A method, device, equipment and readable storage medium for processing an archival task

Abstract

The application relates to the technical field of archive management, and particularly provides a method, device and equipment for processing an archive task and a readable storage medium, the method comprising the following steps: performing preliminary processing on an archive to be processed by using a preset archive processing large model to obtain global information of the archive to be processed, wherein the preliminary processing comprises at least one of global understanding, context analysis, key information extraction, semantic relationship extraction and cross-archive information integration; performing final processing on the global information by using a preset archive processing small model to obtain a processing result, wherein the final processing comprises at least one of keyword extraction, archive classification and archive abstract generation; and processing a target archive task by using the processing result, wherein the target archive task comprises archive management, information inquiry or decision support of the archive to be processed. The method can efficiently and accurately process the archive task.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of archive management, in particular to a method and device for processing archive tasks, and a readable storage medium. BACKGROUND

[0002] In the field of archive management and information processing, with the rapid development of information technology and the sharp increase in the amount of archive data, how to efficiently and accurately process and analyze these archive data has become a problem to be solved. Traditional archive processing methods often rely on manual operation or simple automated tools, which are not sufficient when dealing with large-scale, complex semantic relationship archive tasks. Manual processing is time-consuming and labor-intensive, and is easily affected by personal experience and subjective judgment; simple automated tools can only complete basic classification, retrieval and other tasks, and are difficult to cope with complex tasks involving deep semantic understanding and cross-archive information integration.

[0003] The rapid development of artificial intelligence technology provides a new way for archive processing. In particular, the continuous progress of natural language processing (NLP) and machine learning (ML) technology enables computers to more deeply understand and analyze text data. However, a single model still faces many challenges in processing complex archive tasks, and the analysis is not accurate enough.

[0004] Therefore, how to efficiently and accurately process archive tasks is a technical problem to be solved. SUMMARY

[0005] The purpose of the embodiments of the present application is to provide a method for processing archive tasks, which can efficiently and accurately process archive tasks through the technical solutions of the embodiments of the present application.

[0006] In a first aspect, the embodiments of the present application provide a method for processing archive tasks, comprising: preliminarily processing a to-be-processed archive through a preset archive processing large model to obtain global information of the to-be-processed archive, wherein the preliminary processing includes at least one of global understanding, context analysis, extracting key information, extracting semantic relationship and cross-archive information integration, and the archive processing large model is trained by combining unsupervised learning and supervised learning; finally processing the global information through a preset archive processing small model to obtain a processing result, wherein the final processing includes at least one of extracting keywords, archive classification and generating an archive abstract, and the archive processing small model is trained by supervised learning; and processing a target archive task through the processing result, wherein the target archive task includes archive management, information query or decision support of the to-be-processed archive.

[0007] In the above embodiment, the pre-constructed archive processing large model is used to understand and process the global archive information, and the archive processing small model is used to optimize the global information for specific archive tasks. The combination of the global understanding of the large model and the specific processing of the small model can process complex archive tasks in stages: the large model provides global information and context understanding, and the small model efficiently executes specific tasks and outputs results. This method can efficiently and accurately process archive tasks.

[0008] In some embodiments, the global information includes at least one of the global definition, context relationship, key text, and semantic relationship of the to-be-processed archive; and the processing result includes at least one of the keyword, archive type, and archive summary of the to-be-processed archive.

[0009] In the above embodiment, the acquisition of global information and processing results can provide data basis for subsequent target archive task processing.

[0010] In some embodiments, before the to-be-processed archive is preliminarily processed by the preset archive processing large model to obtain the global information of the to-be-processed archive, the method further includes: constructing an archive set of multiple types of historical archives; preprocessing the archive set to obtain training samples, wherein the preprocessing includes at least one of word segmentation, stop word removal, and stem extraction; using a mask method in unsupervised learning to train a basic neural network architecture using the training samples to obtain an initial archive processing large model; labeling the archive set to obtain a labeled archive set; and using a supervised learning method to adjust the parameters of the initial archive processing large model using the labeled archive set to obtain the archive processing large model.

[0011] In the above embodiment, after constructing the archive set, training the initial archive processing large model using the unsupervised mask method, and fine-tuning the initial archive processing large model using the supervised learning method to obtain the archive processing large model, the dual-supervised learning method can quickly and accurately identify the to-be-processed archive and obtain the global information.

[0012] In some embodiments, before the to-be-processed archive is preliminarily processed by the preset archive processing large model to obtain the global information of the to-be-processed archive, the method further includes: processing the archives in the archive set by the archive processing large model to obtain multiple global information of the archive set; labeling the archive type, summary, and keyword in the multiple global information to obtain a second training sample; and using a supervised learning method to train a basic convolutional neural network using the second training sample to obtain an archive processing small model.

[0013] In the above embodiment, the global information marked can be used as a training sample for supervised learning of the case processing small model, and the obtained case processing small model can process the global information and accurately obtain a processing result.

[0014] In some embodiments, the global information is finally processed by the preset archive processing small model to obtain a processing result, including: obtaining a correlation score between the archive processing large model and the archive processing small model; performing weighted summation on information in the global information according to the correlation score through an attention mechanism or a gating mechanism of the archive processing small model to obtain partial global information; and finally processing the partial global information by the archive processing small model to obtain the processing result.

[0015] In the above embodiment, the information in the global information can be filtered through the attention mechanism or the gating mechanism of the archive processing small model, and the most important partial global information is selected as the data basis for final processing, so that a more accurate processing result can be obtained.

[0016] In some embodiments, the preliminary processing is cross-archive information integration, and the global information of the to-be-processed archive is obtained by preliminarily processing the to-be-processed archive by the preset archive processing large model, including: constructing an initial relationship graph between the to-be-processed archive and the correlative archive; and optimizing the information in the initial relationship graph by the archive processing large model to obtain a relationship graph, and the global information includes the relationship graph.

[0017] In the above embodiment, when processing cross-archive information integration, the relationship graph between the to-be-processed archive and the correlative archive can be used as the global information in priority, which provides a data basis for subsequent archive processing small model processing.

[0018] In some embodiments, the target archive task is processed by using the processing result, including: constructing a task list of the to-be-processed archive by using the processing result, wherein the task list includes archive management, information query and decision support; and matching a target archive task in the task list and automatically executing the target archive task.

[0019] In the above embodiment, the task list can be accurately pre-constructed, the target archive task can be quickly obtained by task matching, and the target archive task can be automatically executed.

[0020] In a second aspect, the embodiments of the present application provide a device for processing an archive task, including:

[0021] The first processing module is configured to perform preliminary processing on the to-be-processed archives by using a preset archive processing large model to obtain global information of the to-be-processed archives, wherein the preliminary processing includes at least one of global understanding, context analysis, key information extraction, semantic relationship extraction, and cross-archive information integration, and the archive processing large model is trained by combining unsupervised learning and supervised learning.

[0022] The second processing module is configured to perform final processing on the global information by using a preset archive processing small model to obtain a processing result, wherein the final processing includes at least one of keyword extraction, archive classification, and archive abstract generation, and the archive processing small model is trained by supervised learning.

[0023] The third processing module is configured to process a target archive task by using the processing result, wherein the target archive task includes archive management, information query, or decision support of the to-be-processed archives.

[0024] Optionally, the global information includes:

[0025] at least one of global definition, context connection, key text, and semantic relationship of the to-be-processed archives;

[0026] The processing result includes at least one of keywords, archive type, and archive abstract of the to-be-processed archives.

[0027] Optionally, the apparatus further includes:

[0028] The first training module is configured to construct an archive set of multiple types of historical archives before the first processing module performs preliminary processing on the to-be-processed archives by using the preset archive processing large model to obtain the global information of the to-be-processed archives.

[0029] The archive set is preprocessed to obtain training samples, wherein the preprocessing includes at least one of word segmentation, stop word removal, and stem extraction.

[0030] The initial archive processing large model is obtained by training a basic neural network architecture by using the training samples in a mask manner of unsupervised learning.

[0031] The archive set is labeled to obtain a labeled archive set.

[0032] The parameters of the initial archive processing large model are adjusted by using the labeled archive set in a manner of supervised learning to obtain the archive processing large model.

[0033] Optionally, the apparatus further includes:

[0034] The second training module is configured to process the archives in the archive set by using the archive processing large model to obtain a plurality of global information of the archive set before the second processing module processes the to-be-processed archive by using the preset archive processing large model to obtain the global information of the to-be-processed archive.

[0035] The archive type, the abstract and the keyword in the plurality of global information are labeled to obtain a second training sample.

[0036] The second training sample is used to train the basic convolutional neural network in a supervised learning manner to obtain the archive processing small model.

[0037] Optionally, the second processing module is specifically configured to:

[0038] The correlation score between the archive processing large model and the archive processing small model is obtained.

[0039] The information in the global information is weighted and summed according to the correlation score by using the attention mechanism or the gating mechanism of the archive processing small model to obtain partial global information.

[0040] The partial global information is finally processed by using the archive processing small model to obtain a processing result.

[0041] Optionally, the preliminary processing is cross-archive information integration, and the first processing module is specifically configured to:

[0042] An initial relationship graph between the to-be-processed archive and the correlative archive is constructed.

[0043] The information in the initial relationship graph is optimized by using the archive processing large model to obtain a relationship graph, and the global information includes the relationship graph.

[0044] Optionally, the third processing module is specifically configured to:

[0045] The task list of the to-be-processed archive is constructed by using the processing result, wherein the task list includes archive management, information query and decision support.

[0046] The target archive task in the task list is matched, and the target archive task is automatically executed.

[0047] In a third aspect, an electronic device is provided, including a processor and a memory, the memory stores computer readable instructions, when the computer readable instructions are executed by the processor, the steps in the method provided in the first aspect are executed.

[0048] In a fourth aspect, a readable storage medium is provided, the readable storage medium stores a computer program, when the computer program is executed by a processor, the steps in the method provided in the first aspect are executed.

[0049] Other features and advantages of the present application will be set forth in the following description, and in part will be apparent from the description, or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be considered as a limitation to the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor.

[0051] Figure 1 A flow chart of a method for processing an archival task provided by an embodiment of the present application;

[0052] Figure 2 A flow chart of an implementation method for processing an archival task provided by an embodiment of the present application;

[0053] Figure 3 A schematic block diagram of an apparatus for processing an archival task provided by an embodiment of the present application;

[0054] Figure 4 A structural schematic block diagram of an apparatus for processing an archival task provided by an embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0056] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0057] First, some of the terms involved in the embodiments of the present application will be described in order to facilitate the understanding of those skilled in the art.

[0058] The Transformer model architecture uses a Self-Attention structure to replace the commonly used RNN network structure in NLP tasks. Its biggest advantage is that it can be calculated in parallel.

[0059] NLP is the English abbreviation for Neuro-Linguistic Programming. In Hong Kong, China, it is also translated as Mind-Body Syntax Programming. N (Neuro) refers to the nervous system, including the brain and the thinking process; L (Linguistic) refers to language, more accurately, from the input of sensory signals to the process of forming meaning; P (Programming) refers to a set of specific instructions to be executed to produce a certain result. That is, our habits in thinking and behavior, just like the programs in the computer, can be changed by updating the software. Therefore, NLP is interpreted as the study of how our brain works. Therefore, NLP is translated as Mind-Body Syntax Programming or Neuro-Linguistic Programming.

[0060] ‌Seq2Seq model (Sequence to Sequence) is a deep learning model mainly used for processing sequence data. Its core idea is to convert input sequences into output sequences through an encoder (Encoder) and a decoder (Decoder).

[0061] ‌ML model (Machine Learning Model) is a model that enables computers to learn from data and make predictions or decisions through algorithms. Machine learning is a subfield of artificial intelligence, and its core idea is to use algorithms to model and learn from data, so as to achieve prediction or classification of new data. The basic steps of machine learning include data collection, data preparation, model selection, training, evaluation and parameter adjustment.

[0062] The present application is applied to the scene of file processing, and the specific scene is to identify global information in the file through a large model, process the global information through a small model to obtain a processing result, and finally realize the task of the target file according to the processing result.

[0063] In the field of archive management and information processing, with the rapid development of information technology and the sharp increase of archive data, how to efficiently and accurately process and analyze these archive data has become a problem to be solved. Traditional archive processing methods often rely on manual operation or simple automated tools, which are not sufficient when dealing with large-scale, complex semantic relationship archive tasks. Manual processing is time-consuming and labor-intensive, and is easily affected by personal experience and subjective judgment; simple automated tools can only complete basic classification, retrieval and other tasks, and are difficult to cope with complex tasks involving deep semantic understanding and cross-archive information integration. The rapid development of artificial intelligence technology provides a new way for archive processing. In particular, the continuous progress of natural language processing (NLP) and machine learning (ML) technology enables computers to more deeply understand and analyze text data. However, a single model still faces many challenges in processing complex archive tasks, and the analysis is not accurate enough.

[0064] To this end, the present application preliminarily processes the to-be-processed archive through a preset archive processing large model to obtain global information of the to-be-processed archive, wherein the preliminary processing includes at least one of global understanding, context analysis, extracting key information, extracting semantic relationship and cross-archive information integration, and the archive processing large model is trained by combining unsupervised learning and supervised learning; the global information is finally processed through a preset archive processing small model to obtain a processing result, wherein the final processing includes at least one of extracting keywords, archive classification and generating an archive abstract, and the archive processing small model is trained by supervised learning; and a target archive task is processed through the processing result, wherein the target archive task includes archive management, information query or decision support of the to-be-processed archive. Through this method, the effect of efficiently and accurately processing archive tasks can be achieved.

[0065] In an embodiment of the present application, the execution subject can be a processing archive task device in a processing archive task system. In actual application, the processing archive task device can be a terminal device and a server, etc. electronic device, which is not limited here.

[0066] The method for processing archive tasks according to the embodiments of the present application will be described in detail below. Figure 1 The method for processing archive tasks according to the embodiments of the present application will be described in detail below.

[0067] Please refer to Figure 1 , Figure 1 A flowchart of a method for processing archive tasks according to an embodiment of the present application is shown in Figure 1 The method for processing archive tasks according to an embodiment of the present application includes:

[0068] Step 110: preliminarily processing the to-be-processed archive through a preset archive processing large model to obtain global information of the to-be-processed archive.

[0069] The preliminary processing includes at least one of global understanding, context analysis, key information extraction, semantic relationship extraction, and cross-file information integration. The preliminary processing can also include file overall meaning analysis and summarization. The global understanding includes global meaning and effect understanding, and can also include global structure layout understanding. The context analysis includes context correlation analysis, meaning analysis, and specific context content analysis. The key information extraction includes extraction of key text, phrase, or word. The semantic relationship can be a context semantic relationship or a semantic relationship between different documents. The cross-file information integration includes information integration of the to-be-processed file and a file uploaded by another system or another department. The to-be-processed file can be a file of various departments and units, such as a school student file, a unit employee file, or a file of a related department, and the application is not limited thereto. The file processing large model is trained by combining unsupervised learning and supervised learning.

[0070] In some embodiments of the application, the global information includes at least one of global definition, context relationship, key text, and semantic relationship of the to-be-processed file, and the processing result includes at least one of keywords, file type, and file summary of the to-be-processed file.

[0071] In the above process, the application can provide data basis for subsequent target file task processing by obtaining global information and processing results.

[0072] The global information can also include at least one of context relationship, key text, and semantic relationship of other files in cross-file information integration, and the names and addresses of the other files. The processing result can also include file directory, file unit, and file purpose.

[0073] In some embodiments of the application, before the global information of the to-be-processed file is obtained by preliminary processing of the to-be-processed file by the preset file processing large model, Figure 1 The method also includes: constructing a file set of a plurality of types of historical files; pre-processing the file set to obtain training samples, wherein the pre-processing includes at least one of word segmentation, stop word removal, and stem extraction; using a mask method in unsupervised learning to train a basic neural network architecture using the training samples to obtain an initial file processing large model; labeling the file set to obtain a labeled file set; and using a supervised learning method to adjust parameters of the initial file processing large model using the labeled file set to obtain the file processing large model.

[0074] In the above process, the initial archive processing large model is trained by constructing an archive set and using an unsupervised mask method, and the archive processing large model is obtained by fine-tuning the initial archive processing large model using a supervised learning method. The archive processing large model trained using a double supervision learning method can quickly and accurately identify the archive to be processed and obtain global information.

[0075] The preprocessing can further include adding, deleting, modifying, querying, truncating, and merging the data in the archives. The training samples include multiple archives. The labeling of the archive set can include labeling the words, types, or other key information of the archives in the archive set. The word segmentation can use a professional word segmentation tool or model to segment the text into words or phrases. The stop word removal includes removing stop words such as "of", "is", "in" and the like in the text to reduce noise. The stem extraction can extract the stem of the English text to restore the word to its basic form. The preprocessed text will be converted into an input form that the model can understand, such as a word embedding vector.

[0076] Optionally, the initial archive processing large model is obtained by training the basic neural network architecture using the training samples in the unsupervised learning mask method. The key words, fields, or texts in the archive set can be masked using a random mask method. The masked archive set is input into the basic neural network architecture for training. The output of the key words, fields, or texts is compared with the masked key words, fields, or texts, and the basic neural network is continuously adjusted so that the final output is the pre-masked content, and the initial archive processing large model is obtained.

[0077] The basic neural network architecture can use a Transformer architecture or BERT, etc. The self-attention mechanism is used to capture long-distance dependencies in the archive text, and a multi-layer encoder-decoder structure is used to achieve complex semantic understanding and generation. The archive processing large-scale archive dataset is collected, including various types of archive files, historical records, reports, etc., to ensure the diversity and comprehensiveness of the data. The collected dataset is used to pre-train the archive processing large model, so that the large model learns the representation and understanding ability of the archive knowledge, including vocabulary, syntax, semantics, and context relationships, etc.

[0078] In some embodiments of the present application, before the global information of the archive to be processed is obtained by using the preset archive processing large model to preliminarily process the archive to be processed, Figure 1The method also includes processing the archives in the archive set by the archive processing large model to obtain a plurality of global information of the archive set; labeling the archive type, abstract and keyword in the plurality of global information to obtain a second training sample; and training the basic convolutional neural network using the second training sample in a supervised learning manner to obtain the archive processing small model.

[0079] In the above process, the global information labeled can be used as a training sample for supervised learning of the archive processing small model, and the archive processing small model obtained can process the global information and accurately obtain the processing result.

[0080] The archive type, abstract and keyword in the plurality of global information can be labeled respectively, and the labeled global information can be used as the second training data. In addition, the labeled information can also include keywords, texts or archive summaries. The archive processing small model selects a convolutional neural network (CNN) for text classification and a sequence-to-sequence (Seq2Seq) model for abstract generation according to the requirements of specific tasks, including specific task execution scenarios. For each specific archive task, collect and prepare a task-related dataset. These datasets should contain specific information required for the task, such as keyword annotations, classification labels or abstract samples. The archive processing small model is trained using the task-related dataset, so that it can efficiently and accurately complete specific tasks based on the global information provided by the archive processing large model. During the training process, transfer learning or fine-tuning techniques can be used to transfer part of the knowledge of the archive processing large model to the small model to improve the performance of the small model. The specific part of the knowledge can be set according to the requirements.

[0081] Optionally, the archive processing small model is obtained by training the basic convolutional neural network using the second training sample in a supervised learning manner, including: inputting the global information with the label into the basic convolutional neural network as the second training sample, comparing the output content with the labeled content, adjusting the model parameters of the basic convolutional neural network according to the comparison result, until the cross-entropy calculated from the output content and the labeled content meets the expected requirement, and then the archive processing small model is obtained.

[0082] The archive processing large model is used for understanding and processing global archive information, and the archive processing small model is used for rapid response to specific archive tasks; the archive processing large model is responsible for processing complex semantic relationships, context understanding, and information integration across archives, and the archive processing small model is used for specific optimization for specific tasks, including keyword extraction, classification, and summary generation; a large-scale archive dataset is used to pre-train the archive processing large model to have archive knowledge understanding and analysis capabilities; for specific archive tasks, collect and prepare task-related datasets to train the archive processing small model, so that it can complete specific tasks based on the global information provided by the large model; the global understanding ability of the archive processing large model and the specific processing ability of the archive processing small model are combined to process the input archive complex task in stages, first providing global information and context understanding by the archive processing large model, and then performing specific task execution and result output by the archive processing small model.

[0083] In some embodiments of the present application, the preliminary processing is cross-archive information integration, which is performed by a preset archive processing large model to obtain global information of the to-be-processed archive, including: constructing an initial relationship graph between the to-be-processed archive and the associated archive; optimizing the information in the initial relationship graph by the archive processing large model to obtain a relationship graph, and the global information includes the relationship graph.

[0084] In the above process, when processing cross-archive information integration, the relationship graph between the to-be-processed archive and the associated archive can be used as global information to provide a data basis for subsequent archive processing small model processing.

[0085] The initial relationship graph includes the relationship between the to-be-processed archive and the associated archive, including directory relationship, content relationship, and type relationship.

[0086] Optionally, the information in the initial relationship graph is optimized by the archive processing large model, including deleting useless information such as auxiliary words and page numbers in the initial relationship graph by the archive processing large model, and extracting the main relationship content in the initial relationship graph to obtain the relationship graph.

[0087] The cross-archive information integration and the associated graph construction can be realized in the following way: for a scenario that requires integration of multiple archive information to complete a task, the application uses the memory network or graph neural network capability of the large model to construct the associated graph between the archives. The specific steps are as follows:

[0088] Entity and relationship extraction: Extract entities (such as names of people, places, events, etc.) and their relationships (such as "A is the father of B", "C happened D event", etc.) from each archive text. This can be achieved through named entity recognition (NER) and relationship extraction techniques.

[0089] Preliminary entity relationship graph construction: Represent the extracted entities and relationships as nodes and edges, and construct a preliminary entity relationship graph. Nodes represent entities, and edges represent relationships between entities. At this time, the graph may contain some incorrect or redundant relationships.

[0090] Optimization of the association graph: Use the semantic understanding ability of the archive processing large model to refine and optimize the preliminary entity relationship graph. The archive processing large model can identify implicit relationships (such as relationships inferred from context) and delete incorrect relationships (such as incorrectly extracted relationships due to text understanding errors). The optimized association graph more accurately reflects the associations between archives.

[0091] Application of the association graph: Use the optimized association graph as one of the inputs of the archive processing small model. When processing specific tasks, the archive processing small model can use information from the association graph, such as the relationships between entities and the temporal relationships between events, to assist in completing the task. For example, in the archive summary generation task, the archive processing small model can generate a more accurate and coherent summary based on key entities and relationships in the association graph.

[0092] Take the classification and summary generation of archives as an example to illustrate the comprehensive application of large and small model combination, attention mechanism and association graph in the invention. The archive processing large model processes archive texts and outputs global information representation and constructs the association graph between archives. The archive processing small model uses the attention mechanism to selectively integrate the global information of the large model when processing classification tasks, improving the accuracy of classification. The small model uses key entities and relationships in the association graph to generate more accurate and coherent summaries when processing summary generation tasks. In this way, the invention realizes efficient processing of complex tasks of archives.

[0093] Optionally, during the use of the archive processing large model and the archive processing small model, user feedback and task execution results can also be collected continuously to optimize and iterate the archive processing large model and the archive processing small model. By introducing new data sets, adjusting model structures or parameters, etc., the performance and accuracy of the model can be improved. Regularly evaluate the synergistic effect of the large model and the small model to ensure that the information flow and task allocation between the two reach the optimal state.

[0094] Step 120: Perform final processing on the global information through the preset archive processing small model to obtain the processing result.

[0095] Among them, the final processing includes at least one of extracting keywords, archiving classification and generating an archive summary. It can also include catalog generation and summary generation, etc., and the archive processing small model is obtained by supervised learning.

[0096] In some embodiments of the present application, the global information is finally processed by a preset archive processing small model to obtain a processing result, including: obtaining a correlation score between the archive processing large model and the archive processing small model; weighting and summing the information in the global information according to the correlation score through the attention mechanism or the gating mechanism of the archive processing small model to obtain part of the global information; and finally processing the part of the global information through the archive processing small model to obtain the processing result.

[0097] In the above process, the archive processing small model can filter the information in the global information through the attention mechanism or the gating mechanism, select part of the most important global information as the data basis for final processing, and obtain a more accurate processing result.

[0098] Specifically, the method for obtaining the global information and the processing result can be obtained by the following steps:

[0099] Suppose the archive processing large model encodes the input archive text and outputs the global information as , wherein is the hidden state vector of the i-th position, i is a positive integer, and n is the length of the output sequence;

[0100] S2: Suppose the current state of the archive processing small model when processing a specific task is , wherein t represents the time step or the number of layers in the small model processing process;

[0101] S3: Dot product attention mechanism is used to calculate the correlation score between the current state of the archive processing small model and each output position of the archive processing large model, and the calculation formula is:

[0102] ;

[0103] , wherein represents the transpose of , and is the output vector of the archive processing large model;

[0104] S4: The correlation score is normalized by using a softmax function to obtain the attention weight , and the sum of the weights of all positions is 1; ​

[0105] = et;

[0106] S5: Weighted sum of the output H of the large model according to the attention weight to obtain the context information required by the archive processing small model ;

[0107] = ;

[0108] S6: Fuse the context information with the current state of the archive processing small model to update the state of the archive processing small model and obtain a new state ;

[0109] S7: Repeat steps S3 to S6 until the archive processing small model completes the processing of a specific task and outputs the final processing result.

[0110] Optionally, the information in the global information is weighted and summed according to the relevance score through the attention mechanism or the gating mechanism of the archive processing small model, to obtain part of the global information; the archive processing small model processes part of the global information to obtain the processing result, which can be realized by the following way:

[0111] Calculate the relevance score: For each current state of the small model when processing the archive text (such as the hidden state in RNN or the query vector in Transformer), calculate the relevance score between it and the output of the large model (such as the output representation of the encoder). This is usually achieved through dot product, bilinear function or trainable feedforward neural network. The higher the score, the more relevant the output of the large model is to the current state of the small model.

[0112] Weighted sum: According to the calculated relevance score, the output of the large model is weighted and summed. The weight is obtained by normalizing the relevance score through the softmax function. In this way, the small model can dynamically select and integrate global information according to the relevance of the large model output, forming context information.

[0113] Application of context information: The context information obtained by weighted sum is used as additional input of the small model, or is spliced, added, etc. with the state of the small model to assist the small model to complete specific archive tasks. For example, in the text classification task, the context information can help the small model to more accurately understand the semantics of the text, thereby improving the accuracy of classification.

[0114] Step 130: Process the target archive task through the processing result.

[0115] The target archive task includes archive management, information query, or decision support of the to-be-processed archive. The target archive task can also include other tasks such as directory query and archive application. The archive management can be classification or storage of the archive. The information query can query key information or archive type in the to-be-processed archive. The decision support includes application, calling, or association of other archives of the target archive.

[0116] In some embodiments of the present application, the target archive task is processed by using the processing result, including: constructing a task list of the to-be-processed archive by using the processing result, wherein the task list includes archive management, information query, and decision support; and matching the target archive task in the task list and automatically executing the target archive task.

[0117] In the above process, the present application can pre-construct an accurate task list, quickly obtain the target archive task by task matching, and automatically execute the target archive task.

[0118] The task list can also include other tasks such as directory query and archive application.

[0119] Optionally, the matching of the target archive task in the task list and the automatic execution of the target archive task can be performed by keyword extraction and keyword similarity matching, or directly by similarity matching according to the name or text of the task to obtain the target archive task.

[0120] In the above Figure 1 In the above process, the present application performs preliminary processing on the to-be-processed archive by using a preset archive processing large model to obtain global information of the to-be-processed archive, wherein the preliminary processing includes at least one of global understanding, context analysis, extraction of key information, extraction of semantic relationship, and cross-archive information integration, and the archive processing large model is trained by combining unsupervised learning and supervised learning; the global information is finally processed by using a preset archive processing small model to obtain a processing result, wherein the final processing includes at least one of keyword extraction, archive classification, and archive abstract generation, and the archive processing small model is trained by using supervised learning; and the target archive task is processed by using the processing result, wherein the target archive task includes archive management, information query, or decision support of the to-be-processed archive. By using the method, the archive task can be efficiently and accurately processed.

[0121] The implementation method of the archive task processing of the embodiments of the present application will be described in detail. Figure 2 The implementation method of the archive task processing of the embodiments of the present application will be described in detail.

[0122] Please refer to Figure 2 , Figure 2A flowchart of an embodiment of a method for processing an archival task is provided in the present application, as shown in Figure 2 The embodiment of the method for processing an archival task includes:

[0123] Step 210: Construct an initial large model.

[0124] Specifically: using the mask method in unsupervised learning, the training sample is used to train the basic neural network architecture to obtain the initial archival processing large model.

[0125] Step 220: Train the initial large model through data to obtain a large model.

[0126] Specifically: using the supervised learning method, the labeled archival set is used to adjust the parameters of the initial archival processing large model to obtain the archival processing large model.

[0127] Step 230: Process the to-be-processed archives through the large model to obtain global information.

[0128] Specifically: the archival processing large model is used to process the archives in the archival set to obtain multiple global information of the archival set.

[0129] Step 240: Construct an initial small model.

[0130] Specifically: the preset convolutional neural network structure is used as the initial small model.

[0131] Step 250: Train the initial small model through data to obtain a small model.

[0132] Specifically: the archival processing large model is used to process the archives in the archival set to obtain multiple global information of the archival set; the archival types, abstracts and keywords in the multiple global information are labeled to obtain a second training sample; the supervised learning method is used to train the basic convolutional neural network using the second training sample to obtain the archival processing small model.

[0133] Step 260: Process the global information through the small model to obtain a processing result.

[0134] Specifically: the correlation score between the archival processing large model and the archival processing small model is obtained; the attention mechanism or the gating mechanism of the archival processing small model is used to weight and sum the information in the global information according to the correlation score to obtain partial global information; the partial global information is finally processed through the archival processing small model to obtain the processing result.

[0135] Step 270: Execute the archival task according to the processing result.

[0136] Specifically: the target archival task in the task list is matched, and the target archival task is automatically executed.

[0137] In addition, Figure 2 The specific methods and steps shown can be found in Figure 1 The methods shown are not described in detail here.

[0138] The foregoing describes a method of processing archival tasks, and the following describes an apparatus for processing archival tasks. Figures 1-2 The foregoing describes a method of processing archival tasks, and the following describes an apparatus for processing archival tasks. Figures 3-4 The foregoing describes a method of processing archival tasks, and the following describes an apparatus for processing archival tasks.

[0139] Please refer to Figure 3 , a schematic block diagram of an apparatus 300 for processing archival tasks provided in an embodiment of the present application, which can be a module, program segment or code on an electronic device. The apparatus 300 corresponds to the method embodiments described above and can perform each step involved in the method embodiments described above. The specific functions of the apparatus 300 can be found in the description below. To avoid repetition, the detailed description is appropriately omitted here. Figure 1 The foregoing describes a method of processing archival tasks, and the following describes an apparatus for processing archival tasks. Figure 1 The foregoing describes a method of processing archival tasks, and the following describes an apparatus for processing archival tasks.

[0140] Optionally, the apparatus 300 comprises:

[0141] A first processing module 310 is configured to perform preliminary processing on the archival data to be processed by using a preset archival processing large model to obtain global information of the archival data to be processed, wherein the preliminary processing includes at least one of global understanding, context analysis, extraction of key information, extraction of semantic relationship and cross-archival information integration, and the archival processing large model is trained by combining unsupervised learning and supervised learning.

[0142] A second processing module 320 is configured to perform final processing on the global information by using a preset archival processing small model to obtain a processing result, wherein the final processing includes at least one of extraction of keywords, archival classification and generation of an archival abstract, and the archival processing small model is trained by supervised learning.

[0143] A third processing module 330 is configured to process a target archival task by using the processing result, wherein the target archival task includes archival management, information query or decision support of the archival data to be processed.

[0144] Optionally, the global information includes at least one of global definition, context connection, key text and semantic relationship of the archival data to be processed, and the processing result includes at least one of keywords, archival type and archival abstract of the archival data to be processed.

[0145] Optionally, the apparatus further comprises:

[0146] The first training module is configured to, before the first processing module performs preliminary processing on the to-be-processed archives by using the preset archive processing large model to obtain global information of the to-be-processed archives, construct an archive set of multiple types of historical archives; perform preprocessing on the archive set to obtain training samples, wherein the preprocessing includes at least one of word segmentation, stop word removal and stem extraction; train a basic neural network architecture by using the training samples in a mask manner in unsupervised learning to obtain an initial archive processing large model; label the archive set to obtain a labeled archive set; and adjust parameters of the initial archive processing large model by using the labeled archive set in a supervised learning manner to obtain the archive processing large model.

[0147] Optionally, the apparatus further includes:

[0148] The second training module is configured to, before the second processing module performs preliminary processing on the to-be-processed archives by using the preset archive processing large model to obtain global information of the to-be-processed archives, process archives in the archive set by using the archive processing large model to obtain multiple global information of the archive set; label archive types, abstracts and keywords in the multiple global information to obtain second training samples; and train a basic convolutional neural network by using the second training samples in a supervised learning manner to obtain an archive processing small model.

[0149] Optionally, the second processing module is specifically configured to:

[0150] Obtain a correlation score between the archive processing large model and the archive processing small model; perform weighted summation on information in the global information according to the correlation score by using an attention mechanism or a gating mechanism of the archive processing small model to obtain partial global information; and perform final processing on the partial global information by using the archive processing small model to obtain a processing result.

[0151] Optionally, the preliminary processing is cross-archive information integration, and the first processing module is specifically configured to:

[0152] Construct an initial relationship graph between the to-be-processed archives and the correlative archives; optimize information in the initial relationship graph by using the archive processing large model to obtain a relationship graph, and the global information includes the relationship graph.

[0153] Optionally, the third processing module is specifically configured to:

[0154] Construct a task list of the to-be-processed archives by using the processing result, wherein the task list includes archive management, information query and decision support; and match a target archive task in the task list and automatically execute the target archive task.

[0155] Please refer to Figure 4A structural schematic block diagram of a device for processing an archive task is provided in the embodiments of the present application. The device can include a memory 410 and a processor 420. Optionally, the device can further include a communication interface 430 and a communication bus 440. The device and the above-mentioned Figure 1 The method embodiments correspond to, and can execute Figure 1 The various steps involved in the method embodiments, the specific functions of the device can be referred to the description below.

[0156] Specifically, the memory 410 is configured to store computer readable instructions.

[0157] The processor 420 is configured to process the readable instructions stored in the memory, and can execute Figure 1 The various steps in the method.

[0158] The communication interface 430 is configured to communicate with other node devices for signaling or data. For example, the communication interface 430 is configured to communicate with a server or a terminal, or communicate with other device nodes, and the embodiments of the present application are not limited thereto.

[0159] The communication bus 440 is configured to realize the direct connection communication of the above-mentioned components.

[0160] In the embodiments of the present application, the communication interface 430 of the device is configured to communicate with other node devices for signaling or data. The memory 410 can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. The memory 410 can optionally be at least one storage device located away from the aforementioned processor. The memory 410 stores computer readable instructions, and when the computer readable instructions are executed by the processor 420, the electronic device executes the above-mentioned Figure 1 The method process. The processor 420 can be used in the device 300 and is configured to execute the functions in the present application. For example, the processor 420 described above can be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, and the embodiments of the present application are not limited thereto.

[0161] The embodiments of the present application also provide a readable storage medium. When the processor executes the computer program, the processor executes the method process performed by the electronic device in the method embodiments as shown in Figure 1 The method process.

[0162] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device can refer to the corresponding process in the foregoing method, and will not be repeated here.

[0163] To sum up, the embodiment of the present application provides a method, device and equipment for processing archive tasks and a readable storage medium, the method comprising: performing preliminary processing on the to-be-processed archive by a preset archive processing large model to obtain global information of the to-be-processed archive, wherein the preliminary processing comprises at least one of global understanding, context analysis, extraction of key information, extraction of semantic relationship and cross-archive information integration, and the archive processing large model is obtained by combining unsupervised learning and supervised learning; performing final processing on the global information by a preset archive processing small model to obtain a processing result, wherein the final processing comprises at least one of keyword extraction, archive classification and archive abstract generation, and the archive processing small model is obtained by supervised learning; and processing a target archive task by the processing result, wherein the target archive task comprises archive management, information query or decision support of the to-be-processed archive. The method can achieve the effect of efficiently and accurately processing archive tasks.

[0164] In several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can also be implemented by other manners. The apparatus embodiments described above are only schematic, for example, the flowcharts and block diagrams in the drawings show the possible implementation architectures, functions and operations of the apparatus, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that, in some alternative implementation manners, the functions noted in the blocks can also occur in different orders from those noted in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and the combination of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system for executing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0165] In addition, each functional module in the embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0166] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts of the prior art that make contributions or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0167] The above merely provides an example of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0168] The above merely provides an example of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0169] It should be noted that, in this document, the terms such as first and second are used merely 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", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the processes, methods, articles or devices including a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

Claims

1. A method of processing archival tasks, characterized by, The method comprises the following steps: The archive processing large model is trained by combining unsupervised learning and supervised learning in the following manner: an initial archive processing large model is obtained by training a basic neural network architecture using a mask method in unsupervised learning and using training samples; An archive set is labeled to obtain a labeled archive set; the parameters of the initial archive processing large model are adjusted using the labeled archive set by a supervised learning method to obtain the archive processing large model; The archive processing small model is trained by a supervised learning method in the following manner: the archive processing large model is used to process the archives in the archive set to obtain a plurality of global information of the archive set; the archive type, abstract and keywords in the plurality of global information are labeled to obtain a second training sample; the second training sample is used to train a basic convolutional neural network by a supervised learning method to obtain the archive processing small model; The archive processing large model is used to preliminarily process the to-be-processed archive to obtain global information of the to-be-processed archive, wherein the preliminary processing includes at least one of global understanding, context analysis, key information extraction, semantic relationship extraction, cross-archive information integration, archive overall meaning analysis and summary; The archive processing small model is used to finally process the global information to obtain a processing result, wherein the final processing includes at least one of keyword extraction, archive classification, archive abstract generation, directory generation and summary generation; The target archive task is processed by using the processing result, wherein the target archive task includes archive management, information query, decision support, directory query and archive application of the to-be-processed archive.

2. The method of claim 1, wherein, The global information includes at least one of global definition, context connection, key text and semantic relationship of the to-be-processed archive. The processing result includes at least one of keywords, archive type and archive abstract of the to-be-processed archive. Before the archive processing large model is used to preliminarily process the to-be-processed archive to obtain global information of the to-be-processed archive, the method further comprises:

3. The method of claim 2, wherein, The archive set of a plurality of types of historical archives is constructed; The archive set is preprocessed to obtain the training sample, wherein the preprocessing includes at least one of word segmentation, stop word removal and stem extraction. The global information is finally processed by the archive processing small model to obtain a processing result, which includes:

4. The method according to any one of claims 1 to 3, characterized in that, The correlation score between the archive processing large model and the archive processing small model is obtained; The information in the global information is weighted and summed according to the correlation score by using the attention mechanism or the gating mechanism of the archive processing small model to obtain partial global information; The partial global information is finally processed by the archive processing small model to obtain the processing result. The preliminary processing is the cross-archive information integration, and the archive processing large model is used to preliminarily process the to-be-processed archive to obtain global information of the to-be-processed archive, which includes:

5. The method according to any one of claims 1 to 3, characterized in that, An initial relationship graph is constructed between the to-be-processed archive and the associated archive; ​ The information in the initial relationship graph is optimized by the archive processing large model to obtain a relationship graph, and the global information includes the relationship graph.

6. The method according to any one of claims 1 to 3, characterized in that, The target archive task in the task list is matched, and the target archive task is automatically executed. The task list of the archive to be processed is constructed by the processing result, wherein the task list includes archive management, information query, and decision support. The target archive task in the task list is matched, and the target archive task is automatically executed.

7. An apparatus for processing archival tasks, characterized by Comprising: The first training module is configured to train the archive processing large model by combining unsupervised learning and supervised learning in the following manner: using a mask method in unsupervised learning, training a basic neural network architecture using training samples to obtain an initial archive processing large model; The archive set is labeled to obtain a labeled archive set; and the parameters of the initial archive processing large model are adjusted using the labeled archive set by a supervised learning method to obtain the archive processing large model; The second training module is configured to train the archive processing small model by a supervised learning method in the following manner: processing the archives in the archive set by the archive processing large model to obtain a plurality of global information of the archive set; labeling the archive type, abstract, and keywords in the plurality of global information to obtain a second training sample; and training a basic convolutional neural network using the second training sample by a supervised learning method to obtain the archive processing small model; The first processing module is configured to preliminarily process the archive to be processed by the archive processing large model to obtain global information of the archive to be processed, wherein the preliminary processing includes at least one of global understanding, context analysis, extracting key information, extracting semantic relationship, and cross-archive information integration; The second processing module is configured to finally process the global information by the archive processing small model to obtain a processing result, wherein the final processing includes at least one of extracting keywords, archive classification, and generating an archive abstract; The third processing module is configured to process a target archive task by the processing result, wherein the target archive task includes archive management, information query, or decision support of the archive to be processed.

8. An electronic device, comprising: Comprising: The memory stores computer readable instructions, and when the computer readable instructions are executed by the processor, the steps in the method of any one of claims 1-6 are run.

9. A computer-readable storage medium, characterized in that, Comprising: The computer program, when running on a computer, causes the computer to execute the method of any one of claims 1-6.

Images