A natural resource document retrieval method and system based on large model

Abstract

The present invention provides a natural resource document retrieval method and system based on a large model, including the steps of data preprocessing, large model fusion, retrieval request processing, retrieval execution, result optimization and intelligent recommendation, user interactive feedback, etc. The method collects files, performs digital processing and key information annotation, uses natural language processing technology to extract key information, and constructs word vectors. A natural language processing model is selected for fine-tuning training, processes user retrieval requests, executes retrieval operations, and optimizes results and intelligently recommends based on semantic similarity. At the same time, a user interactive interface is provided to collect feedback to improve the performance of the retrieval system. The present invention can improve the accuracy and efficiency of natural resource document retrieval, and at the same time, enhance the personalized service capability and continuous optimization capability of the retrieval system through intelligent recommendation and user feedback mechanisms.

Description

Technical Field

[0001] The present invention relates to the field of information retrieval technology, and more specifically, to a natural resource document retrieval method and system based on a large model. Background Art

[0002] Existing natural resource document retrieval systems usually rely on keyword matching and simple database query technology. When dealing with a large number of diverse documents, these systems often find it difficult to accurately understand the user's search intent and the deep meaning of the documents. Due to the complexity of the document language and the widespread use of professional terms, traditional retrieval methods are difficult to achieve high-precision semantic matching, making it difficult for users to quickly find highly relevant document information. In addition, there are complex associations between documents, such as document evolution, supporting relationships, and cross-regional reference relationships. These associations are often ignored in traditional retrieval systems, limiting the relevance and comprehensiveness of retrieval results. With the development of big data and artificial intelligence technologies, especially the advancement of natural language processing technology, it has become possible to conduct deeper semantic understanding and association analysis of documents. However, existing technologies have not yet fully utilized these advanced technical means to achieve smarter and more accurate document retrieval and recommendation services.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: the existing retrieval system cannot effectively handle the complexity and diversity of files, lacks the ability to understand the deep semantics of files and analyze relevance, resulting in insufficient accuracy and relevance of retrieval results, and cannot meet users' needs for efficient and accurate file retrieval. Summary of the invention

[0004] The invention provides a natural resource document retrieval method and system based on a large model.

[0005] In a first aspect of the present invention, a natural resource document retrieval method based on a large model is provided, comprising: Collect natural resource documents from multiple regions, types and at different times, digitize them, convert them into a unified format and extract key information annotations. At the same time, use natural language processing technology to conduct a preliminary analysis of the document text content, mark professional terms, high-frequency words and important semantically related phrases, and construct a word vector representation of the text; Select a natural language processing model, and perform targeted fine-tuning training on it based on the collected and organized natural resource document annotated data samples to obtain a large model. During the training process, design tasks that meet the natural resource document retrieval scenario, use optimization algorithms to adjust model parameters, and set training hyperparameters, including training rounds and batch size; Receive the search request input by the user, and use the basic natural language processing process to pre-process the request statement, including lexical analysis, stop word removal and simple entity recognition. Then input the pre-processed search request into the large model that has completed fine-tuning training, and the large model will perform deep intent analysis and convert it into specific query logic that matches the database structure; According to the query logic output by the big model, search operations are performed in the natural resource file database that has undergone data preprocessing; The preliminary result set obtained from the retrieval is sorted by relevance, and the semantic similarity between the retrieval request and each file text is calculated again using a large model. After sorting according to the semantic similarity, a natural resource file knowledge graph is constructed to analyze and record the correlation between files. Combined with user personalized information, relevant files are intelligently recommended to users based on the correlation information of the file knowledge graph.

[0006] Furthermore, in the data preprocessing step, the key information includes the document title, issuing agency, issuing time, summary of core clauses, scope of application, specific classification of natural resources involved, and other associated document numbers.

[0007] Furthermore, in the retrieval request processing step, the preprocessing operation uses the NLTK toolkit to perform lexical analysis, the stop words are removed using a conventional stop word list, and entity recognition is performed by matching with the regional information pre-stored in the system.

[0008] Furthermore, in the retrieval execution step, if a relational database is used, the index mechanism includes a B-tree index and an inverted index, the semantic similarity calculation is based on the text vector representation output by the large model, calculated by the cosine similarity method, and the semantic similarity score of the file resources included in the preliminary retrieval result set exceeds 0.8.

[0009] Furthermore, in the result optimization and intelligent recommendation steps, the sorting algorithm is bubble sort, and the knowledge graph is constructed using the Neo4j graph database. The relationship information between files is extracted from the file text through manual analysis and natural language processing technology, including file evolution relationships, supporting relationships, and cross-regional reference relationships.

[0010] Furthermore, in the user interactive feedback step, the search result display interface is built using the Vue.js framework, and the feedback entry includes a satisfaction rating star button and an opinion and suggestion text input box.

[0011] In a second aspect of the present invention, a natural resource retrieval system based on a large model is provided, comprising: Data preprocessing module: used to build a natural resource document knowledge base, collect and organize documents, and annotate key information and construct word vectors; Large model fusion module: used to select a large language model and perform targeted fine-tuning training; Retrieval request processing module: used to receive user retrieval requests and perform preprocessing and intent analysis to generate query logic; Retrieval execution module: used to retrieve file resources in the database according to the query logic, and mine related files in combination with semantic similarity calculation; Result optimization and intelligent recommendation module: used to sort the relevance of search results and make intelligent recommendations.

[0012] Furthermore, the data preprocessing module includes a data collection unit and a key information labeling unit. The data collection unit collects files from multiple authoritative channels through a web crawler tool and performs format conversion.

[0013] Furthermore, the large model fusion module includes a model selection unit and a fine-tuning training unit. The model selection unit selects the Qwen 2.5 72B model and adjusts its input and output structure. The fine-tuning training unit uses labeled data samples to construct a training data set, uses the Adam optimization algorithm to perform fine-tuning training, and adjusts parameters according to verification results.

[0014] Furthermore, the training hyperparameters is to optimize the hyperparameters, obtain the optimized hyperparameters by using an optimization algorithm, perform a hyperparameter search according to the selected optimization method, and for each search, use cross-validation to evaluate the performance of the model.

[0015] Furthermore, in the retrieval operation performed in the natural resource file database that has undergone data preprocessing according to the query logic output by the big model, if a relational database is used, a pre-established index mechanism is used to locate the file record or article content; if a non-relational database is used, its own query method for processing text data is used for search, and at the same time, combined with the calculation results of semantic similarity by the big model, semantically related file resources are mined and extracted.

[0016] According to the above-mentioned embodiments of the present invention, at least the following beneficial effects are achieved: the use of a natural resource file retrieval method based on a large model can improve the accuracy and efficiency of retrieval. Through the in-depth analysis of the file text by natural language processing technology, the method can understand professional terms, high-frequency words and important semantically related phrases, and construct a word vector representation of the file text, thereby achieving in-depth intent analysis of the user's search request. In addition, the method fine-tunes the training of the large model and optimizes the algorithm to adjust the model parameters, so that the retrieval system can more accurately match user needs and file content, which can improve the relevance of the search results.

[0017] This method can also provide users with intelligent recommendation services by building a knowledge graph of natural resource files, analyzing and recording the relevance between files, and combining user personalized information. Through the user interactive feedback mechanism, the system can collect user evaluations and suggestions, and continuously optimize retrieval performance and user experience. This method can not only improve the intelligence level of file retrieval, but also enhance the system's adaptability and personalized service capabilities, providing users with a more intuitive and convenient retrieval platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which: Figure 1 A schematic diagram of a flow chart of a natural resource file retrieval method based on a large model provided by an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a natural resource document retrieval system based on a large model provided by an embodiment of the present invention; Figure 3 The schematic diagram schematically shows the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0020] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0021] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0022] Reference below Figure 1 , Figure 1 The following is a flow chart of a natural resource file retrieval method based on a large model provided by an embodiment of the present invention. Figure 1 As shown, a natural resource document retrieval method 100 based on a large model includes: Step 101, data preprocessing: collect natural resource documents from multiple regions, multiple types and released at different times, perform digital processing, convert to a unified format and extract key information annotations, and use natural language processing technology to perform preliminary analysis on the document text content, mark professional terms, high-frequency words and important semantically related phrases, and construct a word vector representation of the document text; Step 102, large model fusion: select a large language model with strong semantic understanding and generation capabilities in the field of natural language processing, and perform targeted fine-tuning training on it based on the collected and organized natural resource document annotated data samples. During the training process, design tasks that meet the natural resource document retrieval scenario, use appropriate optimization algorithms to adjust model parameters, and set reasonable training rounds, batch size and other hyperparameters; Step 103, retrieval request processing: receiving the retrieval request input by the user, first pre-processing the request statement using the basic natural language processing process, including lexical analysis, removal of stop words and simple entity recognition, and then inputting the pre-processed retrieval request into the large model that has completed fine-tuning training, and the large model performs deep intent analysis and converts it into specific query logic that matches the database structure; Step 104, retrieval execution: according to the query logic output by the big model, a search operation is performed in the natural resource file database that has undergone data preprocessing. If a relational database is used, a pre-established efficient indexing mechanism is used to quickly locate the file record or article content. If a non-relational database is used, its own query method suitable for processing text data is used for searching. At the same time, combined with the calculation results of semantic similarity by the big model, semantically related file resources are mined and extracted; Step 105, result optimization and intelligent recommendation: sort the relevance of the preliminary result set obtained by the search, use the big model to calculate the semantic similarity between the search request and each file text again, sort them according to the semantic similarity, build a natural resource file knowledge graph, analyze and record the relevance between files, combine user personalized information, and intelligently recommend relevant files to users based on the relevance information of the file knowledge graph; It should be noted that this embodiment involves a natural resource document retrieval method based on a large model, which includes steps such as data preprocessing, large model fusion, retrieval request processing, retrieval execution, result optimization and intelligent recommendation, and user interaction feedback. Here, data preprocessing refers to the process of digitizing the collected documents and extracting key information; large model fusion refers to the use of a large language model in the field of natural language processing for targeted training to adapt to specific document retrieval scenarios.

[0023] In this application, the big model refers to an advanced machine learning model based on natural language processing (NLP) technology, which has been fine-tuned and trained to adapt to a specific application scenario, namely the retrieval of natural resource documents.

[0024] Specifically, in the data preprocessing step, the files are first converted into electronic format through digital processing, and then the format is unified and converted to extract key information, such as file title, publishing agency, etc. At the same time, natural language processing technology is used to conduct a preliminary analysis of the file text content, mark professional terms, high-frequency words and important semantically related phrases, and construct a word vector representation of the file text. In the large model fusion step, the large language model in the field of natural language processing, such as Qwen 2.572B, is selected. Through targeted fine-tuning training, the model parameters are adjusted to adapt to the natural resource file retrieval scenario.

[0025] Preferably, in the retrieval request processing step, the preprocessing operation can use the NLTK toolkit to perform lexical analysis, the removal of stop words can use a conventional stop word list, and entity recognition can be matched with the regional information pre-stored by the system. In the retrieval execution step, if a relational database is used, efficient indexing mechanisms such as B-tree index and inverted index can be used to quickly locate file records or article content; if a non-relational database is used, the query method suitable for processing text data is used for search. At the same time, combined with the calculation results of semantic similarity by the large model, semantically related file resources are mined and extracted, and file resources with a semantic similarity score exceeding 0.8 will be included in the preliminary retrieval result set.

[0026] In some embodiments, in the data preprocessing step, the key information includes the document title, issuing agency, issuing time, summary of core clauses, scope of application, specific classification of natural resources involved, and other associated document numbers.

[0027] It should be noted that this implementation describes in detail the extraction of key information in the data preprocessing step, which includes the document title, issuing agency, issuing time, summary of core clauses, scope of application, specific classification of natural resources involved, and other related document numbers. Here, key information refers to those data points that are crucial for understanding and retrieving documents, which help to quickly locate and understand the main content and scope of application of the document.

[0028] Specifically, in the data preprocessing step, the extraction of key information involves an in-depth analysis of the document. For example, the document title is the name of the document, the issuing agency indicates the official source of the document, and the publishing time records the date of publication of the document. The core clause summary is a brief description of the most important clauses in the document, the scope of application defines the effective area or object of the document, and the specific classification of natural resources involved details the types of resources affected by the document, such as minerals, forests, etc. The other associated document numbers provide reference relationships between documents, which helps to build a network of connections between documents.

[0029] Preferably, the extraction of key information can be automated through natural language processing technology. For example, named entity recognition technology is used to automatically identify the issuing agency and the time of issuance, and the key terms summary is extracted through keyword extraction technology. The scope of application and the specific classification of natural resources involved can be marked through predefined classification systems and rules. Other related document numbers can be identified by comparing the document content with the existing document database.

[0030] Furthermore, machine learning methods, such as supervised learning, can be used to improve the accuracy of key information extraction by training models to identify and extract these key information points. In actual operations, manual review can also be combined to ensure the accuracy and completeness of the extracted information.

[0031] In some embodiments, in the large model fusion step, the optimization algorithm is the Adam algorithm, the large model is Qwen 2.5 72B, the training rounds are 10 rounds, the batch size is 32, and the initial value of the learning rate is 0.0001.

[0032] It should be noted that this embodiment describes in detail the specific operations in the large model fusion step, including the selection of optimization algorithms, the selection of large models, the setting of training rounds, batch size and learning rate. Here, optimization algorithms refer to mathematical methods used to adjust model parameters to improve model performance; large models refer to complex computational models that can understand and generate natural language text in the field of natural language processing.

[0033] Specifically, in the large model fusion step, the optimization algorithm selected is the Adam algorithm, which is an optimization method based on gradient descent, commonly used in deep learning, and is popular for its adaptive learning rate adjustment feature. The large model selected is Qwen 2.5 72B, a large model with 7.2 billion parameters that can handle complex language understanding and generation tasks. The number of training rounds is set to 10, which means that the entire training set will be traversed 10 times to update the model parameters. The batch size is 32, which refers to the number of data samples used to calculate the gradient before each parameter update. The initial value of the learning rate is set to 0.0001, which is the value used to adjust the parameter step size in the early stage of model training.

[0034] Preferably, the large model fusion step can be further refined or alternative solutions can be provided. For example, in addition to the Adam algorithm, other optimization algorithms such as SGD (Stochastic Gradient Descent) or RMSprop can also be considered to adapt to different training requirements. For the large model Qwen 2.5 72B, if resources are limited or in specific application scenarios, smaller-scale models can be considered, or model distillation techniques can be adopted to transfer the knowledge of the large model to small models. The number of training epochs and batch size can be adjusted according to the actual training effect to achieve better model performance and generalization ability. The learning rate can adopt a learning rate decay strategy, gradually decreasing as the training progresses to refine the solution space and improve the accuracy of the model. In addition, regularization techniques such as dropout or weight decay can be introduced to prevent the model from overfitting.

[0035] In some embodiments, in the retrieval request processing step, the preprocessing operation uses the NLTK toolkit for lexical analysis, the stop words are removed using a conventional stop word list, and entity recognition is performed by matching with the region information pre-stored in the system.

[0036] It should be noted that this embodiment mode details the preprocessing operation in the retrieval request processing step, including the specific implementation details of lexical analysis, stop word removal, and entity recognition. Here, lexical analysis refers to the process of decomposing natural language text into words, phrases, or other meaningful elements; stop words refer to commonly used words that are usually ignored in text processing, such as "de" (of), "shi" (is), etc.; entity recognition refers to identifying specific entities from the text, such as person names, locations, organization names, etc.

[0037] Specifically, in the retrieval request processing step, lexical analysis can be implemented using the NLTK toolkit, which is a widely used natural language processing library that provides rich lexical analysis tools. The operation of removing stop words can be completed by using a conventional stop word list, which contains common and less meaningful words in natural language processing. Entity recognition is achieved by matching with the region information pre-stored in the system, which means that the system will have a database containing various region names for identifying the geographical location information in the user's retrieval request.

[0038] Preferably, the retrieval request processing steps can be further refined or alternatives can be provided. For example, in addition to the NLTK toolkit, lexical analysis can also use other natural language processing tools such as SpaCy and Stanford NLP. The operation of removing stop words can be customized according to the terminology of specific fields to meet the retrieval needs of different fields. In addition to matching regional information, entity recognition can also be extended to other types of entities, such as organizations, events, etc., and more advanced named entity recognition (NER) models, such as BERT-based models, can be used to improve the accuracy and coverage of recognition. In addition, context-aware entity linking technology can be introduced to associate the identified entities with the corresponding entities in the knowledge base to enhance the semantic understanding ability of retrieval.

[0039] In some embodiments, in the retrieval execution step, if a relational database is used, the index mechanism includes a B-tree index and an inverted index, the semantic similarity calculation is based on the text vector representation output by the large model, calculated by the cosine similarity method, and the semantic similarity score of the file resources included in the preliminary retrieval result set exceeds 0.8.

[0040] It should be noted that this embodiment describes the specific technical details of the retrieval execution step, especially the indexing mechanism of the relational database and the calculation method of semantic similarity. Here, the indexing mechanism refers to the data structure used to quickly retrieve data in the database; the semantic similarity calculation refers to the method of evaluating the semantic similarity between two texts.

[0041] Specifically, in the retrieval execution step, if a relational database is used, two efficient indexing mechanisms, B-tree index and inverted index, can be used. B-tree index is a balanced tree structure used to quickly locate records in a database; inverted index is a data structure that maps words appearing in a text to their positions, and is commonly used in text searches. Semantic similarity calculations are based on the text vector representation output by the large model and are calculated using the cosine similarity method, which is a similarity measure that measures the angle between two non-zero vectors and is commonly used to evaluate the similarity of text data. In this embodiment, only when the semantic similarity score of a file resource exceeds 0.8 will it be included in the preliminary retrieval result set.

[0042] Preferably, the retrieval execution step can be further refined or alternatives can be provided. For example, in addition to B-tree index and inverted index, other types of index mechanisms such as hash index or full-text index can also be considered, which are suitable for different query types and data characteristics. The cosine similarity method can be combined with other similarity calculation methods, such as Jaccard similarity or Euclidean distance, to meet different semantic similarity calculation requirements.

[0043] Furthermore, machine learning methods can be introduced to optimize the calculation of semantic similarity, for example, by training a similarity evaluation model to more accurately capture the semantic relationship between file texts. In practical applications, the threshold of semantic similarity can also be dynamically adjusted according to the characteristics of the file and the user's search behavior to optimize the relevance of the search results.

[0044] In some embodiments, in the result optimization and intelligent recommendation steps, the sorting algorithm is bubble sort, and the knowledge graph is constructed using the Neo4j graph database, and the relationship information between files is extracted from the file text through manual analysis and natural language processing technology, including file evolution relationships, supporting relationships, and cross-regional reference relationships.

[0045] It should be noted that this embodiment describes the specific implementation details of the sorting algorithm and knowledge graph construction in the result optimization and intelligent recommendation steps. Here, the sorting algorithm refers to a computational method used to sort the result set according to a specific standard; the knowledge graph refers to a structured semantic knowledge base used to store and reason about the relationship between entities.

[0046] Specifically, in the result optimization and intelligent recommendation steps, the sorting algorithm used is bubble sort, which is a simple sorting algorithm that repeatedly traverses the sequence to be sorted, compares the size of each pair of adjacent elements, and swaps their positions when necessary. The knowledge graph is constructed using the Neo4j graph database, a high-performance graph database that can store complex network structure data and provide efficient graph query capabilities. Through manual analysis and natural language processing technology, the relationship information between files is extracted from the file text, including file evolution relationship, supporting relationship and cross-regional reference relationship. This relationship information is used to enrich the content of the knowledge graph.

[0047] Preferably, the result optimization and intelligent recommendation steps can be further refined or alternatives can be provided. For example, in addition to bubble sort, more efficient sorting algorithms, such as quick sort or merge sort, can also be considered to increase the speed of result set processing. The Neo4j graph database can be compared and selected with other types of graph databases such as Amazon Neptune or ArangoDB to meet different performance and cost requirements. When building a knowledge graph, more natural language processing technologies, such as dependency syntactic analysis and semantic role labeling, can be introduced to more accurately identify and extract relational information in file texts. In addition, machine learning methods, such as graph neural networks, can be used to automatically discover and infer the relationships between files, thereby further enhancing the intelligent recommendation capabilities of the knowledge graph.

[0048] In some embodiments, in the user interactive feedback step, the search result display interface is built using the Vue.js framework, and the feedback entry includes a satisfaction rating star button and an opinion and suggestion text input box.

[0049] It should be noted that this embodiment describes in detail the specific implementation details of the search result display interface and feedback collection mechanism in the user interaction feedback step. Here, the user interaction feedback step refers to the process of information exchange between the system and the user, the search result display interface refers to the interface for users to view the search results, and the feedback collection mechanism refers to the method by which the system collects users' evaluations and suggestions on the search results.

[0050] Specifically, in the user interaction feedback step, the search result display interface is built using the Vue.js framework, which is a progressive JavaScript framework for building user interfaces. It allows developers to build complex user interfaces in a componentized way. The feedback entry includes a satisfaction rating star button and a text input box for opinions and suggestions. The star button allows users to quickly express their satisfaction with the search results by clicking on the star, while the text input box allows users to enter more detailed feedback information.

[0051] Preferably, the user interaction feedback step can be further refined or alternatives can be provided. For example, in addition to the Vue.js framework, other front-end frameworks such as React or Angular can be considered to build the search result display interface. These frameworks have their own characteristics and can be selected according to project requirements and team familiarity. The satisfaction rating star button can be further refined into a dynamic star rating system, allowing users to slide between different levels of satisfaction. The opinion and suggestion text input box can add rich text editing functions, allowing users to submit feedback in a richer format. In addition, automated feedback analysis tools, such as natural language processing technology, can be introduced to automatically classify and summarize user feedback content, thereby driving system optimization more efficiently.

[0052] The above-mentioned embodiments of the present invention have the following beneficial effects: The natural resource file retrieval method based on a large model described in the present invention can improve the accuracy and efficiency of file retrieval. The method can achieve in-depth understanding and semantic matching of files through steps such as data preprocessing, large model fusion, retrieval request processing, retrieval execution, result optimization and intelligent recommendation, and user interactive feedback, so that users can quickly and accurately retrieve the required file information. In addition, by constructing a knowledge graph and calculating semantic similarity using a large model, the method can intelligently recommend relevant files to users and enhance the relevance and comprehensiveness of the retrieval results.

[0053] Furthermore, this method can effectively collect and utilize user feedback to continuously optimize the performance of the retrieval system. Through the user interaction feedback module, the system can collect users' evaluations and suggestions on the retrieval results, and then adjust and improve the large model and various modules of the retrieval process. This iterative optimization mechanism based on user feedback can continuously improve the intelligence level of the system and user satisfaction, making the file retrieval service more in line with the actual needs of users.

[0054] like Figure 2 As shown, a natural resource retrieval system 200 based on a large model in some embodiments includes: Data preprocessing module 201: used to build a natural resource document knowledge base, collect and organize documents, and perform key information annotation and word vector construction; Large model fusion module 202: used to select a large model and perform targeted fine-tuning training; Retrieval request processing module 203: used to receive user retrieval requests and perform pre-processing and intent analysis to generate query logic; Retrieval execution module 204: used to retrieve file resources in the database according to the query logic, and mine related files in combination with semantic similarity calculation; Result optimization and intelligent recommendation module 205: used to sort the search results by relevance and make intelligent recommendations; User interaction feedback module 206: used to provide a search result display interface and collect user feedback to drive system optimization.

[0055] It is understandable that the modules and references recorded in the natural resource document retrieval system 200 based on the large model Figure 1 Therefore, the operations, features and beneficial effects described above for the natural resource document retrieval method based on a large model are also applicable to the natural resource document retrieval system 200 based on a large model and the modules contained therein, and will not be repeated here.

[0056] In some embodiments, the data preprocessing module includes a data collection unit and a key information labeling unit. The data collection unit collects files from multiple authoritative channels through a web crawler tool and performs format conversion. The key information labeling unit manually formulates labeling rules and organizes professionals to label, and uses the Word2Vec model to construct word vectors.

[0057] It should be noted that this embodiment describes in detail the composition and functions of the data preprocessing module, including the data collection unit and the key information labeling unit. The data preprocessing module is the basis of the entire system and is responsible for building a natural resource file knowledge base, involving the collection and organization of files, as well as the labeling of key information and the construction of word vectors. Here, the data collection unit refers to the component used to obtain files from different sources, and the key information labeling unit refers to the part that performs in-depth analysis and information extraction on these files.

[0058] Specifically, the data collection unit in the data preprocessing module collects files from multiple authoritative channels through a web crawler tool and performs format conversion. A web crawler is an automated web browsing tool that can simulate user behavior to crawl web page content on the Internet. The key information annotation unit manually formulates annotation rules and organizes professionals to perform annotation, while using the Word2Vec model to construct word vectors. Word2Vec is a widely used word embedding model that can convert words in a text into numerical vectors for computer processing and analysis.

[0059] Preferably, the operation steps of the data preprocessing module can be further refined or alternative solutions can be provided. For example, in addition to using web crawler tools, the data collection unit can also obtain files directly from the data provider through the API interface to ensure the accuracy and timeliness of the data. The annotation rules of the key information annotation unit can be customized based on the specific structure and content characteristics of the file to improve the accuracy of the annotation. In addition to the Word2Vec model, other word embedding technologies such as GloVe or BERT can also be considered to capture richer semantic information.

[0060] Furthermore, machine learning algorithms can be introduced to automate the labeling process of key information, reduce manual intervention, and improve efficiency. When building word vectors, you can also consider using pre-trained language models, such as BERT or GPT, to further improve the model's ability to understand the semantics of file text.

[0061] In some embodiments, the large model fusion module includes a model selection unit and a fine-tuning training unit. The model selection unit selects the Qwen 2.5 72B model and adjusts its input and output structure. The fine-tuning training unit uses labeled data samples to construct a training data set, uses the Adam optimization algorithm to perform fine-tuning training, and adjusts parameters according to the verification results.

[0062] It should be noted that this embodiment describes in detail the composition and functions of the large model fusion module, including the model selection unit and the fine-tuning training unit. The large model fusion module is the core part of the system, responsible for selecting a suitable large model and performing targeted fine-tuning training to adapt to specific natural resource file retrieval tasks. Here, the model selection unit refers to the component that determines which pre-trained model to use, and the fine-tuning training unit refers to the part that further trains the selected model to optimize its performance.

[0063] Specifically, the model selection unit in the large model fusion module selected the Qwen 2.5 72B model and adjusted its input and output structure to adapt to specific retrieval tasks. Qwen 2.5 72B is a large pre-trained language model with 7.2 billion parameters that can handle complex natural language understanding and generation tasks. The fine-tuning training unit uses labeled data samples to build a training data set and uses the Adam optimization algorithm for fine-tuning training. At the same time, it adjusts parameters based on verification results to optimize model performance. The Adam optimization algorithm is an optimization method for adaptive learning rate, which is often used in deep learning to accelerate model convergence.

[0064] Preferably, the operation steps of the large model fusion module can be further refined or alternatives can be provided. For example, in addition to the Qwen 2.5 72B model, the model selection unit can also consider other pre-trained models, such as BERT or GPT series, to select the most suitable model based on the specific requirements of the task and the performance of the model. The training data set of the fine-tuning training unit can be further refined to include more document domain-specific data to improve the model's understanding of document text and retrieval accuracy. In addition to the Adam optimization algorithm, other optimization algorithms, such as SGD or RMSprop, can also be tried to find the optimization method that best suits the current task.

[0065] Furthermore, regularization techniques, such as dropout or weight decay, can be introduced to prevent model overfitting and improve the generalization ability of the model. During fine-tuning training, different learning rate decay strategies, such as exponential decay or step decay, can be set to dynamically adjust the learning rate and optimize the training effect.

[0066] Reference below Figure 3 , which shows a schematic diagram of a structure 300 of an electronic device suitable for implementing some embodiments of the present invention. The electronic devices in some embodiments of the present invention may include, but are not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 3The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0067] like Figure 3 As shown, the electronic device 300 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 to a random access memory (RAM) 303. In the RAM 303, various programs and data required for the operation of the electronic device 300 are also stored. The processing device 301, the ROM 302, and the RAM 303 are connected to each other via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0068] Typically, the following devices may be connected to the I / O interface 305: an input device 306 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 307 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 308 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 309. The communication device 309 may allow the electronic device 300 to communicate with other devices wirelessly or by wire to exchange data. Figure 3 The electronic device 300 is shown with various devices, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead. Figure 3 Each block shown in the figure may represent one device, or may represent multiple devices as required.

[0069] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0070] The above descriptions are only some preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A natural resource document retrieval method based on a large model, characterized in that: The following steps are involved: Collect natural resource documents from multiple regions, types and at different times, digitize them, convert them into a unified format and extract key information annotations. At the same time, use natural language processing technology to conduct a preliminary analysis of the document text content, mark professional terms, high-frequency words and important semantically related phrases, and construct a word vector representation of the text; Select a natural language processing model, and perform targeted fine-tuning training on it based on the collected and organized natural resource document annotated data samples to obtain a large model. During the training process, design tasks that meet the natural resource document retrieval scenario, use optimization algorithms to adjust model parameters, and set training hyperparameters, including training rounds and batch size; Receive the search request input by the user, and use the basic natural language processing process to pre-process the request statement, including lexical analysis, stop word removal and simple entity recognition. Then input the pre-processed search request into the large model that has completed fine-tuning training, and the large model will perform deep intent analysis and convert it into specific query logic that matches the database structure; According to the query logic output by the big model, search operations are performed in the natural resource file database that has undergone data preprocessing; The preliminary result set obtained from the retrieval is sorted by relevance, and the semantic similarity between the retrieval request and each file text is calculated again using a large model. After sorting according to the semantic similarity, a natural resource file knowledge graph is constructed to analyze and record the correlation between files. Combined with user personalized information, relevant files are intelligently recommended to users based on the correlation information of the file knowledge graph.

2. A natural resource file retrieval method based on a large model according to claim 1, characterized in that: In the data preprocessing step, the key information includes the document title, issuing agency, issuing time, summary of core clauses, scope of application, specific classification of natural resources involved, and other related document numbers.

3. The natural resource file retrieval method based on a large model according to claim 1 is characterized in that: In the retrieval request processing step, the preprocessing operation uses the NLTK toolkit to perform lexical analysis, the conventional stop word list is used to remove stop words, and entity recognition is performed by matching with the regional information pre-stored by the system.

4. A natural resource file retrieval method based on a large model according to claim 1, characterized in that: In the retrieval execution step, if a relational database is used, the index mechanism includes a B-tree index and an inverted index, the semantic similarity calculation is based on the text vector representation output by the large model, calculated by the cosine similarity method, and the semantic similarity score of the file resources included in the preliminary retrieval result set exceeds 0.

8.

5. A natural resource file retrieval method based on a large model according to claim 1, characterized in that: In the result optimization and intelligent recommendation steps, the sorting algorithm is bubble sort, and the knowledge graph is constructed using the Neo4j graph database. The relationship information between files is extracted from the file text through manual analysis and natural language processing technology, including file evolution relationships, supporting relationships, and cross-regional reference relationships.

6. A natural resource file retrieval method based on a large model according to claim 1, characterized in that: In the user interactive feedback step, the search result display interface is built using the Vue.js framework, and the feedback entry includes a satisfaction rating star button and an opinion and suggestion text input box.

7. A natural resource retrieval system based on a large model, characterized in that: It includes data preprocessing module, large model fusion module, search request processing module, search execution module, and result optimization and intelligent recommendation module; The data preprocessing module is used to build a natural resource document knowledge base, collect and organize documents, and perform key information annotation and word vector construction; The large model fusion module is used to select a large language model and perform targeted fine-tuning training; The search request processing module is used to receive user search requests and perform preprocessing and intent analysis to generate query logic; The retrieval execution module is used to retrieve file resources in the database according to the query logic, and mine related files in combination with semantic similarity calculation; The result optimization and intelligent recommendation module is used to sort the relevance of the search results and make intelligent recommendations.

8. The natural resource document retrieval system based on a large model according to claim 7 is characterized in that: The large model fusion module includes a model selection unit and a fine-tuning training unit. The model selection unit selects the Qwen 2.5 72B model and adjusts its input and output structure. The fine-tuning training unit uses labeled data samples to construct a training data set, uses the Adam optimization algorithm to perform fine-tuning training, and adjusts parameters according to verification results.

9. The natural resource document retrieval system based on a large model according to claim 1 is characterized in that: The training hyperparameters is to optimize the hyperparameters, obtain the optimized hyperparameters by using an optimization algorithm, perform a hyperparameter search according to the selected optimization method, and use cross-validation to evaluate the performance of the model for each search.

10. The natural resource document retrieval system based on a large model according to claim 1 is characterized in that: According to the query logic output by the big model, in the retrieval operation in the natural resource file database that has undergone data preprocessing, if a relational database is used, a pre-established index mechanism is used to locate the file record or article content; if a non-relational database is used, its own query method for processing text data is used for search. At the same time, combined with the calculation results of semantic similarity by the big model, semantically related file resources are mined and extracted.

Images