Archive information resource intelligent sharing method and system based on AI

Abstract

The invention relates to the technical field of archive information query, and discloses an AI-based archive information resource intelligent sharing method and system, and the method comprises the steps: firstly constructing a multi-dimensional data module according to a business data type, and the multi-dimensional data module comprises an archive system database, a high-dimensional document vector database and an archive business knowledge graph; and performing deep analysis on user query through a natural language processing tool to form a structured keyword list. Then, performing semantic analysis on the keyword list by utilizing a pre-trained archive information resource understanding large model, and generating an enhanced query text; the query text is recalled from the data modules of the multiple dimensions, candidate answers are obtained, and the candidate answers are fused and optimized through the general large language model. And finally, returning the optimal answer to the user. By introducing the AI technology, the accuracy and comprehensiveness of file information resource sharing are improved, the retrieval result is optimized, and the retrieval efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of archive information query, and in particular to an AI-based intelligent sharing method and system for archive information resources. Background Art

[0002] In the current information age, the management and sharing of archival information resources have become an indispensable part of all walks of life. At present, some archival information resource sharing methods based on keyword matching have been proposed, attempting to search in the archival database through the keywords entered by the user to find relevant archival information. However, the current keyword matching methods lack semantic understanding capabilities and cannot handle synonyms, near-synonyms or context-related information in user queries. This results in the retrieval results often being inaccurate and unable to meet the actual needs of users. For example, a user may use "project information" to query, while the archives may use different expressions such as "project files" or "project archives". Traditional keyword matching methods may therefore miss these relevant information.

[0003] Moreover, existing archival information resource sharing methods are often limited to a single database or document collection, and are unable to perform intelligent retrieval and integration across multiple dimensions of data modules. This limits the comprehensiveness and accuracy of retrieval results, making it difficult for users to obtain complete and accurate archival information. Summary of the invention

[0004] The purpose of the present invention is to provide an AI-based intelligent sharing method and system for archival information resources to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solution: an AI-based intelligent sharing method for archive information resources, the method comprising:

[0006] Step 1: Based on the business data type, pre-build data modules of multiple dimensions, including archive system database module, high-dimensional document vector database module and archive business knowledge graph module;

[0007] Step 2: receiving an archive information resource query request input by a user, and obtaining a first query request text;

[0008] Step 3: parse the first query request text using a natural language processing tool, including word segmentation, part-of-speech tagging, dependency syntax analysis, and keyword extraction, to form a structured keyword list;

[0009] Step 4: passing the structured keyword list into a pre-trained archival information resource understanding model, and receiving the model output result as a second query request text; the archival information resource understanding model is used to parse the semantics of the keyword list and generate an enhanced query text;

[0010] Step 5: Recall the second query request text from the data modules of multiple dimensions to obtain multiple candidate answers;

[0011] Step 6: Encapsulate the second query request text and multiple candidate answers into a question-answer context text according to a preset question-answer template, further pass the question-answer context text into a general large language model, and use the natural language understanding ability of the general large language model to merge and optimize multiple candidate answers;

[0012] Step 7: Get the output of the general large language model as the optimal answer, and return the optimal answer to the current user.

[0013] Preferably, in step 3, the TF-IDF algorithm and the TextRank algorithm are combined to extract keywords. First, the importance weight of each word is calculated using the TF-IDF algorithm, and then the TextRank algorithm is used to iteratively calculate the graph structure constructed based on the dependency syntax tree, and finally the words with higher weights are selected as keywords to form a structured keyword list; for word wi, its TF-IDF value is:

[0014]

[0015] Among them, TF(wi) is the frequency of occurrence of word wi in the text, DF(wi) is the number of documents containing word wi, and N is the total number of documents.

[0016] Preferably, the Transformer architecture is used as the basis for pre-training the large model for understanding archival information resources. The architecture includes multiple encoder and decoder layers, each layer of which is composed of a multi-head self-attention mechanism and a fully connected neural network to capture the contextual information and semantic features of the input text. The steps of model construction include:

[0017] S1: Collect and organize text data of archival information resources in the archives, including historical archives, policy documents, and business reports, as the corpus for the pre-training model;

[0018] S2: Clean, remove duplicates and segment the corpus to form a text sequence for pre-training;

[0019] S3: During the pre-training process, a part of the words in the text sequence in the corpus is randomly masked, and the remaining word sequence is used as input to train the model to predict the masked words. The cross entropy loss function is used as the optimization target, and the model parameters are updated through the gradient descent algorithm;

[0020] Suppose the input text sequence is X = {x1, x2, ..., x n}, some words are masked to get X masked, the probability distribution of the masked words predicted by the model is P model (XX masked ), the distribution of the real masked words is P true (X), the cross entropy loss function is:

[0021]

[0022] S4: After pre-training is completed, the model is fine-tuned on a specific archival information resource understanding task to adapt to different query requirements and business scenarios; during the fine-tuning process, the model output layer is adjusted according to the task requirements, and the corresponding loss function and optimization algorithm are used for training;

[0023] S5: Evaluate the performance of the pre-trained archival information resource understanding large model on the test set, including precision, recall, and F1 score metrics.

[0024] Preferably, in step 4, the keyword list is input into a pre-trained archival information resource understanding model, and the model performs semantic analysis on the keyword list through its internal neural network structure, including identification of relationships between keywords, labeling of semantic roles, and integration of contextual information; based on the semantic analysis results of the keyword list by the model, an enhanced query text is generated, and the enhanced query text not only contains the key information in the user's original query request, but also is expanded and refined through the semantic understanding ability of the model.

[0025] Preferably, in step 5, according to the enhanced query text generated in step 4, a search is performed from pre-constructed data modules of multiple dimensions, the data modules include an archive system database module, a high-dimensional document vector database module and an archive business knowledge graph module, and the data modules of multiple dimensions respectively store different types of archive information resources; the candidate answers recalled from each data module are merged to form a set containing multiple candidate answers.

[0026] Preferably, in each data module, a corresponding retrieval algorithm and matching mechanism are used to recall candidate answers related to the query text:

[0027] For the archive system database module, SQL query statements are used for retrieval;

[0028] For the high-dimensional document vector database module, vector similarity calculation is used to recall similar documents;

[0029] For the archive business knowledge graph module, graph traversal algorithm or semantic matching algorithm is used to find knowledge points related to the query.

[0030] Preferably, in S3, the formula for updating the parameters of the gradient descent algorithm is:

[0031]

[0032] Among them, θ represents the model parameters, including all weights and biases that the model needs to learn; α represents the learning rate, which is used to control the step size of parameter update; It represents the gradient of the loss function J(θ) with respect to the parameter θ, which is a vector pointing to the direction in which the loss function grows fastest; := represents the assignment operation, that is, updating the value of the parameter θ.

[0033] Preferably, in step 6, the constructed question-answer context text is passed into a general large language model, and the natural language understanding ability of the general large language model is used to fuse multiple candidate answers; during the fusion process, the general large language model analyzes the logical relationship, content similarity and information complementarity between the candidate answers, and generates a fused answer that integrates the advantages of each candidate answer.

[0034] Preferably, an AI-based intelligent sharing system for archive information resources comprises:

[0035] A data module construction unit is used to pre-construct data modules of multiple dimensions according to business data types;

[0036] A user request receiving unit, used to receive an archive information resource query request input by a user and obtain a first query request text;

[0037] A natural language processing unit, configured to parse the first query request text using a natural language processing tool, wherein the parsing process includes word segmentation, part-of-speech tagging, dependency syntax analysis, and keyword extraction to form a structured keyword list;

[0038] A semantic parsing unit, including a pre-trained archival information resource understanding model, is used to receive the structured keyword list, parse the semantics of the keyword list, and generate an enhanced query text as the second query request text;

[0039] A candidate answer recall unit, used to recall the second query request text from the data modules of multiple dimensions and obtain multiple candidate answers;

[0040] An answer fusion optimization unit is used to encapsulate the second query request text and multiple candidate answers into a question and answer context text according to a preset question and answer template, and pass the question and answer context text into the general large language model, and use the natural language understanding ability of the general large language model to fuse and optimize the multiple candidate answers;

[0041] The output unit is used to obtain the output of the general large language model, take it as the optimal answer, and return the optimal answer to the current user.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention uses natural language processing tools and a pre-trained archival information resource understanding model to perform in-depth analysis and semantic expansion of user query requests. It can not only accurately identify keywords in user queries, but also understand their synonyms, near-synonyms and context-related information, thereby greatly improving the accuracy and comprehensiveness of queries.

[0044] Traditional methods are often limited to a single database or document collection, while the present invention constructs data modules of multiple dimensions, including an archive system database module, a high-dimensional document vector database module, and an archive business knowledge graph module, enabling the system to perform intelligent retrieval and fusion across multiple dimensions, thereby obtaining more complete and accurate archive information to meet users' diverse query needs.

[0045] By introducing a universal large language model, the present invention can fuse and optimize multiple candidate answers, which can not only improve the accuracy and relevance of the answers, but also make the answers more in line with the user's natural language habits, thereby improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A step diagram of the AI-based intelligent sharing method for archive information resources according to the present invention;

[0047] Figure 2 A flowchart for constructing a large model for understanding archival information resources;

[0048] Figure 3 A flowchart for generating candidate answers;

[0049] Figure 4 Flowchart for generating fused answers. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] See also Figure 1-4 The present invention provides an AI-based intelligent sharing method for archive information resources, the method comprising:

[0052] Step 1: Pre-build data modules of multiple dimensions according to the business data type. The data modules include the archive system database module, which is used to store structured archive information; the high-dimensional document vector database module, which is used to store unstructured or semi-structured archive documents and represent them in a vectorized way for efficient similarity calculation; and the archive business knowledge graph module, which is used to store entities, attributes and relationships in the archive business field, providing a basis for semantic understanding and reasoning.

[0053] Step 2: Receive the archive information resource query request input by the user and obtain the first query request text.

[0054] Step 3: Use natural language processing tools to parse the first query request text, including word segmentation, which is to split the continuous text string into independent words; part-of-speech tagging, assigning a part-of-speech tag to each word; dependency syntax analysis, building a dependency syntax tree, and clarifying the dependency relationship between words; and keyword extraction, combining algorithms such as TF-IDF and TextRank to extract keywords that can represent the user's query intent from the text to form a structured keyword list.

[0055] Step 4: The structured keyword list is passed into the pre-trained archival information resource understanding model to parse the semantics of the keyword list and generate enhanced query text. The enhanced query text not only contains the keywords entered by the user, but also adds synonyms, near synonyms or context-related information related to the query intent through the semantic understanding ability of the model, thereby improving the accuracy and comprehensiveness of the query.

[0056] Step 5: The system recalls the second query request text from the data modules of multiple dimensions and obtains multiple candidate answers. This process is achieved by searching and matching in the archive system database module, the high-dimensional document vector database module, and the archive business knowledge graph module. The system will search for candidate answers related to the enhanced query text in each data module.

[0057] Step 6: Encapsulate the second query request text and multiple candidate answers into a question-and-answer context text according to the preset question-and-answer template. The question-and-answer context text contains the question queried by the user and multiple candidate answers. Then, the question-and-answer context text is passed to the general large language model. The general large language model has a strong natural language understanding ability and can perform in-depth analysis and processing of the question-and-answer context text. Through the reasoning and calculation of the model, multiple candidate answers are integrated and optimized to generate a more accurate and comprehensive answer.

[0058] Step 7: Get the output of the general large language model as the optimal answer, and return the optimal answer to the current user.

[0059] The present invention will be further described below in conjunction with Examples 1 to 3:

[0060] Embodiment 1:

[0061] In step 3, the TF-IDF algorithm and the TextRank algorithm are combined to extract keywords.

[0062] First, the TF-IDF (Term Frequency-Inverse Document Frequency) is calculated for the first query request text after word segmentation, part-of-speech tagging and dependency syntax analysis. Taking a sample text "How to query personal information in the archive system?" as an example, first count the frequency of occurrence (TF) of each word in the text, and calculate the inverse document frequency (IDF) of each word in the entire archive information resource corpus. For the word "personal information", its TF value in the sample text is relatively high, and it is assumed that the DF value in the corpus is relatively low, so its TF-IDF value will be high, indicating the importance of this word in distinguishing and expressing the query request. For word wi, its TF-IDF value is:

[0063]

[0064] Among them, TF(wi) is the frequency of occurrence of word wi in the text, DF(wi) is the number of documents containing word wi, and N is the total number of documents.

[0065] The TextRank algorithm is used to further extract keywords. The TextRank algorithm is a graph-based sorting algorithm that uses the co-occurrence relationship between words to build a graph structure and iteratively calculates the weight of the words. In the specific implementation, first, based on the results of dependency syntactic analysis, a similarity matrix between words is constructed. For the sample text, words such as "how", "query", "archive system", "in", and "personal information data" can be used as nodes, and the edges are determined based on the dependency relationship between them (such as subject-predicate relationship, verb-object relationship, etc.), and the weight of the edge (that is, the similarity between words) is calculated.

[0066] Initialize a weight vector, where the initial weight of each word can be set to the same (e.g., uniform distribution). Through iterative calculation, the weight of each word is continuously updated until the change in the weight vector is less than the preset threshold and reaches a convergence state. During the iteration process, the weight of each word is updated according to the weight and similarity of its adjacent words. The higher the weight, the more likely it is to be a keyword.

[0067] Finally, according to the converged weight vector, words with higher weights are selected as keywords. For the sample text, "query", "archive system" and "personal information data" can be selected as keywords because they have higher TF-IDF values ​​and TextRank weights, which can accurately represent the intent and theme of the query request.

[0068] Embodiment 2:

[0069] The following is a detailed explanation of the construction process of the archival information resource understanding model with examples:

[0070] The Transformer architecture is used as the basis of the large model. Transformer is composed of multiple encoder and decoder layers, each of which contains a multi-head self-attention mechanism and a fully connected neural network. The multi-head self-attention mechanism can capture different semantic features in the input text, while the fully connected neural network is used to perform nonlinear transformation and fusion on the captured features. This design enables the model to efficiently process long text sequences and capture contextual information in the text. The steps of model construction include:

[0071] S1: Collect and organize text data of archival information resources in archives, including historical archives, policy documents, and business reports, as the corpus for pre-training models.

[0072] S2: Clean, remove duplicates and segment the corpus to form a text sequence for pre-training.

[0073] S3: During the pre-training process, randomly mask some words in the text sequence in the corpus. For example, for the text sequence "archival information resource sharing is an important way to improve the efficiency of archive management", the two words "improve" and "efficiency" may be randomly masked to obtain the masked text sequence "archival information resource sharing is an important way to ____archival management ____". Then, use the remaining word sequence as input to train the model to predict the masked words. During the training process, the cross entropy loss function is used as the optimization target, and the model parameters are updated through the gradient descent algorithm.

[0074] Suppose the input text sequence is X = {x1, x2, ..., x n}, some words are masked to get X masked , the probability distribution of the masked words predicted by the model is P model (X|X masked ), the distribution of the real masked words is P true (X), the cross entropy loss function is:

[0075]

[0076] The formula for updating parameters of the gradient descent algorithm is:

[0077]

[0078] Among them, θ represents the model parameters, including all weights and biases that the model needs to learn; α represents the learning rate, which is used to control the step size of parameter update; It represents the gradient of the loss function J(θ) with respect to the parameter θ, which is a vector pointing to the direction in which the loss function grows fastest; := represents the assignment operation, that is, updating the value of the parameter θ.

[0079] S4: After pre-training is completed, the model is fine-tuned on a specific archival information resource understanding task. Taking the user query "How to query personal archives?" as an example, the query request is converted into a text sequence and input into the pre-trained model. Then, the model output layer is adjusted according to the task requirements so that it can generate enhanced query text related to the query request. For example, the model may output "Steps and methods for querying personal archives". During the fine-tuning process, the corresponding loss function and optimization algorithm are used for training to improve the performance of the model on specific tasks.

[0080] S5: Evaluate the performance of the pre-trained archival information resource understanding model on the test set. The performance of the model is comprehensively evaluated using indicators such as accuracy, recall, and F1 score. The evaluation results show that the model performs well in the archival information resource understanding task, can accurately understand the semantics of user query requests, and generate high-quality enhanced query text.

[0081] In step 5, three data modules are mainly relied on: archive system database module, high-dimensional document vector database module and archive business knowledge graph module. Each module stores different types of archive information resources and is equipped with corresponding retrieval algorithms and matching mechanisms.

[0082] Archives system database module: stores structured archive data, such as basic archive information, metadata, etc. This module supports SQL query statements and can accurately retrieve data records that match the query conditions.

[0083] High-dimensional document vector database module: stores unstructured archival documents, such as PDF, Word and other files. Archival documents are converted into high-dimensional vectors and stored in the vector database. Through vector similarity calculation, documents with similar content to the query text can be recalled.

[0084] Archives business knowledge graph module: stores knowledge points, entities and their relationships in the archives business field. This module represents knowledge in the form of a graph and supports graph traversal algorithms and semantic matching algorithms to find knowledge points and entities related to the query.

[0085] Data module retrieval:

[0086] Archive system database module retrieval: Based on the enhanced query text, construct SQL query statements. For example, SELECT*FROM Archives WHERE ProjectName = 'XX project' AND ArchiveType = 'project background'. Execute this query statement to retrieve background archives related to the XX project from the database.

[0087] High-dimensional document vector database module retrieval: The enhanced query text is converted into a vector and similarity is calculated with the document vector in the database. For example, using cosine similarity as a metric, the top N documents with the highest similarity to the query vector are found as candidate answers.

[0088] Archives business knowledge graph module retrieval: Find relevant entities and knowledge points in the knowledge graph based on the keywords and phrases in the enhanced query text. For example, use the graph traversal algorithm to start from the "XX project" node, find directly related nodes (such as "project background", "project leader", etc.), and further obtain relevant information.

[0089] Candidate answer fusion: The candidate answers recalled from each data module are fused to form a set containing multiple candidate answers.

[0090] Assume that the first query request text input by the user is "Search for files about XX project". After processing in step 4, the generated enhanced query text is "Search for files of XX project, including project background, project progress and project leader information".

[0091] In step 5, search each data module separately:

[0092] Retrieve the background files, progress records and project leader information related to the XX project from the archive system database module;

[0093] Recall documents related to the XX project from the high-dimensional document vector database module, such as project reports, meeting minutes, etc.;

[0094] Find entities and knowledge points related to the XX project from the archive business knowledge graph module, such as project name, project time, project leader, etc.

[0095] Embodiment 3:

[0096] In step 6, in order to achieve effective integration of multiple candidate answers, a general large language model plus question-answer context text construction process is adopted to integrate key information in the candidate answers. The implementation method is described below with specific examples.

[0097] Based on the multiple candidate answers obtained in step 5, a question-answering context text is constructed, which not only contains the user's original query request (i.e., the second query request text), but also lists all candidate answers one by one. In order to clearly indicate the relationship between these answers, a short description is added before each candidate answer, such as "Candidate answer 1:", "Candidate answer 2:", etc., so that the general large language model can recognize and distinguish different answers.

[0098] The constructed question-answer context text is passed into the general large language model, which has powerful natural language understanding capabilities and can deeply analyze the semantics, grammar, and contextual information in the text.

[0099] The general large language model first analyzes the content of each candidate answer to understand its meaning and context. Then, the model compares the logical relationships between the candidate answers to identify which answers support each other and which answers are contradictory. For mutually supporting answers, the model will try to merge their content to form a more complete and accurate description. For contradictory answers, the model will judge based on their authority and credibility, and select the most reliable answer as part of the fusion result.

[0100] In addition to logical relationship analysis, the general large language model also calculates content similarity and information complementarity between candidate answers. Content similarity is used to measure the overlap and consistency between answers, while information complementarity is used to evaluate whether the answers contain different but complementary information. With these two indicators, the model can more accurately determine which answers should be merged and how to merge them.

[0101] Finally, the general large language model generates a fusion answer based on the analysis results. The fusion answer combines the advantages of each candidate answer, removes repeated and redundant information, and forms a more accurate, comprehensive and easy-to-understand answer.

[0102] Take the user's query "relevant archives about a historical event" as an example. Assume that in step 5, the following three candidate answers are recalled from the archive system database module, the high-dimensional document vector database module, and the archive business knowledge graph module respectively:

[0103] Candidate answer A: "This historical event took place in XX, the main participants included..., and the important events were..."

[0104] Candidate answer B: "Regarding this historical event, the archive system records..., which describes in detail..."

[0105] Candidate answer C: "In the archive business knowledge graph, this historical event is associated with..., and its background information is..."

[0106] Next, the three candidate answers are packaged together with the second query request text into a question-answer context text in the following format:

[0107] User query: Related files about a historical event

[0108] Candidate answer A: This historical event took place in XX, the main participants included…, and the important events were…

[0109] Candidate answer B: Regarding this historical event, the archive system records..., which describes in detail...

[0110] Candidate answer C: In the archive business knowledge graph, this historical event is associated with..., and its background information is...

[0111] This question-answer context text is then passed into the general large language model. The general large language model uses its powerful natural language understanding capabilities to analyze the logical relationship, content similarity, and information complementarity between candidate answers. It recognizes that candidate answer A provides basic information about the historical event, candidate answer B describes the record in the archive system in detail, and candidate answer C provides background information associated with the historical event.

[0112] Based on these analyses, the general large language model generates a fusion answer that combines the advantages of each candidate answer, including basic information and detailed descriptions of historical events, and also links to relevant background information. The fusion answer is fluent and comprehensive, meeting the user's query needs.

[0113] An AI-based intelligent sharing system for archive information resources, the system comprising:

[0114] A data module construction unit is used to pre-construct data modules of multiple dimensions according to business data types;

[0115] A user request receiving unit, used to receive an archive information resource query request input by a user and obtain a first query request text;

[0116] A natural language processing unit, configured to parse the first query request text using a natural language processing tool, wherein the parsing process includes word segmentation, part-of-speech tagging, dependency syntax analysis, and keyword extraction to form a structured keyword list;

[0117] A semantic parsing unit, including a pre-trained archival information resource understanding model, is used to receive the structured keyword list, parse the semantics of the keyword list, and generate an enhanced query text as the second query request text;

[0118] A candidate answer recall unit, used to recall the second query request text from the data modules of multiple dimensions and obtain multiple candidate answers;

[0119] An answer fusion optimization unit is used to encapsulate the second query request text and multiple candidate answers into a question and answer context text according to a preset question and answer template, and pass the question and answer context text into the general large language model, and use the natural language understanding ability of the general large language model to fuse and optimize the multiple candidate answers;

[0120] The output unit is used to obtain the output of the general large language model, take it as the optimal answer, and return the optimal answer to the current user.

[0121] For the implementation of the system, please refer to the above embodiments.

[0122] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0123] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent sharing method for archival information resources, characterized in that: The method comprises: Step 1: Based on the business data type, pre-build data modules of multiple dimensions, including archive system database module, high-dimensional document vector database module and archive business knowledge graph module; Step 2: receiving an archive information resource query request input by a user, and obtaining a first query request text; Step 3: parse the first query request text using a natural language processing tool, including word segmentation, part-of-speech tagging, dependency syntax analysis, and keyword extraction, to form a structured keyword list; Step 4: passing the structured keyword list into a pre-trained archival information resource understanding model, and receiving the model output result as a second query request text; the archival information resource understanding model is used to parse the semantics of the keyword list and generate an enhanced query text; Step 5: Recall the second query request text from the data modules of multiple dimensions to obtain multiple candidate answers; Step 6: Encapsulate the second query request text and multiple candidate answers into a question-answer context text according to a preset question-answer template, further pass the question-answer context text into a general large language model, and use the natural language understanding ability of the general large language model to merge and optimize multiple candidate answers; Step 7: Get the output of the general large language model as the optimal answer, and return the optimal answer to the current user.

2. The AI-based intelligent sharing method for archival information resources according to claim 1 is characterized in that: In step 3, the TF-IDF algorithm and the TextRank algorithm are combined to extract keywords. First, the TF-IDF algorithm is used to calculate the importance weight of each word, and then the TextRank algorithm is used to iteratively calculate the graph structure constructed based on the dependency syntax tree. Finally, the words with higher weights are selected as keywords to form a structured keyword list; for word wi, its TF-IDF value is: Among them, TF(wi) is the frequency of occurrence of word wi in the text, DF(wi) is the number of documents containing word wi, and N is the total number of documents.

3. The AI-based intelligent sharing method for archival information resources according to claim 2 is characterized in that: The Transformer architecture is used as the basis for pre-training the large model for understanding archival information resources. The architecture contains multiple encoder and decoder layers. Each layer consists of a multi-head self-attention mechanism and a fully connected neural network to capture the contextual information and semantic features of the input text. The steps of model construction include: S1: Collect and organize text data of archival information resources in the archives, including historical archives, policy documents, and business reports, as the corpus for the pre-training model; S2: Clean, remove duplicates and segment the corpus to form a text sequence for pre-training; S3: During the pre-training process, a part of the words in the text sequence in the corpus is randomly masked, and the remaining word sequence is used as input to train the model to predict the masked words. The cross entropy loss function is used as the optimization target, and the model parameters are updated through the gradient descent algorithm; Suppose the input text sequence is X = {x1, x2, ..., x n }, some words are masked to get X masked , the probability distribution of the masked words predicted by the model is P model (X|X masked ), the distribution of the real masked words is P true (X), the cross entropy loss function is: S4: After pre-training is completed, the model is fine-tuned on a specific archival information resource understanding task to adapt to different query requirements and business scenarios; during the fine-tuning process, the model output layer is adjusted according to the task requirements, and the corresponding loss function and optimization algorithm are used for training; S5: Evaluate the performance of the pre-trained archival information resource understanding large model on the test set, including precision, recall, and F1 score metrics.

4. The AI-based intelligent sharing method for archival information resources according to claim 3 is characterized in that: In step 4, the keyword list is input into the pre-trained archival information resource understanding model, and the model performs semantic analysis on the keyword list through its internal neural network structure, including the recognition of the relationship between keywords, the labeling of semantic roles, and the integration of contextual information; based on the semantic analysis results of the keyword list by the model, an enhanced query text is generated, and the enhanced query text not only contains the key information in the user's original query request, but also is expanded and refined through the semantic understanding ability of the model.

5. The AI-based intelligent sharing method for archival information resources according to claim 4 is characterized in that: In step 5, according to the enhanced query text generated in step 4, a search is performed from pre-constructed data modules of multiple dimensions, the data modules include an archive system database module, a high-dimensional document vector database module and an archive business knowledge graph module, and the data modules of multiple dimensions respectively store different types of archive information resources; the candidate answers recalled from each data module are merged to form a set containing multiple candidate answers.

6. The AI-based intelligent sharing method for archive information resources according to claim 5 is characterized in that: In each data module, the corresponding retrieval algorithm and matching mechanism are used to recall candidate answers related to the query text: For the archive system database module, SQL query statements are used for retrieval; For the high-dimensional document vector database module, vector similarity calculation is used to recall similar documents; For the archive business knowledge graph module, graph traversal algorithm or semantic matching algorithm is used to find knowledge points related to the query.

7. The AI-based intelligent sharing method for archive information resources according to claim 3 is characterized in that: In S3, the formula for updating parameters of the gradient descent algorithm is: Among them, θ represents the model parameters, including all weights and biases that the model needs to learn; α represents the learning rate, which is used to control the step size of parameter update; It represents the gradient of the loss function J(θ) with respect to the parameter θ, which is a vector pointing to the direction in which the loss function grows fastest; := represents the assignment operation, that is, updating the value of the parameter θ.

8. The AI-based intelligent sharing method for archival information resources according to claim 1 is characterized in that: In step 6, the constructed question-answer context text is passed into the general large language model, and the natural language understanding ability of the general large language model is used to merge multiple candidate answers; During the fusion process, the general large language model analyzes the logical relationship, content similarity, and information complementarity between candidate answers, and generates a fused answer that integrates the advantages of each candidate answer.

9. An AI-based intelligent sharing system for archive information resources, characterized in that: The system comprises: A data module construction unit is used to pre-construct data modules of multiple dimensions according to business data types; A user request receiving unit, used to receive an archive information resource query request input by a user and obtain a first query request text; A natural language processing unit, configured to parse the first query request text using a natural language processing tool, wherein the parsing process includes word segmentation, part-of-speech tagging, dependency syntax analysis, and keyword extraction to form a structured keyword list; A semantic parsing unit, including a pre-trained archival information resource understanding model, is used to receive the structured keyword list, parse the semantics of the keyword list, and generate an enhanced query text as the second query request text; A candidate answer recall unit, used to recall the second query request text from the data modules of multiple dimensions and obtain multiple candidate answers; An answer fusion optimization unit is used to encapsulate the second query request text and multiple candidate answers into a question and answer context text according to a preset question and answer template, and pass the question and answer context text into the general large language model, and use the natural language understanding ability of the general large language model to fuse and optimize the multiple candidate answers; The output unit is used to obtain the output of the general large language model, take it as the optimal answer, and return the optimal answer to the current user.

Images