Intelligent analysis system for initially created enterprise based on large language model

Abstract

The invention provides a big language model-based intelligent analysis system for a newly created enterprise, and the system comprises a data collection module which is used for collecting data of the newly created enterprise; the data preprocessing module is used for preprocessing the data of the initially created enterprise; the vector database and knowledge graph construction module is used for constructing a vector database and a knowledge graph based on the preprocessed initially created enterprise data; the context information construction module is used for carrying out deep semantic analysis on related information, input by a user, of the to-be-analyzed enterprise, and fusing a semantic analysis result with data of the vector database and the knowledge graph to obtain context information; and the intelligent analysis module is used for carrying out intelligent analysis on the context information based on the initially created enterprise research analysis auxiliary large model, obtaining development state data of the to-be-analyzed enterprise, and completing initially created enterprise intelligent analysis based on the large language model. And comprehensive and efficient retrieval and processing of related data of the initially created enterprise are realized.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to an intelligent analysis system for start-up enterprises based on a large language model. Background Art

[0002] With the rapid development of artificial intelligence technology, especially in the fields of natural language processing and deep learning, a large number of research and applications focus on how to use these technologies for data analysis and business analysis. In the research and analysis of start-ups, industry analysts, market researchers, research analysts and other personnel need to process a large amount of unstructured data, such as market reports, industry news, technical literature, patent data, etc. However, existing technical means have many problems in processing this data.

[0003] Existing startup analytics tools have the following key shortcomings:

[0004] Limited data processing capabilities: Faced with large amounts of unstructured data, existing tools are unable to effectively extract valuable information, resulting in insufficient accuracy and comprehensiveness of analysis results.

[0005] Low analysis efficiency: Traditional methods rely on manual analysis, which is time-consuming and difficult to meet the needs of rapid analysis.

[0006] These technical deficiencies make it difficult for researchers to conduct comprehensive and accurate analysis when faced with large amounts of data. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present invention provides an intelligent analysis system for start-up enterprises based on a large language model to improve data processing efficiency and analysis quality.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] An intelligent analysis system for startups based on a large language model, comprising:

[0010] Data collection module, used to collect data on startups;

[0011] A data preprocessing module, used for preprocessing the start-up enterprise data;

[0012] A vector database and knowledge graph construction module, used to construct a vector database and a knowledge graph based on the preprocessed startup data;

[0013] A context information building module is used to perform deep semantic analysis on the relevant information of the enterprise to be analyzed input by the user, and fuse the semantic analysis results with the data of the vector database and the knowledge graph to obtain context information;

[0014] The intelligent analysis module is used to perform intelligent analysis on the context information based on the large model for assisting the research and analysis of start-ups, obtain the development status data of the enterprise to be analyzed, and complete the intelligent analysis of start-ups based on the large language model.

[0015] Preferably, the start-up enterprise data includes market report data, patent data, technical literature data, industry news data and technology company data.

[0016] Preferably, the data preprocessing module includes:

[0017] A data parsing unit, used to parse the start-up enterprise data to obtain structured text information; wherein the parsing includes PDF parsing, HTML parsing, XML / JSON parsing and table data processing;

[0018] A data cleaning unit, used for denoising and formatting the structured text information to obtain format-standardized data;

[0019] The text standardization unit is used to unify the case of format-standardized data, filter stop words, and perform word form restoration and stem extraction to obtain text-standardized data.

[0020] Preferably, the vector database and knowledge graph construction module includes:

[0021] A knowledge graph construction unit, used to perform entity recognition and relationship extraction on the text standardization data to construct a dynamically updated knowledge graph;

[0022] The vector database construction unit is used to perform vectorization processing and optimization on the text standardization data to obtain the optimized text embedding vector, and perform vector retrieval on the optimized text embedding vector based on the hierarchical navigable small-world algorithm to complete the construction of the vector database.

[0023] Preferably, the process of obtaining the text embedding vector is:

[0024] Processing the text normalization data based on the Ada model to obtain a text embedding vector;

[0025] Based on the self-attention mechanism combined with the loss function, the text embedding vector is linearly transformed to obtain the query Q, key K and value vector V; and the dot product between the query Q and the key K is calculated to obtain the similarity between different parts of the text;

[0026] Use the Softmax function to convert the similarity between different parts of the text into attention weights;

[0027] The attention weight and the value vector V are weightedly summed to obtain an optimized text embedding vector.

[0028] Preferably, the process of performing vector retrieval on the optimized text embedding vector based on the hierarchical navigable small-world algorithm is:

[0029] Use the optimized text embedding vectors as nodes to build a sparse hierarchical graph structure;

[0030] Inserting a new optimized text embedding vector to update the sparse hierarchical graph structure;

[0031] Based on layer-by-layer navigation, a local to global search is performed on the updated sparse hierarchical graph structure nodes to obtain a storage vector that is closest to the user query vector.

[0032] Preferably, the context information building module includes:

[0033] A semantic analysis unit, which is used to perform in-depth semantic analysis on the relevant information of the enterprise to be analyzed input by the user based on a large language model to obtain a semantic analysis result;

[0034] A knowledge graph query unit, used to extract the opponent data of the enterprise to be analyzed in the knowledge graph based on the semantic analysis result;

[0035] A vector database retrieval unit, configured to retrieve industry dynamic data related to the enterprise to be analyzed from the vector database based on the semantic analysis result;

[0036] The data fusion unit is used to reconstruct data by taking the relevant information of the enterprise to be analyzed as the main context and taking the competitor data and the industry dynamic data as supplementary information to obtain the context information.

[0037] Preferably, the intelligent analysis module includes:

[0038] The model building unit is used to fine-tune the large language model based on relevant data in the field of the enterprise to be analyzed, and obtain a large model to assist in the research and analysis of start-ups;

[0039] The intelligent analysis unit is used to analyze the context information based on the start-up enterprise research and analysis auxiliary model to obtain the development status data of the enterprise to be analyzed.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] (1) Improve data processing and analysis efficiency

[0042] Through the data collection and preprocessing modules, the present invention can efficiently clean, format and extract features from raw data from multiple sources. Compared with traditional manual processing and manual sorting, the system can process large amounts of data in a shorter time, significantly reducing the time cost of data preparation. Compared with manual entry, the data processing efficiency of this system is increased by more than 50%, and the processing time of relevant data for each company is reduced from an average of 2 hours to less than 15 minutes.

[0043] (2) Provide highly accurate and logical intelligent analysis

[0044] Improved accuracy and logic: By fine-tuning GPT-4 or Zhipu GLM-4, a large model to assist in research and analysis of startups is created. The system can conduct in-depth searches on market structures, technological development contexts, and competitive conditions.

[0045] (3) Wide range of industry applicability

[0046] This system is suitable for a variety of high-tech industries, especially companies involved in technological innovation, such as artificial intelligence, enterprise software, virtual reality, etc. In the test, the system was able to adapt to the technology and market characteristics of different fields, helping researchers to quickly understand the market and enterprises in the field.

[0047] (4) Economic and social benefits

[0048] This invention not only provides powerful technical support for industry analysts and market researchers, but also has the potential to promote the application of artificial intelligence technology in the field of enterprise research and analysis in the future, with significant economic and social benefits. By improving the accuracy and comprehensiveness of research and analysis of start-ups, the system helps researchers better understand high-potential technology companies and promote technological innovation and economic development. In the long run, this research behavior will help promote the development of emerging technologies and promote industrial upgrading. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0050] Figure 1 This is a schematic diagram of the structure of an intelligent analysis system for start-ups based on a large language model according to an embodiment of the present invention;

[0051] Figure 2 This is a workflow diagram of a context information construction module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Embodiment 1

[0055] like Figure 1 , Figure 2 As shown, an intelligent analysis system for startups based on a large language model includes: a data collection module, a data preprocessing module, a vector database and knowledge graph construction module, a context information construction module and an intelligent analysis module.

[0056] The data collection module is used to collect startup data; in a further embodiment, the startup data includes market report data, patent data, technical literature data, industry news data, and technology company data. The module obtains data from multiple sources through API interfaces, web crawler technology, or manual data import functions, and supports data in different formats.

[0057] In this embodiment, market report data is obtained through APIs (such as Bloomberg API, Thomson Reuters API), including global market trend data, company financial data, existing industry analysis data, etc. Global patent data is obtained through Google Patents, Chinese patent data and Espacenet. Technical literature data is obtained through IEEE Xplore, ArXiv and ScienceDirect; industry news data is obtained through Gartner (market research reports in the IT field, including technology trends and market forecasts), IDC (providing market data, analysis and forecasts for the IT industry) and Forrester (providing market research, industry analysis and technology development trend reports); technology company data is obtained through Crunchbase website (providing company information, financing status, market dynamics and other data), PitchBook website (detailed company financial, market and transaction information database) and CBInsights website (providing industry trends, company comparison, technology development status and other data). Specific data acquisition method: API calls are made by using Python's requests library, and Web crawlers are performed by methods such as BeautifulSoup and Scrapy libraries.

[0058] The data preprocessing module is used to preprocess the startup data. The main function of the data preprocessing module is to normalize, clean, and convert the original and diverse data to provide high-quality and actionable data for the subsequent vector database and knowledge graph construction. Unprocessed raw data usually contains noise, inconsistency, redundancy and other problems. Direct use of this data will reduce the overall analysis capability and accuracy of the system. Therefore, the data preprocessing module ensures that subsequent modules can effectively use the data for in-depth analysis.

[0059] A further implementation is that the data preprocessing module includes:

[0060] A data parsing unit is used to parse the startup data to obtain structured text information; the parsing includes PDF parsing, HTML parsing, XML / JSON parsing and table data processing;

[0061] Specifically, PDF parsing: Use libraries such as pdfplumber or PyMuPDF to parse PDF files, extract text, tables, images, and other content, and separate related page and paragraph information.

[0062] HTML parsing: Parse HTML documents through libraries such as BeautifulSoup or lxml, extract information such as text, title, links, and remove useless elements such as advertisements and navigation.

[0063] XML / JSON parsing: For structured files such as patents and company data, use xml.etree.ElementTree or json library to parse and extract target fields such as invention name, inventor, company financial data, etc.

[0064] Tabular data processing: For CSV or Excel files, use the pandas library to read them directly to ensure that the data format is standardized.

[0065] The data cleaning unit is used to denoise and standardize the format of structured text information to obtain format-standardized data.

[0066] Specifically, remove useless characters and symbols: remove HTML tags, special characters, non-ASCII characters and other irrelevant information.

[0067] Text Deduplication: Detect and remove duplicate data, especially similar news articles or the same company data extracted multiple times.

[0068] Missing value processing: For text data, prediction can be made based on the context.

[0069] Format standardization:

[0070] Time format: All timestamps are unified into the ISO standard format (YYYY-MM-DD).

[0071] Currency unit: Convert all financial data involved (such as financing amount, income) into the same currency unit (such as USD, RMB) and take into account exchange rate adjustments.

[0072] Standardization of technical classification and field labels: For example, technical fields in patent data need to be normalized according to predefined classification standards.

[0073] The text standardization unit is used to unify the case of format-standardized data, filter stop words, and perform word form restoration and stem extraction to obtain text-standardized data.

[0074] Specifically, capitalization is unified: convert all text to lowercase, unless retaining uppercase has semantic meaning (such as company names, proper nouns).

[0075] Stop word filtering: Use the stop word library of nltk or spaCy to remove common meaningless words.

[0076] Lemmatization and stemming: Lemmatization and stemming are used to reduce the complexity of text so that different forms of the same word can be processed uniformly.

[0077] Vector database and knowledge graph construction module, used to build vector database and knowledge graph based on pre-processed startup data;

[0078] A further implementation method is that the vector database and knowledge graph construction module includes:

[0079] The knowledge graph construction unit is used to perform entity recognition and relationship extraction on text standardized data to build a dynamically updated knowledge graph; in this embodiment, the Bert pre-trained model and tools such as Spacy and Stanford NLP are used to identify key entities and their relationships from the data. Entities include companies, market areas, technologies, patents, products, business areas, financing rounds, etc., and relationships include "the company has patents", "technology applied to products", "the company is active in the market", "the company obtains financing", "products involve technology", etc.

[0080] Use the pre-trained BERT model to perform entity recognition tasks and perform relationship extraction between the identified entities, including the following steps: Use tools such as spaCy and StanfordNLP to perform syntactic analysis on the input text to obtain information such as the dependency structure and part-of-speech tagging of the sentence. And design a set of rules or patterns based on the specific relationship. Apply the predefined rules to the input text, identify sentences that meet these patterns, and extract the corresponding entities and relationships. It should be noted that the relationships between the entities extracted by the present invention are all objective data in existing materials.

[0081] Using the graph database Neo4j, the identified entities and relationships are stored in a structured manner to form a dynamically updated knowledge graph. The system regularly extracts the latest data from the information base and updates the knowledge graph to ensure the timeliness and accuracy of the graph content. Through the pre-built knowledge graph, in the intelligent analysis phase, the system can quickly use the information in the graph for analysis after the user enters the company information to be analyzed.

[0082] A vector database construction unit is used to vectorize and optimize text standardization data, obtain optimized text embedding vectors, and perform vector retrieval on the optimized text embedding vectors based on a hierarchical navigable small world algorithm to complete the construction of a vector database. In this embodiment, the text standardization data is vectorized using the Ada model (text-embedding-ada-002) of OpenAI to generate a basic embedding vector. The vector represents the comprehensive semantic content of the input text. In order to further enhance the semantic representation capability of the text, the present invention optimizes the generated embedding vectors using a self-attention mechanism so that it can more accurately capture the semantic associations within the text, especially when processing complex texts, the optimized vectors can better reflect the context and overall semantic structure.

[0083] A further implementation method is that the process of obtaining the text embedding vector is:

[0084] Based on the Ada model, the text normalization data is processed to obtain the text embedding vector; in this embodiment, the input text T is passed to the Ada model to generate the corresponding text embedding vector:

[0085] V Ada =AdaModel(T)=[v1,v2,…,v d ]

[0086] Where T is the input text (text normalization data), V Ada It is a vector generated by the Ada model, which captures the semantic content of the text, but it cannot fully express the complex semantic dependencies in the text. i is the value of each dimension and d is the dimension of the vector.

[0087] In order to further optimize the embedded representation of text, the present invention adopts a self-attention mechanism, which can dynamically adjust the weight of the embedded vector so that the optimized vector not only retains the original semantic information, but also enhances the ability to capture important semantic relationships within the text. Specifically, based on the self-attention mechanism combined with the loss function, the text embedding vector V Ada Perform linear transformation to obtain query Q, key K and value vector V; and calculate the dot product between query Q and key K to obtain the similarity between different parts of the text; specifically, Q = V Ada W Q ,K=V Ada W K ,V=V Ada W V

[0088] Among them, W Q ,W K , W V is a learnable parameter matrix used to transform the embedding vector into query, key, and value vectors.

[0089] The Softmax function is used to convert the similarity between different parts of the text into attention weights; specifically, the similarity between different parts of the text is measured by calculating the dot product between the query Q and the key K:

[0090]

[0091] The result is then converted into an attention weight α through the Softmax function, which is used to measure the importance of different parts of the text:

[0092] α=softmax(Score(Q,K)).

[0093] The attention weight and the value vector V are weighted and summed to obtain the optimized text embedding vector. Specifically, the value vector V is weighted and summed by the weight α of the self-attention mechanism to generate the optimized embedding vector V opt :

[0095] When dealing with long texts or complex semantics, it can enhance the awareness of context.

[0096] In this embodiment, in order to enable the optimization process of the self-attention mechanism to generate more effective embedding vectors, the present invention uses the loss function of contrastive learning for training. By designing and optimizing the loss function, the model can learn the semantic association between similar texts and shorten the distance between their embedding vectors, while distinguishing dissimilar texts. The specific loss function designed and optimized is as follows:

[0097]

[0098] in, Represents the cosine similarity between optimized vectors, y i,j is a similarity label, with a value of 1 indicating similarity and a value of 0 indicating dissimilarity. By optimizing this loss function, the model can learn the similarity relationship between texts and adjust the way embedding vectors are generated, so that the vectors of similar texts are closer and the vectors of dissimilar texts are kept farther apart in the embedding space.

[0099] In particular, since the previously collected text information does not have a similarity label y i,j Therefore, the present invention needs to design similarity labels for the calculation of the loss function. The specific method is to use the LDA (Latent Dirichlet Allocation) model to perform topic modeling on the text, thereby extracting the topic distribution of each text. On this basis, the cosine similarity method is used to measure the similarity between different texts and generate corresponding similarity labels. The specific steps are as follows:

[0100] In the LDA model, each text generates a topic distribution θ T , which represents the probability that a text belongs to different topics. For example, a text may belong to topic 1 with a probability of 0.6, to topic 2 with a probability of 0.3, and to topic 3 with a probability of 0.1. For two texts T i and T j , by calculating their topic distribution and The cosine similarity of is used to determine their semantic similarity. The calculation formula of cosine similarity is as follows:

[0101] in, and It is text T i and text T j The topic distribution vector of . By calculating the cosine similarity, if the similarity value is greater than the set threshold δ LDA , then its similarity label y i,j Set to 1, indicating that the two texts are semantically similar; if the similarity value is less than δ LDA , then y i,j Set to 0, indicating that the semantic differences between the texts are large. In this embodiment, δ LDA Take 0.8.

[0102] Qdrant is an open source, high-performance vector database that uses the HNSW (Hierarchical Navigable Small World) algorithm to achieve efficient vector retrieval and provides a rich API and persistent storage capabilities, making it suitable for large-scale data processing. The system uses Qdrant to store and retrieve optimized text embedding vectors V of text. opt , supporting efficient similarity queries. The optimized text embedding vectors and their associated metadata (such as IDs and tags) are stored in Qdrant for subsequent retrieval and analysis.

[0103] A further implementation method is that the process of performing vector retrieval on the optimized text embedding vector based on the hierarchical navigable small-world algorithm is as follows:

[0104] The optimized text embedding vector is used as a node to construct a sparse hierarchical graph structure. Specifically, a sparse hierarchical graph structure is constructed, where each layer contains a set of points representing different vectors. The higher layers have fewer points, representing the approximate center point of the data, while the lower layers have more points, representing a more detailed point distribution. Each point corresponds to an optimized embedding vector V of a text. opt , these vectors reflect the semantic information of the text. In order to control the level distribution, the parameter β is introduced to adjust the decay rate:

[0105] P(level=1)=e -βl

[0106] By adjusting β, we can better balance the search speed and accuracy.

[0107] Insert the new optimized text embedding vector to update the sparse hierarchical graph structure; specifically, when a new optimized embedding vector V opt When , the HNSW algorithm starts from the top layer of the graph, searches for the node closest to the new vector, and moves down layer by layer. At each layer, the algorithm selects the node that matches the new optimized embedding vector V. opt The closest existing node V j , and in the similarity s(V opt ,V j ) exceeds the threshold θ and embeds the new optimized vector V opt Insert into the appropriate position. Similarity calculation can be done using cosine similarity:

[0108]

[0109] This optimization method helps maintain the sparsity of the graph. In addition, in order to improve the local optimization effect, the penalty term δ is used to dynamically adjust the similarity threshold to optimize the density distribution of inserted nodes:

[0110] s'=s(V opt ,V j )-δ·d(V opt ,V j )

[0111] Where d is the distance between two vectors. This optimization method can ensure that the new vector is inserted at the local optimal position, improving the retrieval efficiency and accuracy.

[0112] Based on layer-by-layer navigation, the updated sparse hierarchical graph structure nodes are searched locally and globally to obtain the storage vector closest to the user query vector. Specifically, during similarity retrieval, the system starts from the top layer of the graph and gradually finds the storage vector closest to the user query vector. By navigating layer by layer, the closest node is found from local to global until the bottom layer is reached, and finally the storage vector V that is most similar to the query vector is returned. opt Each time a new vector is inserted or a similarity search is performed, the system dynamically maintains the hierarchical structure of the graph to ensure that the similarity between vectors is accurately maintained and the retrieval efficiency is not affected. In order to further optimize the efficiency of the system, a cost function C is introduced to strike a balance between sparsity S and search time T:

[0113] C=αS+(1-α)T

[0114] By optimizing this cost function, the flexibility and efficiency of the system in different usage scenarios can be ensured.

[0115] In this system, Qdrant is used to store and retrieve the optimized text embedding vector V opt , the specific process is as follows:

[0116] Vector storage: The optimized embedding vector V opt , the unique identifier (ID) of the text, tags and other metadata are stored in the Qdrant database.

[0117] Similarity query: When a user initiates a similar text retrieval request, the system first generates an embedding vector V of the user input text query (This vector is generated through the same optimization process as the storage vector) and then quickly finds and queries the vector V through the Qdrant API based on the HNSW algorithm query The most similar storage vector V opt The HNSW algorithm effectively finds the closest vector in high-dimensional space through multi-level neighbor node search.

[0118] Query result return: Qdrant returns the same value as the query vector V query The system presents the most similar text records and their related metadata (such as tags and IDs) to the user or for further processing and analysis.

[0119] The context information construction module is used to perform in-depth semantic analysis on the relevant information of the enterprise to be analyzed input by the user, and to integrate the semantic analysis results with the data of the vector database and the knowledge graph to obtain context information.

[0120] A further implementation is that the context information building module includes:

[0121] The semantic analysis unit is used to perform deep semantic analysis on the relevant information of the enterprise to be analyzed input by the user based on the large language model to obtain the semantic analysis results; in this embodiment, OpenAI GPT-4 or Zhipu GLM-4 is used to perform deep semantic analysis on the company information (enterprise to be analyzed) input by the user to extract key information such as company name, technical field, market positioning, patent data, etc., to provide a basis for subsequent association analysis and context generation. This information is first stored in plain text feature form as F, including the extracted text features:

[0122] F=[f 公司名称 ,f 技术领域 ,f 市场定位 ,f 专利数据 ,…]

[0123] Next, convert F into an embedding vector V input :

[0124] V input =Model(F)

[0125] Here, Model refers to the OpenAI Ada model or Zhipu Embedding-3 model used by the system.

[0126] The knowledge graph query unit is used to extract the competitor data of the company to be analyzed in the knowledge graph based on the semantic analysis results; specifically, based on the plaintext feature F extracted from the semantic analysis unit, the graph database Neo4j is queried to extract entity and relationship information such as related technologies, patents, competitors, and market conditions related to the company to be analyzed in the knowledge graph. The information such as the competitive situation, related technology development, and market conditions of the company's industry recorded in the existing data is obtained. The specific method is:

[0127] For each text feature f i Execute query Q in the knowledge graph i , retrieve and return the entity set and relationship set related to the feature. This information includes competitors, related technologies, market conditions, etc. of the company to be analyzed:

[0128] Q i ={E i ,R i}

[0129] Where E i is the same as the characteristic f i The set of related entities, R i It is a collection that describes the relationships between these entities. It combines the relevant data returned by the knowledge graph into a context structure K, covering information such as competition situation and technology market:

[0130] The vector database retrieval unit is used to retrieve industry dynamic data related to the enterprise to be analyzed in the vector database based on the semantic analysis results; the vector database is used to retrieve market, technology, competitor and other data related to the company (enterprise to be analyzed) input by the user, and provide a reference for a wider range of industry, market and technology information to enhance context information. The specific method is:

[0131] The embedding vector V generated by the semantic analysis unit input As input, we perform similarity-based retrieval in the vector database (Qdrant) using the following formula with the embedding vector V in the vector database: db Calculate similarity:

[0132]

[0133] The data with the highest similarity will be used as company and industry related information V rel , converted back to plaintext data M through the index rel , added to the context construction.

[0134] The data fusion unit is used to reconstruct the data by taking the relevant information of the enterprise to be analyzed as the main context and the competitor data and industry dynamic data as supplementary information to obtain context information. Specifically, the data of the knowledge graph and vector database are fused with the company information input by the user, and the relevant company, technology, patent, and competition data are reconstructed and unified into a structured text, and organized in a logical order to construct complete context information for subsequent analysis modules. The specific methods and steps of information fusion are as follows:

[0135] Main company information priority: The company information entered by the user, after deep semantic analysis, the text data F is used as the main context, with the highest priority. It includes basic company information, technological achievements, financial situation, market positioning and patents, etc. This is the core foundation for subsequent information integration.

[0136] Supplementary technology and competition information: The company's technology, patents, competitors and other information K extracted from the knowledge graph are used as supplements. This part of the information supplements the company's technological advantages in the market and the competitive landscape it faces.

[0137] Supplement market background and industry dynamics: The system retrieves companies similar to the target company from the vector database. These companies may be in the same technology field or operate in a similar market. And query relevant market information based on the company's industry field (such as artificial intelligence, medical care, finance, etc.) rel , such as industry growth rate, penetration rate, main challenges and risks, development history and current status, etc.

[0138] In this embodiment, a context verification and correction unit is also included to verify and correct the context information to ensure the accuracy and completeness of the information. This step is particularly important for improving the reliability of the analysis results, especially after the information in the knowledge graph and vector database is fused with the data input by the user, there may be redundant, irrelevant or contradictory information. Therefore, the verification and correction process is designed to clean, filter and optimize the context to provide the best input for subsequent model fine-tuning and reasoning.

[0139] Completeness check: Ensure that the context contains the necessary key information, including company name, technical expertise, market positioning, patents, etc., and that no key parts are omitted. Completeness check ensures that all key information is included in the final generated context, especially those that are critical for analysis and research support, such as company name, technology, patents, etc. If some key information is missing, the system will prompt the user to complete it, or automatically supplement the missing part by re-searching the knowledge graph or vector database.

[0140] Removing irrelevant content: During the information fusion process, the system may extract some information that is irrelevant to the current analysis. This content may interfere with the reasoning of the model, so it is cleaned up through filtering or noise reduction. Filtering or noise reduction uses the LDA (Latent Dirichlet Allocation) topic modeling algorithm to identify topics that match the user input content and filter out other irrelevant topics. The specific method is:

[0141] Using the LDA model, topic modeling analysis is performed on the fused contextual text information to determine the topic for each text.

[0142] Use the LDA model to assign topics to the text entered by the user and find the corresponding topic number.

[0143] Assign topics to the extracted text and filter out content that does not belong to the same topic as the user input.

[0144] The intelligent analysis module is used to intelligently analyze contextual information based on the startup research and analysis auxiliary large model, obtain the development status data of the company to be analyzed, and complete the intelligent analysis of startups based on the large language model. The main purpose of the intelligent analysis and research support engine is to generate in-depth and feasible analysis and research results by combining multi-dimensional data sources and using the fine-tuned large language model, and provide the factual basis behind it. The engine will effectively integrate data such as company information, market structure, technology patents, financing history and expert analysis, and assist industry analysts, researchers, etc. to quickly and comprehensively understand the target company and market environment, and make more reasonable research and analysis judgments.

[0145] A further implementation is that the intelligent analysis module includes:

[0146] A model building unit is used to fine-tune the large language model based on relevant data in the field of the enterprise to be analyzed, so as to obtain a large model for assisting the research and analysis of start-up enterprises; in this embodiment, it includes:

[0147] Fine-tuning data preparation:

[0148] Data collection and integration: Collect data for fine-tuning, including technical literature, patent information, market analysis reports, company-related news, etc. This data should be highly relevant to the target field (for example, for the technology industry) so that the model can learn domain-specific language and knowledge. Most company information already exists in graph databases and vector databases, including company growth history, technical background, financing information, market dynamics, etc. The purpose of this part of data collection is to check for gaps, supplement past startup information, and ensure the completeness and accuracy of the data. In addition, during the fine-tuning process, it is also necessary to create comprehensive data samples that can summarize the above information to help the model learn how to integrate multiple information sources to derive analysis and research basis.

[0149] Sample construction: Based on the company's development results, a certain number of startup company information is divided into success cases and failure cases for fine-tuning training, key company information is retrieved and processed, and samples are annotated in combination with expert opinions. Success cases: companies that have completed multiple rounds of financing, have been listed or acquired. Failure cases: companies that have been liquidated, closed or have been unable to raise funds for a long time. These samples will help the model learn how to generate different types of analysis and auxiliary evidence based on different inputs.

[0150] Expert opinion annotation: Combine the experience and expertise of experts to annotate the information of each company. Experts will provide opinions on the company's technological innovation and market competitiveness, and annotate the reasons and factual basis. These annotations will provide more logical analysis and basis presentation capabilities for the subsequent fine-tuning of the startup research and analysis auxiliary model.

[0151] Fine-tune the dataset format: The dataset will be stored in JSON format, including key dimensions such as company information, market performance, technical analysis, expert opinions, logical analysis, and factual basis, to help the large language model learn the reasons for the success or failure of a company and provide support for subsequent analysis.

[0152] Data partitioning: The labeled dataset is divided into a training set and a validation set in an 80 / 20 ratio to ensure that the performance of the model on the validation set can truly reflect its generalization ability.

[0153] Model fine-tuning:

[0154] Model selection: Use GPT-4 or Zhipu GLM-4 as the base model and upload the prepared fine-tuning data through the API. Each data sample should include:

[0155] Prompt: including company background, market information, core technology, patent information, competitors and other information.

[0156] Completion: Specific auxiliary analysis, with detailed explanation of the logical reasoning, data support and factual basis behind it.

[0157] After the data is uploaded, set parameters such as the number of training rounds and start the fine-tuning job.

[0158] Effect verification and model optimization:

[0159] Fine-tuning process:

[0160] The training set is used to update the model weights. The model sees the fine-tuning data in the training set, calculates the training loss, and backpropagates the error to update the weights.

[0161] After each epoch, the model weights are updated, the system pauses training, and predicts the validation set. The model generates predictions based on the validation set data and compares them with the true annotations to calculate the validation loss.

[0162] Continuously optimize model parameters, such as learning rate, number of training rounds, etc., to ensure that the model has good generalization capabilities and can adapt to the situations of different companies.

[0163] After fine-tuning is completed, the results generated by the model are verified by actually inputting company information to see if they are logically consistent and have sufficient factual basis and data support.

[0164] The intelligent analysis unit is used to analyze the context information based on the large model for assisting the research and analysis of start-up companies, and obtain the development status data of the company to be analyzed. The development status data of the company to be analyzed includes market data, technical data, competition analysis data and potential risk data about the company. It should be noted that the development status data finally obtained by the present invention are all integrated from the objective data existing in the existing materials. Among them, the technical status and patent coverage, market status and industry application, as well as risks and challenges are all summaries and explanations of the integrated objective data, which play a role of assistance and data support for analysts.

[0165] In this embodiment, in order to help users better understand the development status data of the enterprise to be analyzed, the system can display part of the content in the form of charts through visualization tools (such as Plotly or Dash), making the report more intuitive and easy to understand. For example, Plotly can be used to create line charts, bar charts, bubble charts, etc. to represent market share information charts, industry scale change charts, competition comparison charts and other visualization content.

[0166] Embodiment 2

[0167] In the construction and update of knowledge graph, examples of entities and relationships are as follows:

[0168] a) For example, we will use a specific market area, “machine vision”, to demonstrate how entities and relationships in the knowledge graph are constructed. The example shows how entities such as companies, technologies, products, patents, business areas, and financing information, as well as the relationships between them, are organized in the knowledge graph.

[0169] Entity example:

[0170] Company: Company X, Company Y

[0171] technology:

[0172] Computer Vision

[0173] Image Processing

[0174] Products: VisionTech Suite, VisionAI System

[0175] Patents: CN1XXXXXXXXA, CN1YYYYYYYYA, CN1MMMMMMMMA, CN1NNNNNNNNABusiness areas:

[0176] Automated Quality Inspection

[0177] Intelligent Production Line Monitoring

[0178] Industrial Robot Vision Navigation

[0179] Financing information: Angel round, A round, investment company A, investment company B

[0180] Relationship example:

[0181] Company and Technology:

[0182] Company X - uses - Computer Vision

[0183] Company Y - uses - Computer Vision

[0184] Company X - uses - Image Processing

[0185] Company Y - uses - Image Processing

[0186] Company and Products:

[0187] Company X-launches-VisionAI System

[0188] Company Y - Launches - VisionTech Suite

[0189] Technology and products:

[0190] Computer Vision - applied in -

[0191] VisionAI System

[0192] Computer Vision - applied in - VisionTech Suite

[0193] Image Processing-applied in-VisionAI System

[0194] Image Processing-applied in-VisionTech Suite

[0195] Companies and Patents:

[0196] Company X-Owns-Patents CN1XXXXXXXXXA, CN1MMMMMMMMA

[0197] Company Y-owns-patents CN1YYYYYYYYA, CN1NNNNNNNNA

[0198] Products and Patents:

[0199] VisionAI System-based on-patents CN1XXXXXXXXA, CN1MMMMMMMMA

[0200] VisionTech Suite-based on-patents CN1YYYYYYYYA, CN1NNNNNNNNA

[0201] Companies and business areas:

[0202] Company X-operates in-Automated Quality Inspection, Intelligent Production Line Monitoring

[0203] Company Y - operates in - Automated Quality Inspection, Industrial Robot Vision Navigation

[0204] Company and Financing:

[0205] Company X – receives funding from – Investment Company A (Angel Round)

[0206] Company Y – receives funding from – Investment Company B (Round A)

[0207] Company and Company:

[0208] Company X – competes with – Company Y

[0209] Through the above knowledge graph, the system can quickly display the technical applications, product launches, patent coverage, active business areas and financing history in the field of machine vision of Company X and Company Y. This information can be used to analyze the market status of the company, the company's competitiveness in the market and the technological advancement, and ultimately support researchers and analysts to conduct in-depth market and technical evaluations, providing a strong basis for research and analysis.

[0210] Sample examples for building a large model to assist in research and analysis of startups:

[0211] Fine-tuning text data example 1:

[0212] Prompt:

[0213] Company Name: Startup A

[0214] Year founded: 2014

[0215] Financing round:

[0216] Seed round: $1 million (2015)

[0217] Series A: US$5 million (2016)

[0218] Series B: US$20 million (2018)

[0219] Business: Autonomous driving software development

[0220] patent:

[0221] US2018XXXXX1 Autonomous Driving Perception System

[0222] US2019XXXXX2 Road Safety Monitoring System

[0223] Technical data: Text A, Text B, Text C, Text D, ...

[0224] Market: Autonomous driving market

[0225] Market data: TextE, TextF, TextG, TextH, ...

[0226] Competitors: Company B, Company C

[0227] Competitor data: Text I, Text J, Text K, Text L, ...

[0228] Current status of the company: Acquired

[0229] Completion:

[0230] Technology status and patent coverage:

[0231] Startup A’s patent portfolio covers key technologies in the field of autonomous driving, including perception and road safety monitoring systems. These patents provide technical support for high-precision perception and environmental monitoring of autonomous vehicles.

[0232] Patents US2018XXXXX1 and US2019XXXXX2 were authorized in 2018 and 2019 respectively, covering core technologies in the field of autonomous driving.

[0233] Market Status and Industry Applications:

[0234] According to the market report Text E, the market size of autonomous driving technology will reach approximately $XXX billion in 2023 and is expected to grow to $YYY billion in 2025. Startup A has established partnerships with several automakers to apply its patented technology to mass-produced models. In 2023, Startup A's technology is used in more than X% of new mass-produced autonomous vehicles.

[0235] Risks and Challenges:

[0236] Industry standards and regulations for autonomous driving technology are still developing, and there may be a risk of regulatory changes in the future.

[0237] Fine-tuning text data example 2:

[0238] Prompt:

[0239] Company Name: Startup B

[0240] Year founded: 2015

[0241] Financing round:

[0242] Seed round: $2 million (2016)

[0243] Business: VR applied to online education platform

[0244] patent:

[0245] US2016XXXXX1 VR Education Interactive System

[0246] US2017XXXXX1 VR control technology system

[0247] Technical data: Text 1, Text 2, Text 3, Text 4, ...

[0248] Market: VR education market

[0249] Market data: Text5, Text6, Text7, Text8, ...

[0250] Competitor: Company H

[0251] Competitor data: Text 9, Text 10, Text 11, Text 12, ...

[0252] Current status of the company: Closed

[0253] Completion:

[0254] Technology status and patent coverage: Startup B's VR educational interactive system patent (US2016XXXXX1) and VR control technology system patent (US2017XXXXX1) are mainly about VR interactivity and control technology. In actual applications, the implementation of VR technology is relatively complex, especially in educational interactivity and content creation, which requires a lot of cost and technology investment. For example, H Company's technology investment exceeds US$X billion.

[0255] Market conditions and industry applications: According to the market report "Text 5", the global VR market size reached approximately US$XX billion in 2016 and is expected to grow to US$YY billion in 2025. Currently, educational applications account for 5% of the number of VR applications.

[0256] Risks and Challenges: Company B has failed to establish a partnership with a large education platform or content provider, while its competitor Company H has already carried out technology promotion and channel coverage. In addition, Company H has more than one million users worldwide, while Company B has less than 100,000 users. Company B has not conducted a new round of financing for X consecutive years.

[0257] Example of user usage:

[0258] 1. User initial input:

[0259] Company Name: Startup C

[0260] Industry: Artificial Intelligence

[0261] Business: Smart healthcare

[0262] Patent: US20200345678 AI-based intelligent diagnosis system

[0263] Market Data: Text 1

[0264] Competitor: Company M

[0265] The user first enters some information about startup C, such as company name, industry, some patents or competitors, etc. Since the information is incomplete, the system needs to automatically fill in the missing data.

[0266] 2. Completion through graph database:

[0267] During the completion process, the graph database is mainly responsible for identifying company entities and their related financing, patents, market positioning, competitors, etc. Through the established entity relationships and paths, the graph database can effectively retrieve the relevant nodes of company C to help complete the following information.

[0268] 2.1 Complete the year of establishment and financing rounds:

[0269] Through the "Company C" node in the graph database, the system can find the relationship between the company and the financing and founding year. For example, the "Founding Year" in the graph database is associated with Company C, and the system completes it as follows after searching:

[0270] Year of establishment: 2020

[0271] Financing round:

[0272] Seed round: $2 million (2021)

[0273] Series A: US$10 million (2022)

[0274] Series B: US$50 million (2023)

[0275] 2.2 Patent Completion

[0276] The system associates the "Company C" node with the "Patent" node to obtain more relevant patent information. For example:

[0277] patent:

[0278] US20200345678 AI-based intelligent diagnosis system

[0279] US20210567891 Intelligent pathology image analysis technology

[0280] US20220123456 Automated medical data processing system

[0281] 2.3 Competitor Completion

[0282] The market field and competitor nodes associated with "Company C" in the graph database can help complete the market and main competitors. For example, the graph database query finds that the market of Company C is smart healthcare and intelligent diagnosis, and its competitors include not only Company M entered by the user, but also Company N.

[0283] Market: Smart healthcare and intelligent diagnosis

[0284] Competitors: Company M, Company N

[0285] 3. Completion through vector database:

[0286] Vector databases are mainly used to complete unstructured document information, especially market data, technical documents, documents related to competitors, news reports, etc. The system uses the vector database to perform similarity searches based on the input text or company information retrieved from the graph database to complete more detailed document content.

[0287] 3.1 Market data completion

[0288] The company name entered by the user, "Text 1" and the market information retrieved from the graph database are used as the initial query. The system uses the vector database to perform a similarity search to find market reports and industry data related to "Text 1".

[0289] Market data: text1, text2, text3, text4, ...

[0290] 3.2 Technical data completion

[0291] The system uses the vectorization model to generate optimized embedding vectors of these patents based on the patent information input by the user and associated with the graph database, and performs similarity search in the vector database. Through the vector database search operation, the system finds technical documents related to the patent and completes the technical information "Technical Text A", "Technical Text B", "Technical Text C" and "Technical Text D" and other technical documents and reports. The result example is:

[0292] patent:

[0293] US20200345678 AI-based intelligent diagnosis system

[0294] US20210567891 Intelligent pathology image analysis technology

[0295] US20220123456 Automated medical data processing system

[0296] Technical data: Text A, Text B, Text C, Text D, ...

[0297] 3.3 Competitor data completion

[0298] Through user input and competitors retrieved by the graph database (such as company M and company N), the system can pass this information to the vector database to further retrieve relevant industry reports, market dynamics and other documents. For example, by entering "company M" and "company N", the vector database can complete the following data:

[0299] Competitor data: Text E, Text F, Text G, Text H, ...

[0300] Here, the system uses "Company N" retrieved from the graph database as a query condition and passes it to the vector database. After vectorization, the system finds market reports, news documents, or analysis reports related to "Company N" in the vector database. For example, the system retrieves "Text H" as data related to Company N (competitor data).

[0301] Through the above operations, the final completed information example is as follows:

[0302] Company Name: Startup C

[0303] Year of establishment: 2020

[0304] Financing round:

[0305] Seed round: $2 million (2021)

[0306] Series A: US$10 million (2022)

[0307] Series B: US$50 million (2023)

[0308] Business: Smart healthcare

[0309] patent:

[0310] US20200345678 AI-based intelligent diagnosis system

[0311] US20210567891 Intelligent pathology image analysis technology

[0312] US20220123456 Automated medical data processing system

[0313] Technical data: Text A, Text B, Text C, Text D

[0314] Market: Smart healthcare and intelligent diagnosis

[0315] Market data: Text1, Text2, Text3, Text4

[0316] Competitors: Company M, Company N

[0317] Competitor data: Text E, Text F, Text G, Text H, ...

[0318] Current status of the company: Completed 3 rounds of financing in a row

[0319] The completed information is input into the large model as the prompt of the large model, and the final result is as follows:

[0320] Company Name: Startup C

[0321] Year of establishment: 2020

[0322] Financing round:

[0323] Seed round: $2 million (2021)

[0324] Series A: US$10 million (2022)

[0325] Series B: US$50 million (2023)

[0326] Business: Smart healthcare

[0327] patent:

[0328] US20200345678 AI-based intelligent diagnosis system

[0329] US20210567891 Intelligent pathology image analysis technology

[0330] US20220123456 Automated medical data processing system

[0331] Technical data: Text A, Text B, Text C, Text D

[0332] Market: Smart healthcare and intelligent diagnosis

[0333] Market data: Text1, Text2, Text3, Text4

[0334] Competitors: Company M, Company N

[0335] Competitor data: Text E, Text F, Text G, Text H, ...

[0336] Current status of the company: Completed 3 rounds of financing in a row

[0337] Technology status and patent coverage: Startup C's patent portfolio covers core technologies in the field of smart healthcare, especially in pathological image analysis and intelligent diagnosis. Patents "US20200345678" and "US20210567891" were authorized in 2020 and 2021, respectively, providing strong technical support for the company's innovation in intelligent diagnostic systems.

[0338] Market conditions and industry applications: According to market data (Text 1), the smart healthcare market will reach US$15 billion in 2023 and is expected to grow to US$25 billion by 2025. Startup C cooperates with many hospitals, and its smart diagnostic system has been widely used in clinical trials.

[0339] Risks and challenges: Regulations in the field of smart healthcare are still being improved, and Company C may face challenges from policy changes. The technological competition with Company M and Company N has intensified. The two competitors have obtained multiple orders from key customers in the market and have continuously issued related patents. Company C needs to continue to innovate to maintain its competitiveness.

[0340] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. An intelligent analysis system for start-ups based on a large language model, characterized in that: include: Data collection module, used to collect data on startups; A data preprocessing module, used for preprocessing the start-up enterprise data; A vector database and knowledge graph construction module, used to construct a vector database and a knowledge graph based on the preprocessed startup data; A context information building module is used to perform deep semantic analysis on the relevant information of the enterprise to be analyzed input by the user, and fuse the semantic analysis results with the data of the vector database and the knowledge graph to obtain context information; The intelligent analysis module is used to perform intelligent analysis on the context information based on the large model for assisting the research and analysis of start-ups, obtain the development status data of the enterprise to be analyzed, and complete the intelligent analysis of start-ups based on the large language model.

2. The startup enterprise intelligent analysis system based on a large language model according to claim 1 is characterized in that: The startup data includes market report data, patent data, technical literature data, industry news data and technology company data.

3. The startup enterprise intelligent analysis system based on a large language model according to claim 1 is characterized in that: The data preprocessing module comprises: A data parsing unit, used to parse the start-up enterprise data to obtain structured text information; wherein the parsing includes PDF parsing, HTML parsing, XML / JSON parsing and table data processing; A data cleaning unit, used for denoising and formatting the structured text information to obtain format-standardized data; The text standardization unit is used to unify the case of format-standardized data, filter stop words, and perform word form restoration and stem extraction to obtain text-standardized data.

4. The startup enterprise intelligent analysis system based on a large language model according to claim 3 is characterized in that: The vector database and knowledge graph construction module includes: A knowledge graph construction unit, used to perform entity recognition and relationship extraction on the text standardization data to construct a dynamically updated knowledge graph; The vector database construction unit is used to perform vectorization processing and optimization on the text standardization data to obtain the optimized text embedding vector, and perform vector retrieval on the optimized text embedding vector based on the hierarchical navigable small-world algorithm to complete the construction of the vector database.

5. The startup enterprise intelligent analysis system based on a large language model according to claim 4 is characterized in that: The process of obtaining the text embedding vector is: Processing the text normalization data based on the Ada model to obtain a text embedding vector; Based on the self-attention mechanism combined with the loss function, the text embedding vector is linearly transformed to obtain the query Q, key K and value vector V; and the dot product between the query Q and the key K is calculated to obtain the similarity between different parts of the text; Use the Softmax function to convert the similarity between different parts of the text into attention weights; The attention weight and the value vector V are weightedly summed to obtain an optimized text embedding vector.

6. The intelligent analysis system for start-ups based on a large language model according to claim 5 is characterized in that: The process of vector retrieval based on the optimized text embedding vector based on the hierarchical navigable small-world algorithm is as follows: Use the optimized text embedding vectors as nodes to build a sparse hierarchical graph structure; Inserting a new optimized text embedding vector to update the sparse hierarchical graph structure; Based on layer-by-layer navigation, a local to global search is performed on the updated sparse hierarchical graph structure nodes to obtain a storage vector that is closest to the user query vector.

7. The intelligent analysis system for start-ups based on a large language model according to claim 5 is characterized in that: The context information building module includes: A semantic analysis unit, which is used to perform in-depth semantic analysis on the relevant information of the enterprise to be analyzed input by the user based on a large language model to obtain a semantic analysis result; A knowledge graph query unit, used to extract the opponent data of the enterprise to be analyzed in the knowledge graph based on the semantic analysis result; A vector database retrieval unit, configured to retrieve industry dynamic data related to the enterprise to be analyzed from the vector database based on the semantic analysis result; The data fusion unit is used to reconstruct data by taking the relevant information of the enterprise to be analyzed as the main context and taking the competitor data and the industry dynamic data as supplementary information to obtain the context information.

8. The startup enterprise intelligent analysis system based on a large language model according to claim 1 is characterized in that: The intelligent analysis module includes: The model building unit is used to fine-tune the large language model based on relevant data in the field of the enterprise to be analyzed, and obtain a large model to assist in the research and analysis of start-ups; The intelligent analysis unit is used to analyze the context information based on the start-up enterprise research and analysis auxiliary model to obtain the development status data of the enterprise to be analyzed.

Images