graphs, and automated output of multimodal reports through multi-source data integration, knowledge graph construction, and large language model collaboration. This helps users quickly grasp core industry information. The system includes a multi-source data fusion module, a triple extraction module, a knowledge graph construction module, and a multimodal output module.

[0025] Specifically The multi-source data fusion module is configured to: acquire multi-source heterogeneous news text data through large language model (LLM) enhanced retrieval and traditional crawling technology, and perform data cleaning through a hybrid denoising strategy, which includes rule cleaning and large language model self-supervised filtering.

[0026] In this embodiment, the multi-source data fusion module includes: The large-scale crawling unit is configured to: utilize the multi-channel collection and automated workflow technology of the Wiseflow open-source platform to selectively crawl news text data from multiple sources, including news websites, social media, industry forums, and authoritative reports. During the crawling process, natural language processing technology is used to parse the structure of the target websites, dynamically adapting to the HTML layout of different websites to avoid crawling failures due to website redesigns. For example, for social media (such as Twitter), industry-related tweets with hashtags are crawled in real time via API interfaces; for industry forums (such as Zhihu), keyword matching strategies are used to crawl technical discussion content from highly-rated answers. Wiseflow is an open-source AI information mining tool designed to help users extract valuable information from massive amounts of data. The crawler unit is configured to configure dynamic crawler rules through XML Path Language (XPath, a language in the prior art for finding information in XML documents) and Cascading Style Sheets (CSS, a computer language in the prior art for describing the appearance and format of HTML or XML documents) selectors, to perform structured parsing and crawling of news text data from HyperText Markup Language (HTML, a standard markup language in the prior art for creating web pages), Portable Document Format (PDF, a file format in the prior art for presenting documents independently of applications, hardware, and operating systems), and Application Programming Interface (API interface), for example, for HTML pages, the is located through XPath to extract the body content; for PDF documents, PyPDF2 or PDFMiner is used to extract text and regular expressions are used to remove header and footer noise; for API interfaces (such as RESTful APIs of news websites), by constructing a request URL containing parameters such as time range and keywords, the data is obtained page by page and stored in JSON format; The data cleaning unit is configured to clean data through a hybrid denoising strategy, which includes: Primary cleaning: remove web page tags, script codes, and ad templates through regular expressions and Python parsing libraries (such as BeautifulSoup); for example, parse HTML through the BeautifulSoup library of Python to extract the text content under the tags; Fine denoising: based on the self-supervised learning ability of large language models (LLM, a deep learning model in the prior art trained using a large amount of text data), the model is guided to identify and filter noise segments (such as repeated content and non-industry-related text) through a prompt template (Prompt), for example: input prompt template: "The following is a news text, please mark out the advertising content or irrelevant information: [news text]", the model outputs the classification label of the noise segment (such as "advertising" and "repeated content").

[0027] Based on the design of the above multi-source data fusion module, through multi-source channel crawling and dynamic parsing technology, the diversity and integrity of the data source are ensured, unstructured data that cannot be processed by traditional crawlers are covered, and combined with rule cleaning and LLM self-supervised filtering, the proportion of noise data is significantly reduced, and the accuracy of subsequent triple extraction is improved.

[0028] Further, the data cleaning unit is further configured to: In the primary cleaning, the large language model generates noise segment classification labels (such as "advertising content", "redundant information", "irrelevant text", etc.), and based on the classification results, differential processing is carried out, including, for example: discarding "advertising" content directly, replacing "meaningless content" with placeholders after fuzzy matching, and retaining "effective content" and entering the next stage of processing; In the fine denoising, the rule engine and the large language model output results are used to dynamically adjust the denoising threshold to adapt to the text features of different industries. The rule engine generates initial denoising rules based on industry characteristics (such as a medical industry term library); the LLM outputs correction suggestions through context understanding (such as identifying "side effects of a certain drug" as effective content and "promotion information" as noise); the system integrates the outputs of the rule engine and the LLM to dynamically adjust the threshold (such as adjusting the noise confidence threshold from 0.8 to 0.7).

[0029] Based on this, the data cleaning unit dynamically adjusts the denoising threshold to adapt to the text features of different industries (such as professional terms in the financial industry and informal expressions on social media), improving cleaning efficiency and accuracy.

[0030] The triple extraction module is configured to dynamically extract triple data (i.e. entity-relation-entity) from the cleaned news text data based on the hierarchical architecture of the large language model, and eliminate redundancies and conflicts through semantic alignment technology.

[0031] In this embodiment, the triple extraction module includes: The topic extraction unit is configured to guide the large language model through a prompt template to identify core topics (such as the keywords "semiconductor supply chain" and "new energy policy") based on a few-shot prompt, and output a topic label list composed of multiple core topics. An example of identifying core topics can be, for example: input prompt template: "The following is a news text, please extract the core topics and generate a topic label list: [news text]", the model outputs results similar to ["artificial intelligence", "medical industry", "policy support"]; The relationship extraction unit is configured to construct a triple structure [entity A, relationship, entity B] based on the topic label table, and extract triple relationships through a prompt template, large language model parameter optimization, and few-shot prompt to obtain triple data. An example of extracting triple data can be, for example: prompt template example: "Please generate triples based on the following topic labels: [topic labels], and describe their relationships."

[0032] The result fusion unit is configured to: eliminate redundant, mismatched or incorrect triple data by performing context semantic matching on the cross-batch news text data (i.e., calculating the semantic similarity of the cross-batch news text data) through the large language model; and unify the topic labels of the multiple core topics extracted across batches (e.g., unify "AI medical treatment" and "artificial intelligence medical treatment" as "AI medical treatment") using the reasoning ability of the large language model, to obtain the final triple data and output.

[0033] Based on the design of the above-mentioned triple extraction module, the few-shot prompting and parameter optimization can improve the accuracy and coverage of triple data extraction, reduce the cost of manual intervention, and the design of semantic matching and topic label unification can ensure the consistency of cross-batch data and avoid conflicts in the industry relationship graph.

[0034] Further, the topic extraction unit is further configured to: Based on the user input industry keywords, automatically generate a customized prompt template, and the LLM generates industry-related topic labels (e.g., ["blockchain payment", "digital currency regulation", "risk control model"]) according to the template to improve the accuracy of triple data extraction.

[0035] It should be noted that the customized prompt template refers to dynamically generating structured prompt sentences according to user input specific industry keywords (e.g., "photovoltaic industry" and "chip manufacturing") or task requirements to guide the large language model to more accurately complete the triple data extraction task. The goal is to: through pre-defined template structure and variable filling, convert user requirements into LLM understandable instructions to improve the accuracy and consistency of model output. In the implementation of customization, first, parameterize the key information (i.e., insert the user input keywords (e.g., "photovoltaic industry") into the template as variables, identify the corresponding industry keywords (e.g., "new energy policy" and "semiconductor supply chain"), and output fixed formats (e.g., "JSON", "Markdown list", etc.) as task goals such as "entity recognition" and "relation extraction"), then perform template structuring (i.e., the template needs to include clear task description and format requirements to reduce ambiguity), and finally, complete the entire customization process through context injection (i.e., add industry background or domain knowledge to the template to help the model understand the task scope). In addition, the automatic generation of the customized template uses LLM to generate according to user input, which includes the following steps: Input analysis: input the user input industry keywords (e.g., "new energy vehicles") and task requirements (e.g., "extract industry chain relationships") as input; Template generation: LLM generates a structured prompt template according to the training data; Template optimization: optimize the accuracy and practicality of the template through iterative verification (e.g., manual evaluation or model feedback).

[0036] Based on this, the topic extraction unit improves the industry adaptability of topic extraction by customizing the prompt template, avoiding the topic deviation caused by the general template.

[0037] Secondly, the relation extraction unit is further configured to: Through experiments, the influence of different large language model parameters (such as temperature coefficient, maximum generation length) on the extraction effect of triple data is verified, and the configuration of large language model parameters is dynamically adjusted. For example: design a comparative experiment: fix other parameters in LLM, adjust the temperature value (such as 0.5, 0.7, 0.9), and evaluate the F1 score of triple. Based on the experimental results, the optimal parameter combination (such as temperature value 0.7, maximum output length 256) is automatically selected.

[0038] Based on this, the relation extraction unit improves the precision and recall rate of triple data extraction through the design of parameter optimization, and adapts to the complexity requirements of different news texts.

[0039] The knowledge graph construction module is configured to: import the triple data output by the triple extraction module into the graph database, and dynamically construct the industry relationship graph based on the graph structure analysis algorithm.

[0040] In this embodiment, the knowledge graph construction module includes: The data import unit is configured to store triple data through the Neo4j graph database (i.e. a high-performance NoSQL graph database in the prior art, which stores structured data on a graph rather than in a table), and interact between Python and the Neo4j graph database through the py2neo software package, that is, use the py2neo library (i.e. Python interface) of the Neo4j graph database to import triple data in batches; The structure modeling unit is configured to define entity nodes (such as "AI medical care"), attribute features (such as "hotness": 0.8), and relationship edges (such as "related") between nodes, and construct an industry relationship graph; The dynamic analysis unit is configured to perform association analysis on entities and their relationships in the industry relationship graph through the Cypher query language. It should be noted that the association analysis involves the following contents: 1. Direct association between entities, that is, entities (such as enterprises, technologies, policies, products, etc.) in the industry relationship graph are directly connected through explicit relationship edges (such as "investment", "cooperation", "dependence", etc.). Users can dynamically query the direct relationship between two entities (such as "enterprise A→investment→enterprise B") through the Cypher query language, and when the entity attributes or relationship edges in the industry relationship graph change (such as adding a cooperation relationship), the association analysis result reflects these changes in real time.

[0041] 2. Association of multi-step paths, i.e. through the path query function of the graph database, analyze the associated paths of multi-hop relationships between entities (such as "Enterprise A → depends on → Technology X → applied to → Product Y"). Users can filter paths by path length, relationship type (such as "technology dependence" "supply chain") or time range (such as "newly added relationships in 2023") through the Cypher query language, and can adjust the query conditions according to user needs (such as "only display associations with path length ≤ 3"), and return the matching results in real time.

[0042] Based on the above, the knowledge graph construction module ensures real-time dynamic updating of the industry knowledge graph through the efficient graph storage and Cypher query capability of Neo4j, thereby realizing rapid analysis of large-scale industry relationships.

[0043] Further, the dynamic analysis unit is also configured to: Based on community discovery algorithms (such as the algorithm for discovering community structure in social networks (Girvan-Newman algorithm) in the prior art or the efficient algorithm for community discovery (Louvain algorithm) in the prior art), identify the clustering relationship of industry entities (i.e. input nodes and edges in the industry knowledge graph, output clustering results (such as cluster 1 contains "AI medical", "blockchain payment", and cluster 2 contains "traditional finance")), which are used for industry segmentation analysis (such as identifying the cross-field of medical AI and financial technology)), quantify the importance of relationships between entities through page ranking algorithms or node degree centrality analysis (i.e. input the adjacency matrix of the industry knowledge graph, output the PageRank value (such as the PageRank value of "AI medical" is 0.95, higher than that of "traditional finance" 0.65), and the PageRank value is used to recommend core entities (such as recommending the entity with the highest PageRank value to the user)), and based on the timestamp field, update the entity attributes and relationship edges in the industry relationship graph periodically (i.e. update the entity attributes through the timestamp field every period (such as the "hotness" value of the data is updated according to the exponential decay formula: hotness = 0.9 * original hotness + 0.1 * new data hotness)).

[0044] Based on this, the dynamic analysis unit provides multi-dimensional industry relationship analysis capability through clustering relationships and the importance of relationships between entities, so that decision makers can identify key trends and potential risks.

[0045] The multi-modal output module is configured to: based on the industry relationship graph, call a large language model to generate a structured industry report, supplement key data through a networked data completion tool, and generate dynamic charts through a visualization tool, and render an interactive industry knowledge graph based on a front-end framework, to complete the multi-modal collaborative output of the structured industry report, dynamic charts and industry relationship graph.

[0046] In this embodiment, the multi-modal output module comprises: A report generation unit configured to intelligently parse the triple data based on a large language model, and generate a structured industry report (covering market pattern, technology trend, competition situation, risk factors, etc. topic dimensions) through a prompt template; for example: the prompt template example is: "Please generate an industry report on AI medical according to the following triple data, including market trends, technical challenges and policy recommendations.", the LLM outputs a structured report (such as a chaptered Markdown format, including headings, sections, data references); A data chart generation unit configured to: a network data completion tool (such as the BOCHA tool of the open source platform Dify for building AI applications in the prior art) performs directional network retrieval and supplements key data (mainly some scenarios need to supplement unstructured or non-public data (such as enterprise internal data, sensitive industry data)); through a SearXNG open source search engine to build a collaborative retrieval mechanism, obtain missing time series data or key fields from external data sources (such as public databases, industry reports, academic papers, etc.); and through a visualization tool (such as the open source visualization library ECharts implemented using JavaScript in the prior art) to generate dynamic charts in a standard format, the dynamic charts include a proportion pie chart and a trend line chart; A front-end display unit configured to: build a front-end platform based on a front-end framework (such as Django), dynamically render the structured industry report, dynamic chart and industry relationship graph through a force-directed layout algorithm to obtain an industry knowledge graph, and trigger an asynchronous request through natural language query to update the graph data in the industry knowledge graph in real time.

[0047] Based on the above, the multi-modal output module improves the intuitiveness and operability of information transmission through multi-modal output of text + chart + interactive graph, and meets the user's demand for diversified reports.

[0048] Further, the front-end display unit is further configured to: Filter the nodes and edges in the industry relationship graph by industry category, time range or relationship strength (such as "PageRank>0.8"); analyze the user's input natural language instruction (such as "Display 2023 new energy automobile industry chain relationship") through a pre-trained model (such as a bidirectional encoder representation model BERT, etc.), and return the matching subgraph data in the industry relationship graph in real time; and when the user clicks on any entity in the industry knowledge graph, the related content in the structured industry report and the corresponding dynamic chart are highlighted synchronously.

[0049] Based on this, the front-end display unit improves user experience through natural language interaction and real-time updating, enabling users to quickly obtain the required information and dynamically adjust the analysis perspective.

[0050] In summary, the LLM-based industry graph multi-modal report generation system of the embodiment has the following technical features: (1) Multi-source data fusion and enhanced retrieval: Through the combination of multi-channel collection technology of Wiseflow open source platform and traditional crawlers, efficient crawling of multi-source heterogeneous data is realized, and the data blind area of single technology is made up; A hybrid denoising strategy is adopted. Regular expressions and Python parsing libraries are used to remove HTML tags, script codes, and ad templates. Then, LLM prompt templates are used to identify noise segments, significantly reducing the noise ratio (such as ad content filtering rate > 95%) and improving the accuracy and generalization ability of data cleaning; The denoising threshold is dynamically adjusted based on the rule engine and LLM output results to adapt to the text features of different industries such as finance, healthcare, and technology.

[0051] (2) Dynamic extraction of triplets based on hierarchical LLM architecture LLM is guided by a few sample examples to identify core topics and output a list of topic labels, improving the industry-specificity of extraction; Based on the topic labels, a triplet structure is constructed, combining prompt templates, parameter optimization, and a few sample examples to achieve dynamic extraction across different industry scenarios; LLM's context semantic matching capability is used to eliminate redundant or incorrect triplets, and reasoning capability is used to unify cross-batch topic labels; The influence of different LLM parameters on extraction results is verified through experiments, and dynamic adjustment is made to balance diversity and accuracy.

[0052] (3) Dynamic knowledge graph construction and analysis Triple data is stored in Neo4j graph database, and py2neo software package is used to realize the interaction between Python and Neo4j, realizing the rapid construction of large-scale industry relationship graph; Community discovery algorithm identifies the clustering relationship of industry entities, assisting users in discovering potential submarkets; Through page ranking algorithm or node degree centrality, the importance of the relationship between entities is analyzed to identify key trends; Based on the timestamp field, entity attributes are updated regularly to ensure the timeliness of the graph.

[0053] (4) Multi-modal output and interactive visualization LLM-based prompt templates intelligently parse triple data, generating structured reports containing market trends, technical challenges, policy recommendations, and other sections, improving information density and readability. BOCHA tool on Dify platform for targeted network search of missing data, combined with SearXNG search engine for collaborative search mechanism to solve the limitations of single data source. Based on Django framework, the front-end platform is built, and the force-directed layout algorithm is used to dynamically render the industry knowledge graph, providing node dragging, scaling, and relationship path highlighting functions. After the user inputs the instruction, the pre-trained model analyzes the intent and triggers an asynchronous request, returning the matching subgraph data in real time. When the user clicks on an entity in the graph, the related content in the report and the corresponding dynamic chart are highlighted simultaneously.

[0054] (5) End-to-end automation and process optimization From data collection, cleaning, triple extraction to report generation, the networking data completion tool workflow is used to achieve flexible arrangement and reduce manual intervention. Each module can be optimized or replaced independently (such as upgrading the LLM model to a larger parameter version), improving the system's scalability and adaptability. Through asynchronous requests and caching mechanisms, real-time user interaction is achieved, reducing query response time.

[0055] Compared with traditional methods, this system improves data integrity through LLM-enhanced search + dynamic crawler, improves semantic understanding accuracy through hierarchical LLM architecture, improves interpretability through multi-modal output, and improves efficiency through end-to-end automation. Secondly, this system supports the customization needs of multiple industries and, through clustering analysis and page sorting quantification of industry relationship graphs, assists enterprises in quickly identifying market opportunities and risks. At the same time, combined with LLM's dynamic parameter optimization, semantic alignment, and community discovery algorithms, it leaves room for future expansion. Finally, this system improves the efficiency, accuracy, and visualization level of industry analysis, providing strong technical support for supply chain management, financial decision-making, and other fields.

[0056] In the embodiments provided by the present application, it should be understood that the embodiments described herein can be realized by hardware, software, firmware, middleware, codes or any proper combination thereof. For hardware implementation, the processor can be realized in one or more of the following components: an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a processor, a controller, a microcontroller, a microprocessor, other electronic units designed to perform the functions described herein, or a combination thereof. For software implementation, the procedures described herein can be implemented with a computer program that is written in any suitable programming language. The program can be stored in a computer readable storage medium or transmitted as one or more instructions or codes on the computer readable storage medium. The computer readable storage medium includes any storage medium that can be accessed by a computer. The computer readable storage medium can include but is not limited to the following media: a RAM, a ROM, an EEPROM, a CD-ROM or other optical disc storage, a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program codes in the form of instructions or data structures and can be accessed by a computer.

[0057] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, modifications or equivalent replacements of some technical features described in the foregoing embodiments can be made by those skilled in the art, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An LLM-based industry graph multi-modal report generation system, characterized in that, The method comprises the following steps: a multi-source data fusion module configured to obtain multi-source heterogeneous news text data through large language model enhancement retrieval and traditional crawler technology, and perform data cleaning through a hybrid denoising strategy, wherein the hybrid denoising strategy comprises rule cleaning and large language model self-supervised filtering; a triple extraction module configured to perform triple data dynamic extraction on the cleaned news text data based on a hierarchical architecture of a large language model, and eliminate redundancy and conflicts through semantic alignment technology; a knowledge graph construction module configured to import the triple data output by the triple extraction module into a graph database, and dynamically construct an industry relationship graph based on a graph structure analysis algorithm; a multi-modal output module configured to generate a structured industry report based on the industry relationship graph, supplement key data through a networked data completion tool, generate dynamic charts through a visualization tool, and render an interactive industry knowledge graph based on a front-end framework to complete multi-modal collaborative output of the structured industry report, the dynamic charts and the industry relationship graph.

2. The LLM-based industry graph multi-modal report generation system of claim 1, wherein, The multi-source data fusion module comprises: a large model crawling unit configured to directionally crawl news text data from multi-source channels through multi-channel collection and automatic process technology of the Wiseflow open source platform, wherein the multi-source channels comprise news websites, social media, industry forums and authoritative reports; a crawler crawling unit configured to configure dynamic crawler rules through XML path language and cascading style sheet selectors, and structure and crawl news text data from hypertext markup language, portable document format and application programming interface data; a data cleaning unit configured to perform data cleaning through a hybrid denoising strategy, wherein the data cleaning comprises: primary cleaning: removing web page tags, script codes and advertisement templates through regular expressions and Python parsing libraries; fine denoising: based on the self-supervised learning ability of the large language model, guiding the model to identify and filter noise segments through a prompt template.

3. The LLM-based industry graph multi-modal report generation system of claim 2, wherein, The data cleaning unit is further configured to: generate noise segment classification labels through the large language model in the primary cleaning, and perform differential processing based on the classification results; in the fine denoising, dynamically adjust the denoising threshold to adapt to the text features of different industries through the output results of the rule engine and the large language model.

4. The LLM-based industry graph multi-modal report generation system of claim 1 or 2, wherein, The triple extraction module comprises: a topic extraction unit configured to guide the large language model to identify core topics through a prompt template based on a few-shot prompt, and output a topic label list composed of multiple core topics; a relationship extraction unit configured to construct a triple structure based on the topic label list, and perform triple relationship extraction through a prompt template, large language model parameter optimization and few-shot prompt to obtain triple data; a result fusion unit configured to perform context semantic matching on cross-batch news text data through the large language model to eliminate redundant, mismatched or incorrect triple data, and unify the topic labels of multiple core topics extracted across batches using the reasoning ability of the large language model to obtain and output final triple data.

5. The LLM-based industry graph multi-modal report generation system of claim 4, wherein, The subject extraction unit is further configured to: Automatically generate a customized prompt template based on the industry keyword input by the user.

6. The LLM-based industry graph multi-modal report generation system of claim 4, wherein, The relationship extraction unit is further configured to: Verify the influence of different large language model parameters on the extraction effect of triple data through experiments, and dynamically adjust the configuration of the large language model parameters.

7. The LLM-based industry graph multi-modal report generation system of claim 1, wherein, The knowledge graph construction module includes: A data import unit configured to store triple data through a Neo4j graph database, and interact between Python and the Neo4j graph database through a py2neo software package; A structure modeling unit configured to define entity nodes, attribute features, and relationship edges between nodes, and construct an industry relationship graph; A dynamic analysis unit configured to perform correlation analysis on entities and their relationships in the industry relationship graph through a Cypher query language.

8. The LLM-based industry graph multi-modal report generation system of claim 7, wherein, The dynamic analysis unit is further configured to: Identify the clustering relationship of industry entities based on a community discovery algorithm, quantify the importance of the relationship between entities through a page ranking algorithm or node degree centrality analysis, and regularly update the entity attributes and relationship edges in the industry relationship graph based on a timestamp field.

9. The LLM-based industry graph multi-modal report generation system of claim 1, wherein, The multi-modal output module includes: A report generation unit configured to intelligently analyze triple data based on a large language model, and generate a structured industry report through a prompt template; A data chart generation unit configured to perform targeted network retrieval through a network data completion tool, and supplement key data; build a collaborative retrieval mechanism through a SearXNG open source search engine to obtain missing time series data or key fields from external data sources; and generate dynamic charts including proportion pie charts and trend line charts through a visualization tool standard format; A front-end display unit configured to build a front-end platform based on a front-end framework, dynamically render the structured industry report, dynamic charts, and industry relationship graph through a force-directed layout algorithm to obtain an industry knowledge graph, and trigger asynchronous requests through natural language queries to update the graph data in the industry knowledge graph in real time.

10. The LLM-based industry graph multi-modal report generation system of claim 9, wherein, The front-end display unit is further configured to: Filter the nodes and edges in the industry relationship graph through industry categories, time ranges, or relationship strengths; analyze the natural language instructions input by the user through a pre-training model, and return the matching subgraph data in the industry relationship graph in real time; and when the user clicks on any entity in the industry knowledge graph, highlight the related content in the structured industry report and the corresponding dynamic charts.

Images