Enterprise-level knowledge base construction method based on large model

Abstract

The invention discloses an enterprise-level knowledge base construction method based on a large model. The enterprise-level knowledge base construction method comprises the following specific steps: acquiring original data of an enterprise and preprocessing the original data to obtain first target data; performing knowledge extraction and representation based on the first target data to obtain second target data; based on the random-degree large model and the second target data, performing knowledge graph construction and storage to obtain an enterprise-level knowledge base; performing knowledge retrieval and recommendation on the enterprise-level knowledge base by using a modular retrieval enhancement generation method; and periodically updating and maintaining the enterprise-level knowledge base. According to the method, keywords and concepts in various knowledge carriers such as internal documents, reports and communication records of enterprises can be automatically identified based on the strong analysis capability of the random-degree large model, and automatic association among knowledge elements is realized; the enterprise-level knowledge base is quickly constructed, the knowledge base is periodically updated and maintained in time, the content of the enterprise-level knowledge base is enriched, and a foundation is laid for subsequent enterprise knowledge search.

Description

Technical Field

[0001] The present invention relates to the field of knowledge base construction, and in particular to a method for constructing an enterprise-level knowledge base based on a large model. Background Art

[0002] At present, with the expansion of enterprise scale and diversification of business, how to efficiently build, manage and utilize the huge knowledge resources within the enterprise is crucial. The enterprise knowledge base serves as a self-service desk for customers and employees to store, organize and share information. It helps employees complete their work more effectively, allowing them to focus on their tasks and what is really important, without having to conduct a large amount of heavy retrieval for a thing or task, that is, they can refine the key points of things or tasks. However, traditional knowledge bases and their management methods are difficult to meet the needs of enterprises for deep mining and application of knowledge; in addition, the existing large language models require massive databases, and the process of searching for data is time-consuming and laborious, and it is impossible to quickly build an enterprise-level knowledge base, and the construction efficiency is low; at the same time, it takes a long time to update and maintain the knowledge base, and the cost is high, which brings a certain burden to the enterprise. Summary of the invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for constructing an enterprise-level knowledge base based on a big model. Based on the powerful analysis ability of the big model, it can automatically identify keywords and concepts in various knowledge carriers such as internal enterprise documents, reports, communication records, etc., and realize automatic association between knowledge elements; while quickly constructing an enterprise-level knowledge base, it also regularly and timely updates and maintains the knowledge base, enriches the content of the enterprise-level knowledge base, and lays a foundation for subsequent convenient search of enterprise knowledge.

[0004] The technical solution adopted by the present invention is:

[0005] A method for constructing an enterprise-level knowledge base based on a large model includes the following specific steps:

[0006] S1. Obtain the original data of the enterprise and preprocess it to obtain the first target data;

[0007] S2. Extracting and representing knowledge based on the first target data to obtain second target data;

[0008] S3, constructing and storing a knowledge graph based on the Renduda model and the second target data to obtain an enterprise-level knowledge base;

[0009] S4. Use modular retrieval enhancement generation method to perform knowledge retrieval and recommendation on enterprise knowledge base;

[0010] S5. Regularly update and maintain the enterprise-level knowledge base.

[0011] Furthermore, the step S1 specifically includes:

[0012] S11. Obtaining the original data of the enterprise:

[0013] 1) Obtain the electronic document data, database data, business report data, market data, customer feedback data and product information data within the enterprise through the API interface;

[0014] 2) Use web crawlers to obtain external data on papers and journals, industry reports, news and information, and social media;

[0015] S12, storing the obtained raw data in a unified manner in the unstructured database MongoDB;

[0016] S13, preprocessing the original data to obtain first target data, the process includes:

[0017] 1) Identify and correct invalid or erroneous raw data entries; Identify and delete invalid or erroneous raw data entries; Use the aggregation framework and query operations in MongoDB to remove duplicate records and clean the raw data;

[0018] 2) Use the $set update operation in MongoDB to update the fields of the original data to a unified format to achieve normalization of the original data;

[0019] 3) Determine the missing values ​​of the original data and fill the missing values ​​using the $ifNull operator in MongoDB;

[0020] 4) Use the type conversion function in MongoDB to convert the string in the original data into a date type or the text into a numeric value;

[0021] 5) Use the $lookup aggregation phase in MongoDB to connect and integrate data scattered in different database tables;

[0022] 6) Use the create Index command to create a single key index or a composite index for the original data;

[0023] 7) Determine the shard key and use the sh.add Shard and sh.shard Collection commands to prepare the shard collection;

[0024] 8) Use the mongoexport tool to export data from MongoDB to CSV format for further external processing; use the mongoimport tool to import the processed data into MongoDB;

[0025] 9) Place frequently accessed fields at the top of MongoDB documents to speed up document access and optimize document structure.

[0026] Furthermore, the step S2 specifically includes:

[0027] S21, entity recognition: using NLP methods to identify named entities in the text of the first target data, and improving the accuracy of entity recognition through the Random Model;

[0028] S22, relationship identification;

[0029] S23, attribute value extraction;

[0030] S24. Knowledge representation.

[0031] Furthermore, the step S22 specifically includes:

[0032] 1) Generation of candidate items for relation recognition: After entity recognition, search for entity pairs in the text and filter and generate candidate items for relation recognition;

[0033] 2) Feature extraction: Use machine learning methods to extract part-of-speech tagging features, dependency path features, and context information features from the candidate items for relationship recognition, and use them as input data for training the classification model;

[0034] 3) Model training: Use the Renduda model to perform model learning and training to output the relationship between entity pairs;

[0035] 4) Relationship classification: Apply a classification model to each candidate entity pair to determine whether there is a predefined relationship type between them and classify the relationship type;

[0036] 5) Model optimization and evaluation: Continuously optimize model parameters during model training and evaluate model performance through evaluation indicators such as precision, recall, and F1 score;

[0037] 6) Handling entity relationship overlap: The TP Linker model is used to convert the joint extraction task into a Token pair linking problem; the start and end positions of the entities are marked using the Token linking matrix, and the entity relationships are aligned using the hand shake tagging scheme, effectively solving the entity relationship overlap problem.

[0038] Furthermore, the step S23 specifically includes:

[0039] 1) Feature extraction: Extract features from text data based on a variant of long short-term memory network, namely bidirectional LSTM or Bi-LSTM. Bi-LSTM can capture the contextual information in the text and provide rich contextual features for the recognition of attribute values.

[0040] 2) Multiple instance learning: In the multiple instance learning framework, each entity may correspond to multiple instances, i.e., text descriptions. The algorithm can learn from these instances and identify the attribute values ​​related to the entity.

[0041] 3) Multi-label classification: Since an entity may have multiple attribute values, attribute value extraction can be modeled as a multi-label classification problem; each attribute value is used as a label, and the model predicts all possible labels in the text;

[0042] 4) Deep learning model training: Use the labeled training data to train the deep learning model. During the training process, the model learns to identify attribute values ​​from the text;

[0043] 5) Attention mechanism: In order to further improve the performance of the model, the attention mechanism is introduced to enable the model to pay more attention to the text fragments related to the attribute value;

[0044] 6) Model evaluation and optimization: Use accuracy, recall, and F1 score indicators to evaluate the performance of the model on the validation set, and adjust and optimize the model based on the evaluation results;

[0045] 7) Application and deployment: Deploy the trained model to actual applications to automatically identify and extract attribute values ​​from new text data.

[0046] Furthermore, the step S24 specifically includes:

[0047] 1) Multimodal named entity recognition: using image and text multimodal data to identify and link entities and improve recognition accuracy;

[0048] 2) Multimodal Relation Extraction: Extract relations from text and images, and verify and supplement them in multimodal data to achieve a more comprehensive understanding of relations;

[0049] 3) Multimodal event extraction: Identify events from multimodal data sources and extract event-related information, including time, location, and participants;

[0050] 4) Multimodal entity alignment: align the information describing the same entity in different modalities to ensure the consistency of knowledge representation;

[0051] 5) Multimodal entity linking and disambiguation: linking entities in multimodal data to the correct knowledge graph entities and solving entity disambiguation problems;

[0052] 6) Multimodal knowledge graph completion: Use multimodal data to complete the missing links and attributes in the knowledge graph to improve the completeness of the knowledge graph;

[0053] 7) Multimodal knowledge graph reasoning: Applying multimodal data for knowledge reasoning, including inferring attributes or relationships not explicitly mentioned in the text through image content;

[0054] 8) Transformer-based encoder: The Transformer-based encoder is used to extract the initial features of the image, text, and triple surface forms to achieve effective encoding between modalities;

[0055] 9) Interaction modeling between modalities: The interaction features between image and text modalities are learned through co-attention Transformer, and product attributes and attribute value information are aggregated on the target product entity through structural aggregation;

[0056] 10) Multimodal pre-training task: Pre-training of image modality, text modality and knowledge modality, including masked object model, masked language model and link prediction model, to optimize the model’s understanding of multimodal information.

[0057] Furthermore, the step S3 specifically includes:

[0058] S31. Entity linking and disambiguation:

[0059] 1) Candidate entity generation: Generate a set of possible candidate entities through various methods, including dictionary matching, synonym expansion, etc.

[0060] 2) Entity disambiguation: Using the deep learning model BERT, combined with contextual information and structured information in the knowledge graph, the most appropriate entity is selected for each mention. This includes using the attention mechanism to capture the most relevant part of the context to the entity, and considering the consistency between entities to achieve joint disambiguation.

[0061] S32, Relationship Extraction:

[0062] 1) Use conditional random fields (CRF) or Transformer-based models such as BERT to extract relations;

[0063] 2) Identify and classify the types of relationships between entity pairs by training the model;

[0064] S33, Automated Knowledge Completion:

[0065] 1) After the knowledge graph is constructed, the system automatically discovers and fills in the missing parts of the graph;

[0066] 2) Using graph database queries to identify isolated entities or incomplete relationship paths;

[0067] 3) Use web crawler technology to crawl relevant information from the Internet and extract new entities and relationships through information extraction technology;

[0068] 4) Apply machine learning models to predict and complete missing links, including the graph-based embedding model GraphSAGE;

[0069] S34. Knowledge graph storage:

[0070] 1) Use Neo4j graph database to store knowledge graph;

[0071] 2) When importing large amounts of data, use the LOADCSV command to import data, and optimize read and write performance by trading space for time.

[0072] Furthermore, the step S4 specifically includes:

[0073] S41, Knowledge Retrieval: By intelligently integrating keyword search, semantic search, and vector search technologies, the vector encoding model is used to convert user input into a vector, the similarity between question embedding and document block embedding is calculated, and the top K document blocks with the highest similarity are selected as enhanced context information for the current question;

[0074] S42. Find similar memories: Create an infinite memory pool, use the memory capacity of the large model itself to enhance retrieval, and find the memory most similar to the current input;

[0075] S43, additional generation: using LLM to generate the required contextual retrieval; removing redundancy and noise from the retrieved content;

[0076] S44, Task Adaptation: Focus on transforming RAG to adapt to downstream tasks;

[0077] S45, Alignment: Add a trainable Adapter module to the retriever to alleviate the alignment problem between query and text;

[0078] S46. Verification: An additional verification module is introduced after retrieving documents to evaluate the relevance between the retrieved documents and the query and enhance the robustness of RAG.

[0079] Furthermore, the step S5 specifically includes:

[0080] S51. New knowledge discovery: Use web crawler technology to regularly scan predefined information sources, including industry websites, professional databases, and social media to capture the latest content; use API interfaces to achieve real-time docking with external data sources and automatically import updated data and information;

[0081] S52. Archiving new knowledge: Use text mining and NLP techniques to evaluate the relevance and value of collected sentiment analysis and topic identification content; use NLP to automatically classify and label, and archive new knowledge into the correct category of the knowledge base;

[0082] S53. New Knowledge Review: Deploy machine learning models to identify outdated or inaccurate information. The models are trained based on historical data to understand the timeliness and accuracy of the content. Use deep learning technology to deeply understand and review the content.

[0083] S54. Version control and update: Use the version control system Git to automatically record each content change and ensure the traceability of historical versions; implement automated deployment processes to ensure seamless release of knowledge base updates and reduce human errors.

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] 1) Through the analysis capability of the Renduda model, keywords and concepts in various knowledge carriers such as internal documents, reports, and communication records of the enterprise can be automatically identified, and the automatic association between knowledge elements can be achieved to build an interconnected knowledge network;

[0086] 2) Dynamically adjust and optimize the structure of the knowledge graph according to the updates and changes of the enterprise knowledge base to ensure that the knowledge graph always reflects the latest knowledge status and relationships;

[0087] 3) Using the prediction and reasoning capabilities of the Renduda model, the incomplete or missing parts of the knowledge base are intelligently completed to improve the integrity and availability of the knowledge base;

[0088] 4) By building an efficient enterprise-level knowledge base, the efficiency of enterprise knowledge retrieval, sharing and application can be significantly improved, helping enterprises to quickly find the required knowledge and promoting the effective dissemination and utilization of knowledge;

[0089] 5) Provide a rich knowledge base for the innovation activities of enterprises by deeply exploring the potential connections between knowledge;

[0090] 6) Provide enterprise management with a decision support system based on a comprehensive knowledge graph, and assist enterprises in making more scientific and reasonable decisions by analyzing the complex relationships between knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

[0092] Figure 1 It is a flow chart of a method for constructing an enterprise-level knowledge base based on a large model according to the present invention. DETAILED DESCRIPTION

[0093] Example

[0094] The specific details of the Ren Du Da model mentioned in this embodiment are as follows:

[0095] (1) Any large model

[0096] The "Rendu Big Model" is a large language model independently developed by Language Network (Wuhan) Information Technology Co., Ltd. with zero open source dependencies. The R&D team of the Rendu Big Model recoded and developed the Transformer-like neural network framework zANN, and optimized the network structure to maintain more efficient parallel computing capabilities. The Rendu Big Model is based on a parameter scale of 11 billion, and can quickly analyze, understand and generate large-scale text data in a relatively low computing power environment. It also has good knowledge understanding reasoning, mathematics, programming guidance and intelligent agent interaction capabilities, and has demonstrated excellent intelligence in various fields.

[0097] At the same time, based on multi-level learning and fusion training, the "Rendu Big Model" also has cross-modal understanding capabilities. It can effectively combine the semantic understanding of multimodal data such as text, voice, images and videos, and realize cross-modal reasoning and knowledge transfer. This enables Rendu to not only understand user natural language in conversational interactions, but also understand and parse multi-modal content, providing richer understanding and generation capabilities.

[0098] The “Ren Duda Model” has passed the inspection of China Academy of Information and Communications Technology, report number: 24V01K000255-001.

[0099] (2) zANN Architecture

[0100] The zANN framework is a deep learning neural network framework independently developed by Yuwang Network (Wuhan) Information Technology Co., Ltd.

[0101] like Figure 1 As shown, a method for constructing an enterprise-level knowledge base based on a large model specifically includes the following steps:

[0102] S1. Obtain the original data of the enterprise and preprocess it to obtain the first target data:

[0103] Specifically, S11, obtaining the original data of the enterprise:

[0104] 1) Obtain the electronic document data, database data, business report data, market data, customer feedback data and product information data within the enterprise through the API interface;

[0105] 2) Use web crawlers to obtain external data on papers and journals, industry reports, news and information, and social media;

[0106] In addition, sensors are deployed to obtain specific data collected by specific industries, such as the deployment of sensors in the smart manufacturing industry to collect temperature data and / or humidity data;

[0107] S12, storing the obtained raw data in a unified manner in the unstructured database MongoDB;

[0108] S13, preprocessing the original data, specifically including the following steps:

[0109] 1) Identify and correct invalid or erroneous raw data entries; Identify and delete invalid or erroneous raw data entries; Use the aggregation framework and query operations in MongoDB to remove duplicate records and clean the raw data;

[0110] 2) Use the $set update operation in MongoDB to update the fields of the original data to a unified format to achieve normalization of the original data;

[0111] 3) Determine the missing values ​​of the original data and fill the missing values ​​using the $ifNull operator in MongoDB;

[0112] 4) Use the type conversion function in MongoDB to convert the string in the original data into a date type or the text into a numeric value;

[0113] 5) Use the $lookup aggregation phase in MongoDB to connect and integrate data scattered in different database tables;

[0114] 6) Use the create Index command to create a single key index or a composite index for the original data;

[0115] 7) Determine the shard key and use the sh.add Shard and sh.shard Collection commands to prepare the shard collection;

[0116] 8) Use the mongoexport tool to export data from MongoDB to CSV format for further external processing; use the mongoimport tool to import the processed data into MongoDB;

[0117] 9) Place frequently accessed fields at the top of MongoDB documents to speed up document access and optimize document structure;

[0118] In this embodiment, before the data is cleaned, the data is evaluated: the quality of the original data stored in MongoDB is evaluated, including the consistency, accuracy and completeness of the original data;

[0119] Use mongoexport and mongoimport tools to export and import data;

[0120] S2. Extracting and representing knowledge based on the first target data to obtain second target data:

[0121] Specifically, S21, entity recognition:

[0122] Using NLP methods to identify named entities in the text of the first target data, and improving the accuracy of entity recognition through a large model;

[0123] S22. Relationship Identification:

[0124] 1) Generation of candidate items for relation recognition: After entity recognition, search for entity pairs in the text and filter and generate candidate items for relation recognition;

[0125] 2) Feature extraction: Use machine learning methods to extract part-of-speech tagging features, dependency path features, and context information features from the candidate items for relationship recognition, and use them as input data for training the classification model;

[0126] 3) Model training: Use the Renduda model to perform model learning and training to output the relationship between entity pairs;

[0127] 4) Relationship classification: Apply a classification model to each candidate entity pair to determine whether there is a predefined relationship type between them and classify the relationship type;

[0128] 5) Model optimization and evaluation: Continuously optimize model parameters during model training and evaluate model performance through evaluation indicators such as precision, recall, and F1 score;

[0129] 6) Entity relationship overlap processing: The TP Linker model is used to transform the joint extraction task into a token pair linking problem; the start and end positions of the entities are marked using the token linking matrix, and the entity relationships are aligned using the hand shake tagging scheme, effectively solving the entity relationship overlap problem;

[0130] S23, attribute value extraction:

[0131] 1) Feature extraction: Extract features from text data based on a variant of long short-term memory network, namely bidirectional LSTM or Bi-LSTM. Bi-LSTM can capture the contextual information in the text and provide rich contextual features for the recognition of attribute values.

[0132] 2) Multiple instance learning: In the multiple instance learning framework, each entity may correspond to multiple instances, i.e., text descriptions. The algorithm can learn from these instances and identify the attribute values ​​related to the entity.

[0133] 3) Multi-label classification: Since an entity may have multiple attribute values, attribute value extraction can be modeled as a multi-label classification problem; each attribute value is used as a label, and the model predicts all possible labels in the text;

[0134] 4) Deep learning model training: Use the labeled training data to train the deep learning model. During the training process, the model learns to identify attribute values ​​from the text;

[0135] 5) Attention mechanism: In order to further improve the performance of the model, the attention mechanism is introduced to enable the model to pay more attention to the text fragments related to the attribute value;

[0136] 6) Model evaluation and optimization: Use accuracy, recall, and F1 score indicators to evaluate the performance of the model on the validation set, and adjust and optimize the model based on the evaluation results;

[0137] 7) Application and deployment: Deploy the trained model to actual applications to automatically identify and extract attribute values ​​from new text data;

[0138] S24. Knowledge Representation:

[0139] 1) Multimodal named entity recognition: using image and text multimodal data to identify and link entities and improve recognition accuracy; in this embodiment, in the relationship extraction task, images can be attached to significantly improve the recognition performance of attributes and relationships;

[0140] 2) Multimodal Relation Extraction: Extract relations from text and images, and verify and supplement them in multimodal data to achieve a more comprehensive understanding of relations;

[0141] 3) Multimodal event extraction: Identify events from multimodal data sources and extract event-related information, including time, location, and participants;

[0142] 4) Multimodal entity alignment: align the information describing the same entity in different modalities to ensure the consistency of knowledge representation;

[0143] 5) Multimodal entity linking and disambiguation: linking entities in multimodal data to the correct knowledge graph entities and solving entity disambiguation problems;

[0144] 6) Multimodal knowledge graph completion: Use multimodal data to complete the missing links and attributes in the knowledge graph to improve the completeness of the knowledge graph;

[0145] 7) Multimodal knowledge graph reasoning: Applying multimodal data for knowledge reasoning, including inferring attributes or relationships not explicitly mentioned in the text through image content;

[0146] 8) Transformer-based encoder: The Transformer-based encoder is used to extract the initial features of the image, text, and triple surface forms to achieve effective encoding between modalities;

[0147] 9) Interaction modeling between modalities: The interaction features between image and text modalities are learned through co-attention Transformer, and product attributes and attribute value information are aggregated on the target product entity through structural aggregation;

[0148] 10) Multimodal pre-training task: Pre-training of image modality, text modality and knowledge modality, including masked object model, masked language model and link prediction model, to optimize the model’s understanding of multimodal information.

[0149] S3. Based on the Renduda model and the second target data, the knowledge graph is constructed and stored to obtain an enterprise-level knowledge base:

[0150] Specifically, S31, entity linking and disambiguation:

[0151] 1) Candidate entity generation: Generate a set of possible candidate entities through various methods, including dictionary matching, synonym expansion, etc.

[0152] 2) Entity disambiguation: Using the deep learning model BERT, combined with contextual information and structured information in the knowledge graph, the most appropriate entity is selected for each mention. This includes using the attention mechanism to capture the most relevant part of the context to the entity, and considering the consistency between entities to achieve joint disambiguation.

[0153] S32, Relationship Extraction:

[0154] 1) Use conditional random fields (CRF) or Transformer-based models such as BERT to extract relations;

[0155] 2) Identify and classify the types of relationships between entity pairs by training the model;

[0156] S33, Automated Knowledge Completion:

[0157] 1) After the knowledge graph is constructed, the system automatically discovers and fills in the missing parts of the graph;

[0158] 2) Using graph database queries to identify isolated entities or incomplete relationship paths;

[0159] 3) Use web crawler technology to crawl relevant information from the Internet and extract new entities and relationships through information extraction technology;

[0160] 4) Apply machine learning models to predict and complete missing links, including the graph-based embedding model GraphSAGE;

[0161] S34. Knowledge graph storage:

[0162] 1) Use Neo4j graph database to store knowledge graph;

[0163] 2) When importing a large amount of data, use the LOADCSV command to import data and use the strategy of trading space for time to optimize read and write performance;

[0164] In addition, indexes are created for commonly queried attributes to improve query efficiency. At the same time, UNIQUE constraints are used to avoid data redundancy and ensure data consistency.

[0165] S4. Use modular retrieval augmented generation method, namely Retrieval Augmented Generation (RAG), to retrieve and recommend knowledge in enterprise knowledge base;

[0166] S41. Knowledge retrieval:

[0167] By intelligently integrating keyword search, semantic search, and vector search technologies, the vector encoding model is used to convert user input into a vector, calculate the similarity between question embedding and document block embedding, and select the top K document blocks with the highest similarity as enhanced context information for the current question;

[0168] S42. Find similar memories:

[0169] Create an infinite memory pool, using the memory capacity of the large model itself to enhance retrieval and find the memory most similar to the current input;

[0170] S43, additional generation:

[0171] Use LLM to generate the required contextual search;

[0172] Remove redundancy and noise from the retrieved content;

[0173] S44, Task Adaptation:

[0174] Focus on converting RAG to suit downstream tasks;

[0175] In this embodiment, UPRISE automatically retrieves hints for a given zero-shot task input from a pre-built data pool, enhancing the generalizability across tasks and models;

[0176] S45, Alignment:

[0177] Add a trainable Adapter module to the retriever to alleviate the alignment problem between query and text;

[0178] S46. Verification:

[0179] An additional verification module is introduced after retrieving documents to evaluate the relevance between the retrieved documents and the query, thus enhancing the robustness of RAG;

[0180] S5. Regularly update and maintain the enterprise knowledge base:

[0181] Specifically, S51, new knowledge discovery:

[0182] Utilize web crawler technology to regularly scan predefined information sources, including industry websites, professional databases, and social media, to capture the latest content;

[0183] Use API interfaces to achieve real-time connection with external data sources and automatically import updated data and information;

[0184] S52. New knowledge archive:

[0185] Apply text mining and NLP techniques to assess the relevance and value of collected content for sentiment analysis and topic identification;

[0186] Use NLP to automatically classify and label new knowledge and file it into the correct category of the knowledge base;

[0187] S53, New Knowledge Review:

[0188] Deploy machine learning models to identify outdated or inaccurate information. The models are trained on historical data to understand the timeliness and accuracy of the content.

[0189] Use deep learning technology to deeply understand and review content;

[0190] S54. Version control and update:

[0191] Adopt the version control system Git to automatically record each content change and ensure the traceability of historical versions;

[0192] Implement automated deployment processes to ensure seamless release of knowledge base updates and reduce human errors.

[0193] The embodiments described above are merely descriptions of preferred implementation modes of the present invention and are not intended to limit the scope of the present invention. Without departing from the principles and essence 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 scope of protection determined by the claims of the present invention.

Claims

1. A method for constructing an enterprise-level knowledge base based on a large model, characterized in that: The specific steps include: S1. Obtain the original data of the enterprise and preprocess it to obtain the first target data; S2. Extracting and representing knowledge based on the first target data to obtain second target data; S3, constructing and storing a knowledge graph based on the Renduda model and the second target data to obtain an enterprise-level knowledge base; S4. Use modular retrieval enhancement generation method to perform knowledge retrieval and recommendation on enterprise knowledge base; S5. Regularly update and maintain the enterprise-level knowledge base.

2. The method for constructing an enterprise-level knowledge base based on a large model according to claim 1, characterized in that: The step S1 specifically includes: S11. Obtaining the original data of the enterprise: 1) Obtain the electronic document data, database data, business report data, market data, customer feedback data and product information data within the enterprise through the API interface; 2) Use web crawlers to obtain external data on papers and journals, industry reports, news and information, and social media; S12, storing the obtained raw data in a unified manner in the unstructured database MongoDB; S13, preprocessing the original data to obtain first target data, the process includes: 1) Identify and correct invalid or erroneous raw data entries; Identify and delete invalid or erroneous raw data entries; Use the aggregation framework and query operations in MongoDB to remove duplicate records and clean the raw data; 2) Use the $set update operation in MongoDB to update the fields of the original data to a unified format to achieve normalization of the original data; 3) Determine the missing values ​​of the original data and fill the missing values ​​using the $ifNull operator in MongoDB; 4) Use the type conversion function in MongoDB to convert the string in the original data into a date type or the text into a numeric value; 5) Use the $lookup aggregation phase in MongoDB to connect and integrate data scattered in different database tables; 6) Use the create Index command to create a single key index or a composite index for the original data; 7) Determine the shard key and use the sh.add Shard and sh.shard Collection commands to prepare the shard collection; 8) Use the mongoexport tool to export data from MongoDB to CSV format for further external processing; use the mongoimport tool to import the processed data into MongoDB; 9) Place frequently accessed fields at the top of MongoDB documents to speed up document access and optimize document structure.

3. The method for constructing an enterprise-level knowledge base based on a large model according to claim 1, characterized in that: The step S2 specifically includes: S21, entity recognition: using NLP methods to identify named entities in the text of the first target data, and improving the accuracy of entity recognition through the Random Model; S22, relationship identification; S23, attribute value extraction; S24. Knowledge representation.

4. The method for constructing an enterprise-level knowledge base based on a large model according to claim 3, characterized in that: The step S22 specifically includes: 1) Generation of candidate items for relation recognition: After entity recognition, search for entity pairs in the text and filter and generate candidate items for relation recognition; 2) Feature extraction: Use machine learning methods to extract part-of-speech tagging features, dependency path features, and context information features from the candidate items for relationship recognition, and use them as input data for training the classification model; 3) Model training: Use the Renduda model to perform model learning and training to output the relationship between entity pairs; 4) Relationship classification: Apply a classification model to each candidate entity pair to determine whether there is a predefined relationship type between them and classify the relationship type; 5) Model optimization and evaluation: Continuously optimize model parameters during model training and evaluate model performance through evaluation indicators such as precision, recall, and F1 score; 6) Handling entity relationship overlap: The TP Linker model is used to convert the joint extraction task into a token pair linking problem. The start and end positions of the entities are marked using the token linking matrix, and the entity relationships are aligned using the hand shake tagging scheme, effectively solving the entity relationship overlap problem.

5. The method for constructing an enterprise-level knowledge base based on a large model according to claim 3, characterized in that: The step S23 specifically includes: 1) Feature extraction: Extract features from text data based on a variant of long short-term memory network, namely bidirectional LSTM or Bi-LSTM. Bi-LSTM can capture the contextual information in the text and provide rich contextual features for the recognition of attribute values. 2) Multiple instance learning: In the multiple instance learning framework, each entity may correspond to multiple instances, i.e., text descriptions. The algorithm can learn from these instances and identify the attribute values ​​related to the entity. 3) Multi-label classification: Since an entity may have multiple attribute values, attribute value extraction can be modeled as a multi-label classification problem; each attribute value is used as a label, and the model predicts all possible labels in the text; 4) Deep learning model training: Use the labeled training data to train the deep learning model. During the training process, the model learns to identify attribute values ​​from the text; 5) Attention mechanism: In order to further improve the performance of the model, the attention mechanism is introduced to enable the model to pay more attention to the text fragments related to the attribute value; 6) Model evaluation and optimization: Use accuracy, recall, and F1 score indicators to evaluate the performance of the model on the validation set, and adjust and optimize the model based on the evaluation results; 7) Application and deployment: Deploy the trained model to actual applications to automatically identify and extract attribute values ​​from new text data.

6. The method for constructing an enterprise-level knowledge base based on a large model according to claim 3, characterized in that: The step S24 specifically includes: 1) Multimodal named entity recognition: using image and text multimodal data to identify and link entities and improve recognition accuracy; 2) Multimodal Relation Extraction: Extract relations from text and images, and verify and supplement them in multimodal data to achieve a more comprehensive understanding of relations; 3) Multimodal event extraction: Identify events from multimodal data sources and extract event-related information, including time, location, and participants; 4) Multimodal entity alignment: align the information describing the same entity in different modalities to ensure the consistency of knowledge representation; 5) Multimodal entity linking and disambiguation: linking entities in multimodal data to the correct knowledge graph entities and solving entity disambiguation problems; 6) Multimodal knowledge graph completion: Use multimodal data to complete the missing links and attributes in the knowledge graph to improve the completeness of the knowledge graph; 7) Multimodal knowledge graph reasoning: Applying multimodal data for knowledge reasoning, including inferring attributes or relationships not explicitly mentioned in the text through image content; 8) Transformer-based encoder: The Transformer-based encoder is used to extract the initial features of the image, text, and triple surface forms to achieve effective encoding between modalities; 9) Interaction modeling between modalities: The interaction features between image and text modalities are learned through co-attention Transformer, and product attributes and attribute value information are aggregated on the target product entity through structural aggregation; 10) Multimodal pre-training task: Pre-training of image modality, text modality and knowledge modality, including masked object model, masked language model and link prediction model, to optimize the model’s understanding of multimodal information.

7. The method for constructing an enterprise-level knowledge base based on a large model according to claim 1, characterized in that: The step S3 specifically includes: S31. Entity linking and disambiguation: 1) Candidate entity generation: Generate a set of possible candidate entities through various methods, including dictionary matching, synonym expansion, etc. 2) Entity disambiguation: Using the deep learning model BERT, combined with contextual information and structured information in the knowledge graph, the most appropriate entity is selected for each mention. This includes using the attention mechanism to capture the most relevant part of the context to the entity, and considering the consistency between entities to achieve joint disambiguation. S32, Relationship Extraction: 1) Use conditional random fields (CRF) or Transformer-based models such as BERT to extract relations; 2) Identify and classify the types of relationships between entity pairs by training the model; S33, Automated Knowledge Completion: 1) After the knowledge graph is constructed, the system automatically discovers and fills in the missing parts of the graph; 2) Using graph database queries to identify isolated entities or incomplete relationship paths; 3) Use web crawler technology to crawl relevant information from the Internet and extract new entities and relationships through information extraction technology; 4) Apply machine learning models to predict and complete missing links, including the graph-based embedding model Graph SAGE; S34. Knowledge graph storage: 1) Use Neo4j graph database to store knowledge graph; 2) When importing large amounts of data, use the LOADCSV command to import data, and optimize read and write performance by trading space for time.

8. The method for constructing an enterprise-level knowledge base based on a large model according to claim 1, characterized in that: The step S4 specifically includes: S41, Knowledge Retrieval: By intelligently integrating keyword search, semantic search, and vector search technologies, the vector encoding model is used to convert user input into a vector, the similarity between question embedding and document block embedding is calculated, and the top K document blocks with the highest similarity are selected as enhanced context information for the current question; S42. Find similar memories: Create an infinite memory pool, use the memory capacity of the large model itself to enhance retrieval, and find the memory most similar to the current input; S43, additional generation: using LLM to generate the required contextual retrieval; removing redundancy and noise from the retrieved content; S44, Task Adaptation: Focus on transforming RAG to adapt to downstream tasks; S45, Alignment: Add a trainable Adapter module to the retriever to alleviate the alignment problem between query and text; S46. Verification: An additional verification module is introduced after retrieving documents to evaluate the relevance between the retrieved documents and the query and enhance the robustness of RAG.

9. The method for constructing an enterprise-level knowledge base based on a large model according to claim 1, characterized in that: The step S5 specifically includes: S51. New knowledge discovery: Use web crawler technology to regularly scan predefined information sources, including industry websites, professional databases, and social media to capture the latest content; use API interfaces to achieve real-time docking with external data sources and automatically import updated data and information; S52. Archiving new knowledge: Use text mining and NLP techniques to evaluate the relevance and value of collected sentiment analysis and topic identification content; use NLP to automatically classify and label, and archive new knowledge into the correct category of the knowledge base; S53. New Knowledge Review: Deploy machine learning models to identify outdated or inaccurate information. The models are trained based on historical data to understand the timeliness and accuracy of the content. Use deep learning technology to deeply understand and review the content. S54. Version control and update: Use the version control system Git to automatically record each content change and ensure the traceability of historical versions; implement automated deployment processes to ensure seamless release of knowledge base updates and reduce human errors.

Images