A method and system for intelligent compilation and research of archives based on intelligent agents

Abstract

This invention discloses an intelligent archival compilation and research method and system based on intelligent agents, belonging to the field of archival data processing technology. By extracting keywords from archival data, and then generating corresponding summaries based on the extracted keywords, intelligent agents are used to realize intelligent archival compilation and research based on the keywords and summaries. This can significantly improve compilation and research efficiency, reduce compilation and research costs, promote the development and utilization of archival resources, and make intelligent archival compilation and research more intelligent and personalized, bringing more possibilities to archival work.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of archive data processing, and particularly relates to an archive intelligent compiling and researching method and system based on an agent. BACKGROUND

[0002] Archive data is an important carrier of recording historical, cultural, social, economic and other information, including text, image, audio, video and electronic archives and other types. Its characteristics are authenticity, integrity, diversity, long-term and confidentiality. The sources are extensive, covering government agencies, enterprises and institutions, social organizations and individuals. Archive data has multiple values of history, law, scientific research, education and decision-making, and is an important basis for research, proof, reference, inheritance and decision-making. With the development of information technology, the digital and intelligent management of archive data has become a trend, and efficient processing and utilization of archive data have become the key to archive management. Archive compiling and researching is an important work of mining archive value and serving the society. The traditional archive compiling and researching method is low in efficiency and time-consuming, and is difficult to meet the needs of the information explosion era. SUMMARY

[0003] The application provides an archive intelligent compiling and researching method and system based on an agent, to solve the problem of low efficiency and long time consumption of the archive compiling and researching method in the prior art, which is difficult to meet the needs of the information explosion era.

[0004] In one aspect, the application provides an archive intelligent compiling and researching method based on an agent, comprising:

[0005] Retrieving archive data from a specified data source or a specified database, and preprocessing the archive data to obtain preprocessed archive data;

[0006] Performing a simplification process on the preprocessed archive data to obtain simplified archive data, and extracting an embedding vector in the simplified archive data;

[0007] After performing a keyword extraction operation on the embedding vectors in all archive data, constructing the extracted keywords into an archive keyword library;

[0008] For any one archive data, determining at least one target keyword corresponding to the archive data based on the archive keyword library;

[0009] Based on the at least one target keyword corresponding to the archive data, extracting a summary corresponding to the archive data using an agent to obtain an archive directory corresponding to the archive data;

[0010] According to the at least one target keyword corresponding to the archive data and the archive directory, intelligently compiling and researching the archive data to complete the archive intelligent compiling and researching based on the agent.

[0011] In a possible implementation, the archive data is retrieved from a specified data source or a specified database, and the archive data is preprocessed to obtain preprocessed archive data, including:

[0012] The archive data is retrieved from a specified data source or a specified database to obtain a plurality of archive data; wherein the archive data includes text data;

[0013] The archive data is de-duplicated to obtain de-duplicated archive data;

[0014] The de-duplicated archive data is format-unified to obtain format-unified archive data, and the format-unified archive data is taken as the preprocessed archive data.

[0015] In a possible implementation, the preprocessed archive data is simplified to obtain simplified archive data, including:

[0016] The image and table data in the preprocessed archive data are removed to obtain initial processed archive data;

[0017] The initial processed archive data is text-cleaned to obtain text-cleaned archive data;

[0018] The text-cleaned archive data is segmented and stop words are removed, and the segmented and stop words-removed archive data is taken as the simplified archive data.

[0019] In a possible implementation, the embedding vectors in the simplified archive data are extracted, including:

[0020] The simplified archive data is input into a BERT language model to process the simplified archive data through a word embedding function of the BERT language model, and embedding vectors in the simplified archive data are obtained.

[0021] In a possible implementation, after the embedding vectors in all archive data are subjected to keyword extraction, the extracted keywords are constructed into an archive keyword library, including:

[0022] The embedding vectors in the simplified archive data are subjected to dimension reduction processing by using a UMAP algorithm to obtain dimension-reduced embedding vectors;

[0023] Based on the dimension-reduced embedding vectors, clustering processing is performed by using an HDBSCAN algorithm to obtain a plurality of clustering topics;

[0024] For any one cluster topic, the c-TD-IDF value corresponding to each word in the cluster topic is obtained by using a c-TD-IDF algorithm;

[0025] For any one cluster topic, the word with the largest c-TD-IDF value is taken as the keyword corresponding to the cluster topic;

[0026] For any one cluster topic, the MMR value corresponding to each word except the keyword is calculated in turn, and the word with the largest MMR value is taken as a new keyword. This step is repeated until a keyword extraction end condition is met, and all keywords corresponding to each cluster topic are obtained.

[0027] According to all keywords corresponding to all cluster topics, an archive keyword library is constructed.

[0028] In a possible implementation, for any one archive data, at least one target keyword corresponding to the archive data is determined based on the archive keyword library, including:

[0029] For any one archive data to be processed, a weighted word graph corresponding to the archive data is obtained.

[0030] Based on the weighted word graph corresponding to the archive data and the archive keyword library, a score corresponding to each word in the archive data to be processed is obtained by using a TextRank algorithm.

[0031] According to the score corresponding to each word in the archive data to be processed, at least one target keyword corresponding to the archive data to be processed is determined.

[0032] In a possible implementation, based on the at least one target keyword corresponding to the archive data, an agent is used to extract a summary corresponding to the archive data, and an archive directory corresponding to the archive data is obtained, including:

[0033] The at least one target keyword corresponding to the archive data and the archive data are taken as inputs of a first agent to obtain a target keyword vector and an archive data text vector; the first agent includes an XLNet model.

[0034] The archive data text vector is processed by using a second agent to obtain a forward semantic feature text vector and a reverse semantic feature text vector; the second agent includes a BiGRU model.

[0035] The target keyword vector, the archive data text vector, the forward semantic feature text vector, and the reverse semantic feature text vector are processed by using a third agent to obtain an archive directory corresponding to the archive data; the third agent includes a decoder based on an attention mechanism.

[0036] In a possible implementation, before the intelligent agent is used, the method further includes training hyperparameters of the intelligent agent.

[0037] In a possible implementation, the intelligent compiling and researching of the archive data according to the at least one target keyword corresponding to the archive data and the archive directory includes:

[0038] According to the at least one target keyword corresponding to the archive data, the N keywords with the highest scores are taken as the title corresponding to the archive data.

[0039] The title corresponding to the archive data and the archive directory corresponding to the archive data are taken together as the personalized compiling and researching outline corresponding to the archive data, and the intelligent compiling and researching of the archive based on the intelligent agent is completed.

[0040] In another aspect, the present application provides an intelligent compiling and researching system of archives based on an intelligent agent, which includes a data preprocessing module, a vector extraction module, a keyword library construction module, a keyword extraction module, an abstract extraction module and an intelligent compiling and researching module.

[0041] The data preprocessing module is configured to retrieve archive data from a specified data source or a specified database, and preprocess the archive data to obtain the archive data after preprocessing.

[0042] The vector extraction module is configured to perform a simplification process on the archive data after preprocessing to obtain archive data after simplification, and extract embedded vectors in the archive data after simplification.

[0043] The keyword library construction module is configured to construct the extracted keywords as an archive keyword library after performing a keyword extraction operation on the embedded vectors in all archive data.

[0044] The keyword extraction module is configured to determine at least one target keyword corresponding to archive data based on the archive keyword library for any one archive data.

[0045] The abstract extraction module is configured to extract an abstract corresponding to the archive data based on the at least one target keyword corresponding to the archive data, and obtain an archive directory corresponding to the archive data.

[0046] The intelligent compiling and researching module is configured to intelligently compile and research the archive data according to the at least one target keyword corresponding to the archive data and the archive directory, and complete the intelligent compiling and researching of the archive based on the intelligent agent.

[0047] The application provides an intelligent agent-based archive intelligent compilation and research method and system, extracts keywords of archive data, generates a summary corresponding to the archive data based on the extracted keywords, and finally realizes archive intelligent compilation and research by using an intelligent agent according to the keywords and the summary of the archive data, so that the compilation and research efficiency is greatly improved, the compilation and research cost is reduced, the development and utilization of archive resources are promoted, the archive intelligent compilation and research is more intelligent and personalized, and more possibilities are brought to archive work. BRIEF DESCRIPTION OF DRAWINGS

[0048] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the application and, together with the description, further serve to explain the principles of the application.

[0049] Figure 1 A flowchart of an intelligent agent-based archive intelligent compilation and research method provided by an embodiment of the application.

[0050] Figure 2 A structural schematic diagram of an intelligent agent-based archive intelligent compilation and research system provided by an embodiment of the application.

[0051] Through the above drawings, the specific embodiments of the application have been shown, and will be described in more detail hereinafter. These drawings and the written description are not intended to limit the scope of the inventive concept in any way, but to illustrate the inventive concept to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0052] The exemplary embodiments will be described in detail herein with reference to the attached drawings. When the description below refers to accompanying drawings, unless otherwise noted, the same numbers in different drawings refer to the same or similar elements. The implementations described in the following exemplary embodiments are not meant to represent all implementations consistent with the application. Rather, they are merely examples of apparatus and methods consistent with some aspects of the application as detailed in the appended claims.

[0053] The embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0054] As shown in the drawings, an intelligent agent-based archive intelligent compilation and research method provided by an embodiment of the application comprises: Figure 1 S101, archive data is retrieved from a specified data source or a specified database, and the archive data is preprocessed to obtain archive data after preprocessing;

[0055]

[0056] ​Before the agent processing, the pre-processing of the archive data is a crucial step. As a historical record and information resource carrier, the archive data often contains rich knowledge and value, but its original form is not always suitable for direct input into the agent for analysis and mining. Therefore, pre-processing is an indispensable link to ensure data quality and improve model performance. The primary task of pre-processing is data cleaning. Because the archive data may come from different historical periods and different recording systems, its format, standard and quality are often uneven. First, data cleaning is performed, and the purpose is to remove impurities, redundant information and errors in the data, such as deleting duplicate entries, correcting spelling errors, standardizing formats, etc., to ensure the correctness and uniformity of the data.

[0057] S102, performing a simplification processing on the archive data after the pre-processing, obtaining archive data after the simplification processing, and extracting an embedding vector in the archive data after the simplification processing;

[0058] The archive data often has multiple data types, and some irrelevant data not only has no effect on the archive research, but also affects the implementation of the archive research scheme, therefore, some irrelevant data in the archive data can be simplified to improve the accuracy of the archive research.

[0059] In order to realize the archive research, the embedding vector in the archive data after the simplification processing needs to be extracted to realize the identification and extraction of the data.

[0060] S103, after performing a keyword extraction operation on the embedding vectors in all archive data, constructing the extracted keywords into an archive keyword library;

[0061] One of the focuses of the archive research in the embodiment of the application is to extract the archive catalog. Although the research on automatic text summary has gradually increased in recent years, some archive research data sets in specific fields are still relatively rare, therefore, the embodiment of the application can construct an archive keyword library based on a batch of archive data, and then perform archive research based on the archive keyword library, so as to improve the accuracy of the archive research.

[0062] S104, for any one archive data, determining at least one target keyword corresponding to the archive data based on the archive keyword library;

[0063] The target keyword can be selected according to a preset number, so as to select a preset number of target keywords; or the target keyword can be selected according to a preset condition, so as to select a target keyword meeting the condition.

[0064] S105. Based on at least one target keyword corresponding to the archival data, use an intelligent agent to extract the summary corresponding to the archival data to obtain the archival catalog corresponding to the archival data.

[0065] Traditional extractive summarization methods are widely used in the field of automatic text summarization, forming summaries by extracting key sentences or phrases from the original text. While this method shows certain advantages in preserving the meaning of the original text, building a robust model capable of understanding and capturing the deeper meaning of text remains a challenging problem. In particular, effectively extracting key information from long texts and generating concise, accurate, and fluent summaries is crucial to the quality of text summarization. With the rapid development of Natural Language Processing (NLP), generative summarization methods, especially agent-based summarization methods, have gradually become a research hotspot. Therefore, this invention employs an agent to extract summaries corresponding to archival data to improve the accuracy of summarization generation.

[0066] S106. Based on at least one target keyword corresponding to the archival data and the archival catalog, perform intelligent compilation and research on the archival data to complete intelligent compilation and research of archives based on intelligent agents.

[0067] For example, the compilation and research conditions can be preset as follows: use one or more target keywords as the titles corresponding to the archival data (such as setting titles and subtitles, that is, using the target keyword with the highest score as the title and the target keyword with the second highest score as the subtitle), and then use the order in which the archival data is stored as the corresponding serial number of the archival data, while supplementing it with an archival catalog to realize the compilation and research of archival data.

[0068] This invention provides an intelligent archival compilation and research method based on intelligent agents. By extracting keywords from archival data, generating summaries based on the extracted keywords, and then using intelligent agents to achieve intelligent archival compilation and research based on the keywords and summaries, the method can significantly improve compilation and research efficiency, reduce compilation and research costs, promote the development and utilization of archival resources, and make intelligent archival compilation and research more intelligent and personalized, bringing more possibilities to archival work.

[0069] In one possible implementation, retrieving archival data from a specified data source or a specified database, and preprocessing the archival data to obtain preprocessed archival data includes:

[0070] Retrieve archive data from a specified data source or database to obtain multiple archive data; the archive data includes text data;

[0071] The archive data is deduplicated to obtain the deduplicated archive data. For example, a deduplication threshold can be set in advance. When the text similarity between two archive data exceeds this threshold, it indicates that the two archive data are duplicates, and one of them or the archive data that is older can be randomly removed.

[0072] The deduplicated archival data is then formatted to obtain formatted archival data, which is then used as the preprocessed archival data.

[0073] Archival data from different sources may exist in different formats. To standardize processing, archival data can be converted into a unified format for easier handling.

[0074] In one possible implementation, the preprocessed archival data is simplified to obtain simplified archival data, including:

[0075] The images and tables in the preprocessed archival data are removed to obtain the initially processed archival data. Since the images and tables in the archival data are not very useful for archival compilation and research, this embodiment of the invention can remove these useless data, thereby improving the accuracy of archival compilation and research.

[0076] Text cleaning is performed on the initially processed archival data to obtain the cleaned archival data. For example, special symbols, numerical values, and punctuation marks can be removed from the initially processed archival data. These elements often interfere with archival compilation and research, so they can be removed to ensure the accuracy of the final result.

[0077] The document data after text cleaning is segmented and stop words are removed. The segmented and stop word removed document data is then used as the simplified document data.

[0078] In one possible implementation, the embedding vector is extracted from the streamlined archival data, including:

[0079] The simplified archival data is input into the BERT language model to process the simplified archival data through the word embedding function of the BERT language model and obtain the embedding vector in the simplified archival data.

[0080] The BERT model, pre-trained on a large amount of text, is able to understand complex patterns and contextual relationships in language. Sentences in a document are transformed into a fixed-length vector. This vector represents the semantic content of the document in a multi-dimensional space. When the archival dataset is input into the model, the model's output layer generates the embedding vector of this text.

[0081] In one possible implementation, after extracting keywords from the embedding vectors in all archival data, the extracted keywords are used to construct an archival keyword library, including:

[0082] The UMAP algorithm is used to reduce the dimensionality of the embedding vectors in the simplified archival data, resulting in the dimensionality-reduced embedding vectors.

[0083] UMAP (Uniform Manifold Approximation and Projection) is a non-linear dimensionality reduction algorithm for visualizing high-dimensional data. It aims to map high-dimensional data to a low-dimensional space while preserving the local and global structure of the data. The UMAP algorithm has been widely applied in many fields, such as bioinformatics, image processing, and natural language processing.

[0084] The main features of the UMAP algorithm include: (1) Nonlinear dimensionality reduction: UMAP can capture the nonlinear structure in high-dimensional data and map it to a low-dimensional space. (2) Preservation of local and global structure: UMAP considers both the local and global structure of the data during the dimensionality reduction process, so that the low-dimensional representation can better reflect the characteristics of the high-dimensional data. (3) Efficiency: The UMAP algorithm is relatively efficient in computation and can handle large-scale datasets. (4) Scalability: UMAP can adapt to datasets of different sizes and complexities. (5) Fewer parameters: The UMAP algorithm has relatively few parameters, which are easy to adjust and optimize.

[0085] Based on the dimensionality-reduced embedding vectors, the HDBSCAN algorithm is used for clustering to obtain multiple clustering topics;

[0086] HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) is a density-based spatial clustering algorithm that can identify clusters of arbitrary shapes and handle noisy data. HDBSCAN does not require pre-specifying the number of clusters, making it very useful in many practical applications.

[0087] Using UMAP text embedding dimension reduction and the HDBSCAN algorithm for document embedding clustering, similar documents are grouped into the same category, making data processing more efficient.

[0088] For any given cluster topic, the c-TD-IDF algorithm is used to obtain the c-TD-IDF values ​​corresponding to the words in each cluster topic;

[0089] c-TD-IDF (Cluster-based Term Frequency-Inverse Document Frequency) is an improved TF-IDF algorithm for text feature representation. Traditional TF-IDF algorithms evaluate the importance of a word in a document set or corpus using term frequency (TF) and inverse document frequency (IDF). c-TD-IDF, however, introduces clustering information to better capture the semantic features of the text.

[0090] For any given cluster topic, the word with the largest c-TD-IDF value is taken as the keyword corresponding to that cluster topic;

[0091] For any cluster topic, calculate the MMR value of each word except the keyword in turn, and select the word with the largest MMR value as the new keyword. Repeat this step until the keyword extraction end condition is met (such as when the number of keywords reaches the requirement or the MMR value does not meet the threshold requirement), and obtain all the keywords corresponding to each cluster topic.

[0092] The Maximum Marginal Relevance (MMR) value typically refers to the maximum marginal relevance. MMR is a technique used in information retrieval and document summarization that aims to balance relevance and diversity. It attempts to return the most relevant results while avoiding redundancy. Therefore, the archival keyword library constructed in this embodiment of the invention can make keyword extraction more accurate, while considering the relevance of keywords to the topic, ensuring that keywords accurately describe the topic. It also considers the differences between newly selected keywords and previously selected keywords, making the keyword list more diverse and avoiding duplication. The archival keyword library is constructed based on all keywords corresponding to all clustered topics.

[0093] In one possible implementation, for any given set of archival data, based on the archival keyword database, at least one target keyword corresponding to the archival data is determined, including:

[0094] For any given set of archive data to be processed, obtain the weighted vocabulary graph corresponding to the archive data;

[0095] Based on the weighted vocabulary graph corresponding to the archival data and the archival keyword database, the TextRank algorithm is used to obtain the score corresponding to each word in the archival data to be processed.

[0096] The keywords to be extracted from the archival data are matched with the archival keyword database. Words that appear in both databases are assigned a higher initial weight, and the weights of the edges in the graph are adjusted accordingly. This reflects that the co-occurrence relationship between keywords is more important than other words in agricultural literature. Edges where two keywords co-occur can be assigned two levels of weight, improving the accuracy of keyword importance assessment.

[0097] TextRank is a graph-based ranking algorithm for text processing that ranks words or sentences in a text based on the co-occurrence relationships between local words. Inspired by PageRank, TextRank can be used for tasks such as keyword extraction and text summarization.

[0098] Based on the score corresponding to each word in the archive data to be processed, at least one target keyword corresponding to the archive data to be processed is determined.

[0099] For example, you can select the N words with the highest scores as at least one target keyword for the archive data to be processed. However, it is worth noting that other conditions can also be set to select target keywords.

[0100] Optionally, by extracting target keywords, embodiments of the present invention can automatically generate a well-structured and logically clear compilation outline based on the relevance of archival data content, such as including chapter titles and / or subheadings.

[0101] Optionally, users can set conditions to extract information on individuals and time to enhance the personalization of archival research and ultimately achieve better results.

[0102] In one possible implementation, based on at least one target keyword corresponding to the archival data, an intelligent agent is used to extract a summary corresponding to the archival data to obtain an archival catalog corresponding to the archival data, including:

[0103] The first intelligent agent takes at least one target keyword corresponding to the archive data and the archive data as input to obtain the target keyword vector and the archive data text vector; wherein, the first intelligent agent includes the XLNet model;

[0104] The XLNet model is a general-purpose pre-trained language model based on the Transformer, jointly developed by researchers from Google and Carnegie Mellon University (CMU). During the pre-training phase, the XLNet model employs a hybrid approach of autoregression and autoencoders, aiming to combine the advantages of autoencoder models like BERT with those of traditional autoregressive models. This embodiment of the invention uses the XLNet model to extract text vectors, providing a foundation for subsequent summary extraction.

[0105] A second intelligent agent is used to process the text vectors of the archive data to obtain positive semantic feature text vectors and negative semantic feature text vectors; wherein, the second intelligent agent includes a BiGRU model;

[0106] The archival data text vectors are fed into the BiGRU model, which consists of two GRU units: one for forward extraction of semantic features from the archival data text vectors and the other for backward extraction of semantic features from the archival data text vectors. The model is then trained using multiple GRU hidden units to obtain the contextual features of the two archival data text vectors. The forward semantic feature text vector and the backward semantic feature text vector are the contextual features of the archival data text vectors after being trained by BiGRU forward and backward extraction, respectively.

[0107] Optionally, the second agent may also include a multi-layer attention mechanism model, which introduces a multi-layer attention mechanism after BiGRU to enhance the model's attention to important parts of the text. Based on the previous keyword extraction and the semantic features of the text vectors of the archive data, the attention mechanism and the coverage mechanism are combined to generate the archive catalog.

[0108] A third intelligent agent is used to process the target keyword vector, the archive data text vector, the positive semantic feature text vector, and the negative semantic feature text vector to obtain the archive catalog corresponding to the archive data; the third intelligent agent includes a decoder based on an attention mechanism.

[0109] Optionally, the attention-based decoder may include a sequentially connected self-attention mechanism, a first normalization layer, an attention layer, a second normalization layer, a feed-forward layer, a third normalization layer, a linear layer, and a softmax layer. The input and output of the self-attention mechanism are simultaneously used as the input of the first normalization layer, the input and output of the attention layer are simultaneously used as the input of the second normalization layer, and the input and output of the feed-forward layer are simultaneously used as the input of the third normalization layer.

[0110] This invention introduces an attention-based decoder to further optimize the processed text vectors. This attention mechanism enhances the model's ability to focus on key information, more accurately capturing the core content of the text while avoiding redundant or repetitive summary information. To integrate and optimize all the above information, residual connections and layer normalization are used to fuse and standardize various feature vectors. Residual connections help solve the gradient vanishing problem in deep networks, while layer normalization ensures smoother data flow within the network, promoting the stability and efficiency of model training. Finally, the decoder converts the fused feature vectors into predicted word probability distributions through linear layers, and then uses softmax layers to transform these probability distributions into the final text summary output. This process ensures that the generated summary not only concisely and accurately reflects the core content of the original text but also flows smoothly and naturally, meeting the target requirements for summary generation.

[0111] In one possible implementation, the process further includes training the hyperparameters of the agent before its use.

[0112] The training process may include:

[0113] A1. Construct a main task node and multiple sub-task nodes for the intelligent agent, and allocate private memory and a public memory for each task node; each node is deployed on a different core or server;

[0114] A2. Initialize the iteration counter t =1, initialize the scale matrix =I; where I represents the unit-scale matrix;

[0115] A3. Obtain the error function value corresponding to the agent. and the error function value It is placed in a common memory for use by various task nodes; among them, the error function value Obtained through an agent on any task node. Indicates the first t Hyperparameters of the agent during training (such as connection weights between network layers);

[0116] A4. Obtain the gradient matrix of the agent. and the scale matrix and gradient matrix Placed in common memory for use by various task nodes; among them, the gradient matrix Each column in the table represents the gradient of a hyperparameter in the agent; where, Represents the gradient operator, forM Square array M This represents the total number of hyperparameter dimensions of the agent;

[0117] A5. Each task node runs in parallel, and the scaling matrix is ​​retrieved from shared memory by each task node. and gradient matrix Determine the first m Search direction of the agent on each task node scale matrix The Middle m Row data and gradient matrix The Middle m The column data is multiplied and accumulated; where m = 1, 2, ..., M, and the search direction is... Stored in the first m In the private memory corresponding to each task node;

[0118] A6. Retrieve the search direction from the private memory of each task node. According to the search direction The golden section method was used to determine the first [section] in the agent model at this task node. m The update step size corresponding to each hyperparameter And will update the step size Stored in the first m In the private memory corresponding to each task node;

[0119] Optional, depending on the search direction The golden section method was used to determine the first [section] in the agent model at this task node. m The update step size corresponding to each hyperparameter ,include:

[0120] Use search direction Construct a search matrix with only one row. D The m-th element in the search matrix represents the search direction. All other elements are 0, and the total number of elements in the search matrix is ​​the same as the total number of hyperparameters in the agent model;

[0121] Construct a step size matrix with only one column. The m-th element in the step size matrix is ​​the update step size to be determined. All other elements are 0, and the total number of elements in the step matrix is ​​the same as the total number of hyperparameters in the agent model;

[0122] According to the search matrix D and search matrix D Constructing solution conditions The solution conditions are then solved to determine the update step size. ;

[0123] A7. Retrieve the update step size from the private memory of each task node. And based on the update step size For the agent model at the task node, the first m Update the hyperparameters according to the first hyperparameter; m The agent model on the nth task node m The updated hyperparameters are obtained. m One gradient component;

[0124] Optional, based on update step size For the agent model at the task node, the first m The hyperparameters are updated as follows:

[0125]

[0126] in, In the intelligent agent model, the first... m One hyperparameter, Indicates the updated ;

[0127] A8. Each task node runs in parallel, according to the... m The i-th gradient component is redefined. m Gradient matrix corresponding to each task node And based on the gradient matrix For the m The scaling matrix corresponding to each task node Update;

[0128] Optional, according to the first m The i-th gradient component is redefined. m Gradient matrix corresponding to each task node And based on the gradient matrix For the m The scaling matrix corresponding to each task node Updates will be made, including:

[0129] With gradient matrix Based on this, the gradient matrix is ​​updated using the m-th gradient component. The Middle m The element is obtained. m Gradient matrix corresponding to each task node ;

[0130] According to the first m Gradient matrix corresponding to each task node , for the m The scaling matrix corresponding to each task node Updated to:

[0131]

[0132] in, Indicates the updated , W Represents the weight difference matrix. W = , Indicates based on update step size The updated weight vector of the agent model. This represents the weight vector of the agent model before the update. T Z represents the transpose. Z represents the gradient difference matrix;

[0133] A9. Output the scaling matrix through each task node. The Middle m Column elements and the corresponding first agent model m One hyperparameter is sent to the main task node;

[0134] A10. Through the main task node m The updated hyperparameter vector is constructed from the hyperparameters. It is then determined whether the error function value corresponding to the updated hyperparameter vector is less than the preset threshold. If so, the training end requirement is met and training is completed; otherwise, proceed to step A11.

[0135] A11. Determine the iteration counter t If the count value is greater than the preset maximum number of training iterations, then the training ends and the training is complete; otherwise, the iteration counter is reset. t Increment the count value by one, and output the scale matrix for each task node. The Middle m The column elements form the scale matrix The updated hyperparameter vector is used as the hyperparameters of the agent model on all task nodes, and the process returns to step A3.

[0136] The embodiments of the present invention provide a training algorithm that enables multiple parties to collaborate on updates, effectively improving the algorithm execution speed. At the same time, it enables devices or processors with weak computing power to perform collaborative training, thereby improving hyperparameter optimization capabilities.

[0137] In one possible implementation, the archival data is intelligently compiled based on at least one target keyword corresponding to the archival data and the archival catalog, including:

[0138] Based on at least one target keyword corresponding to the archival data, the N keywords with the highest scores are used as the titles corresponding to the archival data.

[0139] The titles and catalogs corresponding to the archival data are used together as a personalized compilation outline for the archival data, thus completing intelligent archival compilation based on intelligent agents.

[0140] like Figure 2 As shown, the present invention provides an intelligent archive compilation and research system based on intelligent agents, including: a data preprocessing module 201, a vector extraction module 202, a keyword database construction module 203, a keyword extraction module 204, an abstract extraction module 205, and an intelligent compilation and research module 206;

[0141] The data preprocessing module 201 is used to retrieve archive data from a specified data source or a specified database, and preprocess the archive data to obtain preprocessed archive data.

[0142] The vector extraction module 202 is used to simplify the preprocessed archive data to obtain simplified archive data, and extract the embedded vectors from the simplified archive data.

[0143] The keyword library construction module 203 is used to extract keywords from the embedded vectors in all archive data and then construct the extracted keywords into an archive keyword library.

[0144] The keyword extraction module 204 is used to determine at least one target keyword corresponding to any archival data based on the archival keyword database.

[0145] The abstract extraction module 205 is used to extract the abstract corresponding to the archive data based on at least one target keyword corresponding to the archive data, and to obtain the archive catalog corresponding to the archive data.

[0146] The intelligent compilation module 206 is used to intelligently compile and research archival data based on at least one target keyword corresponding to the archival data and the archival catalog, thereby completing intelligent archival compilation and research based on intelligent agents.

[0147] The intelligent archive compilation system based on intelligent agents provided in this embodiment of the invention can execute the above-mentioned method and technical solution. Its principle and beneficial effects are similar, and will not be repeated here.

[0148] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] Those skilled in the art will understand that all or part of the steps in the above facts and methods can be implemented by a program instructing related hardware. The program or the program described therein can be stored in a computer-readable storage medium. When the program is executed, it includes the following steps: at this time, the corresponding method steps are introduced. The storage medium can be ROM / RAM, magnetic disk, optical disk, etc.

[0153] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An agent-based archive intelligent compiling and researching method, characterized in that, The application relates to an intelligent archive processing method and device. The method comprises the following steps: Retrieving archive data from a specified data source or a specified database, and preprocessing the archive data to obtain preprocessed archive data; Performing simplification processing on the preprocessed archive data to obtain simplified archive data, and extracting embedding vectors in the simplified archive data; After extracting keywords from all embedding vectors in the archive data, the extracted keywords are constructed into an archive keyword library; For any one piece of archive data, at least one target keyword corresponding to the archive data is determined based on the archive keyword library; Based on the at least one target keyword corresponding to the archive data, an agent is used to extract a summary corresponding to the archive data, and an archive directory corresponding to the archive data is obtained; According to the at least one target keyword corresponding to the archive data and the archive directory, intelligent compilation and research are performed on the archive data to complete archive intelligent compilation and research based on the agent; Before the agent is used, the hyperparameters of the agent are trained; The training process comprises the following steps: A2, initialize iteration counter t = 1, initialize scale matrix = I; where I denotes an identity scale matrix A3, obtaining the error function value corresponding to the intelligent agent and putting the error function value into the public memory for calling by each task node; wherein the error function value is obtained by the intelligent agent on any task node, wherein the intelligent agent is trained by the first training data set t and the hyperparameters of the intelligent agent during training A4、obtaining a gradient matrix of the intelligent agent and putting the scale matrix and the gradient matrix into a common memory for calling by each task node; wherein each column in the gradient matrix is the gradient of one hyperparameter in the intelligent agent; wherein represents a gradient operator, is an M order square matrix, M represents the total number of hyperparameter dimensions of the intelligent agent; A5、each task node runs in parallel, and each task node calls a scale matrix from a public memory and a gradient matrix to determine a search direction of an agent on the m th task node The product accumulation data of the data in the th row of the scale matrix m and the data in the th column of the gradient matrix m ; wherein m = 1, 2, …, M, and the search direction is stored in the private memory corresponding to the m th task node A6. calling the search direction from the private memory of each task node , according to the search direction and using the golden section method to determine the update step size corresponding to the first m hyperparameter in the agent model on the task node , and storing the update step size in the private memory corresponding to the first m task node; According to the search direction And using the golden section method to determine the update step of the agent model on the task node m The first super parameter corresponds to , comprising: Adopting search direction Constructing a search matrix with only one row D , the mth element of which is the search direction , and other elements are 0. The total number of elements of the search matrix is the same as the total number of hyperparameters of the agent model. Construct a step matrix with only one column , the mth element of which is the to-be-determined update step , and other elements are 0. The total number of elements of the step matrix is the same as the total number of hyperparameters of the agent model; According to the search matrix D And the search matrix D Constructing a solution condition And solving the solution condition to determine the update step ; A7. Retrieve the update step size from the private memory of each task node. And based on the update step size For the agent model at the task node, the first m Update the hyperparameters according to the first hyperparameter; m The agent model on the nth task node m The updated hyperparameters are obtained. m One gradient component; According to the update step For the agent model at the task node, the first m The hyperparameters are updated as follows: wherein, represents the i-th hyperparameter in the agent model, m represents the updated ;​ A8、Each task node runs in parallel, according to the first m Gradient component, re-determine the first m Task node corresponding gradient matrix , According to the gradient matrix The first m Task node corresponding scale matrix Update; According to the m The gradient component is redefined. m Gradient matrix corresponding to each task node And based on the gradient matrix For the m The scaling matrix corresponding to each task node Updates will be made, including: Based on the gradient matrix , the mth gradient component is used to update the gradient matrix , the mth gradient component is used to update the gradient matrix m , the mth gradient component is used to update the gradient matrix m , the mth gradient component is used to update the gradient matrix ; According to the first m Gradient matrix corresponding to each task node , for the m The scaling matrix corresponding to each task node Updated to: wherein, denotes the updated , W denotes a weight delta matrix, W = , denotes the weight vector of the agent model after the update according to the update step size , denotes the weight vector of the agent model before the update, T denotes a transpose, Z= , Z denotes a gradient delta matrix; A9. outputting the scale matrix by each task node In the present application m Column elements and the corresponding first m Hyperparameters of the agent model to the main task node; A10、 through the main task node m An updated hyperparameter vector is formed by the hyperparameters, and it is determined whether an error function value corresponding to the updated hyperparameter vector is less than a preset threshold. If yes, the training end requirement is met, the training is completed, otherwise, step A11 is entered. A11. Determine the iteration counter t If the count value is greater than the preset maximum number of training iterations, then the training ends and the training is complete; otherwise, the iteration counter is reset. t Increment the count value by one, and output the scale matrix for each task node. The Middle m The column elements form the scale matrix The updated hyperparameter vector is used as the hyperparameters of the agent model on all task nodes, and the process returns to step A3. 2.The agent-based archive intelligence research method according to claim 1, wherein, A1, a main task node and multiple sub-task nodes are constructed for the agent, and a private memory is divided for each task node and a public memory is divided; wherein each node is deployed on a different core or server; The preprocessed archive data is simplified to obtain simplified archive data, which comprises the following steps: The images and table data in the preprocessed archive data are removed to obtain initial processed archive data; Text cleaning processing is performed on the initial processed archive data to obtain text cleaning processed archive data; 3.The agent-based archive intelligence research method according to claim 1, wherein, The text cleaning processed archive data is subjected to word segmentation and stop word removal processing, and the archive data after word segmentation and stop word removal processing is taken as the simplified archive data. The embedding vectors in the simplified archive data are extracted, which comprises the following steps: 4.The agent-based archive intelligence summarization method of claim 3, wherein, The simplified archive data is input into a BERT language model to process the simplified archive data through the word embedding function of the BERT language model, and the embedding vectors in the simplified archive data are obtained. After extracting keywords from all embedding vectors in the archive data, the extracted keywords are constructed into an archive keyword library, which comprises the following steps: The embedding vectors in the simplified archive data are subjected to dimension reduction processing by using a UMAP algorithm to obtain dimension reduction processed embedding vectors; Based on the dimension reduction processed embedding vectors, clustering processing is performed by using an HDBSCAN algorithm to obtain multiple clustering topics; For any one clustering topic, a c-TD-IDF value corresponding to each word in the clustering topic is obtained by using a c-TD-IDF algorithm; For any one clustering topic, the word with the largest c-TD-IDF value is taken as the keyword corresponding to the clustering topic; For any one clustering topic, the MMR values of other words except the keyword are calculated in sequence, and the word with the largest MMR value is taken as a new keyword. This step is repeated until a keyword extraction end condition is met, and all keywords corresponding to each clustering topic are obtained. According to all keywords corresponding to all clustering topics, an archive keyword library is constructed.

5. The agent-based archive intelligence curation method of claim 4, wherein, For any one archive data, at least one target keyword corresponding to the archive data is determined based on the archive keyword library, including: For any one archive data, a weighted vocabulary graph corresponding to the archive data is obtained; Based on the weighted vocabulary graph corresponding to the archive data and the archive keyword library, a score corresponding to each vocabulary in the archive data is obtained by using a TextRank algorithm; According to the score corresponding to each vocabulary in the archive data, at least one target keyword corresponding to the archive data is determined.

6. The agent-based archive intelligence curation method of claim 5, wherein, Based on the at least one target keyword corresponding to the archive data, an agent is used to extract a summary corresponding to the archive data, to obtain an archive directory corresponding to the archive data, including: The at least one target keyword corresponding to the archive data and the archive data are taken as inputs of a first agent to obtain a target keyword vector and an archive data text vector; wherein the first agent includes an XLNet model; A second agent is used to process the archive data text vector to obtain a forward semantic feature text vector and a reverse semantic feature text vector; wherein the second agent includes a BiGRU model; A third agent is used to process the target keyword vector, the archive data text vector, the forward semantic feature text vector and the reverse semantic feature text vector to obtain the archive directory corresponding to the archive data; wherein the third agent includes a decoder based on an attention mechanism.

7. The agent-based archive intelligence curation method of claim 6, wherein, Before the agent is used, the hyperparameters of the agent are trained. 8.The agent-based archive intelligence summarization method of claim 6, wherein, According to the at least one target keyword corresponding to the archive data and the archive directory, the archive data is intelligently compiled and researched, including: According to the at least one target keyword corresponding to the archive data, the top N keywords with the highest scores are taken as a title corresponding to the archive data; The title corresponding to the archive data and the archive directory corresponding to the archive data are taken together as a personalized compilation and research outline corresponding to the archive data to complete the archive intelligent compilation and research based on the agent.

9. An agent-based archive intelligent compiling and researching system, which is capable of executing the agent-based archive intelligent compiling and researching method according to any one of claims 1-8, characterized in that, including: a data preprocessing module, a vector extraction module, a keyword library construction module, a keyword extraction module, a summary extraction module and an intelligent compilation and research module; The data preprocessing module is configured to retrieve archive data from a specified data source or a specified database, and to preprocess the archive data to obtain preprocessed archive data; The vector extraction module is configured to perform simplification processing on the preprocessed archive data to obtain simplified archive data, and to extract embedded vectors in the simplified archive data; The keyword library construction module is configured to construct extracted keywords as an archive keyword library after performing a keyword extraction operation on the embedded vectors in all archive data; The keyword extraction module is configured to determine at least one target keyword corresponding to archive data based on the archive keyword library for any one archive data; The summary extraction module is configured to extract a summary corresponding to the archive data by using an agent based on the at least one target keyword corresponding to the archive data, to obtain an archive directory corresponding to the archive data; The intelligent compiling module is configured to compile and research the archive data according to at least one target keyword corresponding to the archive data and an archive directory, and complete the archive intelligent compiling based on the intelligent agent. The intelligent compiling module is configured to compile and research the archive data according to at least one target keyword corresponding to the archive data and an archive directory, and complete the archive intelligent compiling based on the intelligent agent.

Images