Power report generation method based on large model and multi-document curriculum learning framework

Abstract

This application relates to a method for generating electricity reports based on a large-scale model and a multi-document learning framework. The method includes: responding to a question-and-answer operation on an electricity customer service platform, obtaining the query text for a specified electricity item; using a pre-defined large-scale language model to perform semantic understanding on the query text and extract key query information; the key query information indicates the target query scope, target query object, and target report type; retrieving multiple related documents for specified electricity items based on the target query scope and target query object in the key query information; integrating the multiple related documents for specified electricity items using a multi-document learning framework to generate simplified document information; and generating an electricity-specific project report based on the simplified document information and the target report type, serving as the query result of the electricity-specific project query text. This method can help provide high-quality, multi-demand electricity-specific report generation services.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power, and in particular to an electric power report generation method and device based on a large model and a multi-document course learning framework, a computer device, a computer readable storage medium, and a computer program product. BACKGROUND

[0002] Currently, when a customer service platform in the electric power industry processes specific report generation requirements, various problems and requirements may be encountered, such as user inquiries and queries on electric power equipment failure, electric power load forecasting, and electric power supply schemes.

[0003] In related technologies, traditional specific report generation methods can support answering single-element question requirements of users, but the processing effect is poor for multi-element composite problems. In addition, due to the complex information stored in the database, the traditional retrieval method has low search efficiency, resulting in poor quality of the generated report content. SUMMARY

[0004] Therefore, it is necessary to provide an electric power report generation method and device based on a large model and a multi-document course learning framework, a computer device, a computer readable storage medium, and a computer program product, which can improve the efficiency and effect of electric power specific report generation.

[0005] In a first aspect, the present application provides an electric power report generation method based on a large model and a multi-document course learning framework, comprising:

[0006] In response to a question and answer operation on an electric power customer service platform, an electric power specified project query text of the question and answer operation is obtained;

[0007] A preset large language model is used to perform semantic understanding on the electric power specified project query text, and query key information is extracted; the query key information is used to indicate a target query range, a target query object, and a target report type;

[0008] According to the target query range and the target query object in the query key information, a plurality of electric power specified project associated files are retrieved;

[0009] A multi-document course learning framework is used to integrate the plurality of electric power specified project associated files to generate simplified file information;

[0010] According to the simplified file information and the target report type, an electric power specified project report is generated as a query result of the electric power specified project query text.

[0011] In one embodiment, the method further comprises:

[0012] Construct training data and training prompt words based on the power designated project query task;

[0013] Obtain a basic large language model for model training, adjust the model parameters of the basic large language model using the training data and the training prompt words until the training end condition is met, and obtain the preset large language model.

[0014] In one of the embodiments, the method further comprises:

[0015] Based on the power designated project query task, obtain sample power designated project query texts of different project types and sample query key information corresponding to each of the sample power designated project query texts as the training data; the sample query key information has labeled labels;

[0016] Based on the power designated project query task, design execution instructions, power project association information, and input text indication information for the basic large language model as the training prompt words.

[0017] In one of the embodiments, the method further comprises:

[0018] Apply a curriculum learning strategy to a multi-document summarization task, combine a multi-document difficulty scoring function, and perform model training on a preset backbone model to construct the multi-document curriculum learning framework.

[0019] In one of the embodiments, the method further comprises:

[0020] Obtain a plurality of sample document sets, and calculate the difficulty scores of each of the sample document sets using the multi-document difficulty scoring function;

[0021] According to the difficulty scores of each of the sample document sets, sort the plurality of sample document sets;

[0022] Based on the sorted plurality of sample document sets, train the preset backbone model in a rolling window curriculum manner.

[0023] In one of the embodiments, the method further comprises:

[0024] Based on the number of documents in the document set, the total length of the documents, and the document length distribution, determine the multi-document difficulty scoring function.

[0025] In a second aspect, the application further provides an electric power report generation device based on a large model and a multi-document curriculum learning framework, comprising:

[0026] The query text acquisition module is configured to acquire power specified item query text of a question and answer operation of the power customer service platform in response to the question and answer operation.

[0027] The key information extraction module is configured to perform semantic understanding on the power specified item query text by using a preset large language model to extract query key information, which is used to indicate a target query range, a target query object, and a target report type.

[0028] The retrieval file obtaining module is configured to perform retrieval according to the target query range and the target query object in the query key information to obtain a plurality of power specified item associated files.

[0029] The simplified information integration module is configured to integrate the plurality of power specified item associated files by using a multi-document course learning framework to generate simplified file information.

[0030] The power report generation module is configured to generate a power specified item report according to the simplified file information and the target report type, as a query result of the power specified item query text.

[0031] In a third aspect, the present application further provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0032] The query text acquisition module is configured to acquire power specified item query text of a question and answer operation of the power customer service platform in response to the question and answer operation.

[0033] The key information extraction module is configured to perform semantic understanding on the power specified item query text by using a preset large language model to extract query key information, which is used to indicate a target query range, a target query object, and a target report type.

[0034] The retrieval file obtaining module is configured to perform retrieval according to the target query range and the target query object in the query key information to obtain a plurality of power specified item associated files.

[0035] The simplified information integration module is configured to integrate the plurality of power specified item associated files by using a multi-document course learning framework to generate simplified file information.

[0036] The power report generation module is configured to generate a power specified item report according to the simplified file information and the target report type, as a query result of the power specified item query text.

[0037] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0038] In response to the question and answer operation on the power customer service platform, obtain power designated project query text of the question and answer operation;

[0039] Understand the semantics of the power designated project query text by using a preset large language model, and extract query key information; the query key information is used to indicate a target query range, a target query object, and a target report type;

[0040] According to the target query range and the target query object in the query key information, retrieve a plurality of power designated project associated files;

[0041] Integrate the plurality of power designated project associated files by using a multi-document course learning framework to generate simplified file information;

[0042] According to the simplified file information and the target report type, generate a power designated project report as a query result of the power designated project query text.

[0043] In a fifth aspect, the present application also provides a computer program product comprising a computer program which, when executed by a processor, implements the following steps:

[0044] In response to the question and answer operation on the power customer service platform, obtain power designated project query text of the question and answer operation;

[0045] Understand the semantics of the power designated project query text by using a preset large language model, and extract query key information; the query key information is used to indicate a target query range, a target query object, and a target report type;

[0046] According to the target query range and the target query object in the query key information, retrieve a plurality of power designated project associated files;

[0047] Integrate the plurality of power designated project associated files by using a multi-document course learning framework to generate simplified file information;

[0048] According to the simplified file information and the target report type, generate a power designated project report as a query result of the power designated project query text.

[0049] The power report generation method, device, computer device, computer readable storage medium and computer program product based on the large model and multi-document course learning framework, by responding to the question and answer operation of the power customer service platform, obtaining the power specified project query text of the question and answer operation, then using the preset large language model to perform semantic understanding on the power specified project query text, extracting the query key information, the query key information is used to indicate the target query range, the target query object and the target report type, searching according to the target query range and the target query object in the query key information, obtaining a plurality of power specified project associated files, and then integrating the plurality of power specified project associated files by using the multi-document course learning framework to generate the simplified file information, generating the power specified project report according to the simplified file information and the target report type, as the query result of the power specified project query text, realizing the automatic generation optimization of the power special report, by using the large model and the multi-document course learning framework collaborative mode, the query text can be understood and processed based on the large model to extract the key information, and the multi-document course learning framework can be used to quickly integrate and generate the simplified information from the plurality of retrieved files, so that the special report generation can be faster and more accurate, which helps to provide high-quality and multi-demand special report generation service, and effectively improves the quality and efficiency of the power report. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other related drawings without creative labor based on these drawings.

[0051] Figure 1 A flowchart of a power report generation method based on a large model and a multi-document course learning framework in an embodiment;

[0052] Figure 2 A schematic diagram of a power report generation process based on the cooperation of a large model and a multi-document course learning framework in an embodiment;

[0053] Figure 3 A flowchart of a power report generation method based on a large model and a multi-document course learning framework in another embodiment;

[0054] Figure 4 A structural block diagram of a power report generation device based on a large model and a multi-document course learning framework in an embodiment;

[0055] Figure 5 An internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0056] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0057] In one exemplary embodiment, as shown in Figure 1 A power report generation method based on a large model and a multi-document curriculum learning framework is provided. The embodiment illustrates the method applied to a terminal. It should be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and can be realized through the interaction of the terminal and the server. In the embodiment, the method includes the following steps 101 to 104. Wherein:

[0058] Step 101, in response to a question and answer operation on a power customer service platform, obtaining a power designated project query text of the question and answer operation.

[0059] In actual application, as shown in Figure 2 The user input query information can be obtained as the power designated project query text in response to the question and answer operation on the power customer service platform.

[0060] Step 102, using a preset large language model to perform semantic understanding on the power designated project query text, and extracting query key information.

[0061] The query key information can be used to indicate the target query range, the target query object, and the target report type.

[0062] After obtaining the power designated project query text, as shown in Figure 2 The preset large language model can be used to understand and retrieve the query key information, such as keywords indicating the target query range, the target query object, and the target report type, to further input the keywords extracted by the model into the database for retrieval to find the corresponding document.

[0063] Exemplarily, in the semantic understanding stage, the preset large language model can be used to perform an entity recognition task. By performing semantic understanding on the input power designated project query text, the user specified report type (i.e. target report type) and the keywords required for report generation (i.e. target query range, target query object) can be extracted, such as the time range of the data to be queried, the data set name to which the data to be queried belongs, etc. Optionally, for different report types, the extracted multiple keywords can be classified and sorted, such as according to time sequence, according to different data sets, etc.

[0064] Step 103, retrieving according to the target query range and target query object in the query key information, obtaining a plurality of power specified project associated files.

[0065] In a specific implementation, as shown in Figure 2 , the relevant documents in the information database can be retrieved according to the target query range and target query object in the query key information, such as the time range involved in the data to be queried, the data set name to which the data to be queried belongs, and the like, and then a plurality of power specified project associated files can be obtained.

[0066] Step 104, integrating the plurality of power specified project associated files by using a multi-document curriculum learning framework to generate concise file information.

[0067] As an example, the multi-document curriculum learning framework can be a curriculum learning (CL) enhanced framework for multi-document summarization (MDS), namely MDS-CL.

[0068] After the plurality of power specified project associated files, as shown in Figure 2 , the multi-document curriculum learning framework can be used to integrate the plurality of power specified project associated files retrieved, and then concise file information such as a concise report summary can be generated.

[0069] Specifically, for the multi-document curriculum learning framework, the task of generating a summary according to the documents retrieved from the information database can be regarded as a multi-document summarization task, so as to generate a concise summary based on a plurality of documents.

[0070] Step 105, generating a power specified project report according to the concise file information and the target report type, as the query result of the power specified project query text.

[0071] As an example, the report type can include but is not limited to a regular report, a specific special report, a special summary report, and a specific task data analysis report.

[0072] In actual application, according to the special report type required by the user (i.e., the target report type), such as Figure 2 , and the generated concise summary (i.e., the concise file information), the special report required by the user can be generated, and the model based on the specified report form to generate the corresponding report content is realized.

[0073] In an example, the technical solution of the embodiment can adopt a large language model to read and understand user input information, retrieve keywords into an information database, and efficiently generate concise summary content by using a multi-document summary course learning enhancement framework, thereby combining the data understanding ability of the multi-document course learning enhancement framework, the language retrieval and generation ability of the large model, efficiently extracting, analyzing and reporting generation of multi-source data in the power field, making the power special report generation process more intelligent and efficient, improving the generation quality and efficiency of the power special report, and better meeting the industry demand.

[0074] In the above power report generation method based on the large model and the multi-document course learning framework, the power specified project query text is obtained in response to the question and answer operation of the power customer service platform, and then the preset large language model is used for semantic understanding of the power specified project query text to extract the query key information. The target query range and target query object in the query key information are retrieved to obtain a plurality of power specified project associated files, and then the multi-document course learning framework is used to integrate the plurality of power specified project associated files to generate concise file information. The power specified project report is generated according to the concise file information and the target report type, which is the query result of the power specified project query text, realizes the optimization of automatic generation of power special report, and through the collaborative mode of large model and multi-document course learning framework, the query text can be understood and processed based on the large model to extract key information, and the multi-document course learning framework can be used to quickly integrate and generate concise information from the retrieved multiple files, so that the special report generation can be faster and more accurate, which helps to provide high-quality and multi-demand special report generation service, and effectively improves the quality and efficiency of the power report.

[0075] In an exemplary embodiment, the following steps can also be included:

[0076] The training data and the training prompt word based on the power specified project query task are constructed, the basic large language model used for model training is obtained, the model parameters of the basic large language model are adjusted by using the training data and the training prompt word, until the training end condition is met, and the preset large language model is obtained.

[0077] In practical application, the key information extraction can be regarded as a named entity recognition problem, and the key information extraction based on the basic large language model can adopt a prompt fine-tuning method. By using an open source large language model as a base framework and performing prompt fine-tuning based thereon, the model can adjust the parameters by comparing the generated answer and the correct answer during the prompt fine-tuning process, so as to better adapt to specific tasks.

[0078] In an example, in the prompt fine-tuning stage, fine-tuning data can be collected and a prompt can be designed based on the power designated project query task, that is, training data and training prompts; through prompt engineering for few-shot fine-tuning, the large language model can effectively deal with the named entity recognition problem of the special report generation, and after training the model using only a small amount of labeled data, a high accuracy can be maintained in the test stage, thereby realizing the specific task fine-tuning of the large model based on the prompt engineering and the fine-tuning data set.

[0079] In the embodiment, by constructing the training data and the training prompt based on the power designated project query task, a basic large language model for model training is obtained, the model parameters of the basic large language model are adjusted using the training data and the training prompt until the training end condition is met, and a preset large language model is obtained, which can adjust the parameters of the large language model based on the prompt and the training data for the power designated project query task.

[0080] In an exemplary embodiment, the training data and the training prompt based on the power designated project query task can include the following steps:

[0081] Based on the power designated project query task, sample power designated project query texts of different project types and sample query key information corresponding to each sample power designated project query text are obtained as the training data; the sample query key information has labeled labels; based on the power designated project query task, execution instructions, power project association information, and input text indication information for the basic large language model are designed as the training prompts.

[0082] In a specific implementation, a fine-tuning data set for power special reports can be constructed and a prompt can be designed to adjust the parameters of the large language model using the prompt and the fine-tuning data set. For example, in order to enable the model to complete the named entity recognition problem, the fine-tuning data set can be a task-specific data set with labels, which can include multi-class user input content and labeled key information corresponding to each piece of content, that is, sample power designated project query texts of different project types and sample query key information corresponding to each sample power designated project query text, and the sample query key information has labeled labels.

[0083] For example, the prompt word (i.e., the training prompt word) can include multiple elements: instructions, background information, input text. Among them, the instruction (i.e., the execution instruction) can be a short sentence for controlling the model to perform a certain task, such as "as an electric power special report clerk, extract key information according to user input content" when processing an electric power special report; the background information (i.e., the electric power project associated information) can be the relevant information provided by the few-shot learning, and the background information of the prompt word can include "the key words include time such as x year / x month / x day to y year / y month / y day, the data set name to which the data to be queried belongs, the report type; the input text (i.e., the input text indication information) can be used to determine the text that needs to be processed by the model, such as the user input content in the fine-tuning data.

[0084] In this embodiment, by querying the electric power specified project based on the electric power specified project query task, sample electric power specified project query texts of different project types and sample query key information corresponding to each sample electric power specified project query text are obtained as training data, and then based on the electric power specified project query task, the execution instruction, the electric power project associated information, and the input text indication information for the basic large language model are designed as the training prompt word, which provides data support for parameter adjustment of the large language model for the electric power specified project query task.

[0085] In one exemplary embodiment, the following steps can also be included:

[0086] By applying the curriculum learning strategy to the multi-document summarization task and combining the multi-document difficulty scoring function, the preset backbone model is trained to construct the multi-document curriculum learning framework.

[0087] In an example, the curriculum learning enhanced framework for multi-document summarization can enhance the multi-document summarization based on the curriculum learning strategy by applying the curriculum learning strategy to the multi-document summarization task. To distinguish between simple examples and difficult examples, a multi-document difficulty scoring function can be designed for multiple source documents, which can be composed of three difficulty heuristics, such as the number of documents, the total length of the documents, and the length distribution. Then, a rolling window curriculum approach can be used to arrange the sorted dataset, thereby effectively fine-tuning the preset backbone model.

[0088] In this embodiment, by applying the curriculum learning strategy to the multi-document summarization task and combining the multi-document difficulty scoring function, the preset backbone model is trained to construct the multi-document curriculum learning framework, which can efficiently generate concise summary content based on the data understanding ability of the multi-document curriculum learning enhanced framework, and improve the generation quality and efficiency of the electric power special report.

[0089] In an exemplary embodiment, the model training of the preset backbone model by applying the course learning strategy to the multi-document summarization task and in combination with the multi-document difficulty scoring function can include the following steps:

[0090] Obtaining a plurality of sample document sets, calculating the difficulty scores of each of the sample document sets using the multi-document difficulty scoring function, sorting the plurality of sample document sets according to the difficulty scores of each of the sample document sets, and training the preset backbone model in a rolling window course manner based on the sorted plurality of sample document sets.

[0091] In actual application, for constructing a multi-document course learning framework, the difficulty scores of each document set (i.e., sample document set) can be calculated using the multi-document difficulty scoring function, and then each document set can be sorted according to the difficulty scores, and the preset backbone model can be trained in a rolling window course manner.

[0092] In an example, by defining the multi-document difficulty scoring function, the preset backbone model can be trained in a rolling window course manner. In this stage, the training data can be arranged into a learning course based on the sorted document sets to train the adopted model.

[0093] In an optional embodiment, the following course methods can be used to verify the effectiveness of course learning:

[0094] 1. No course: randomly sample from the original data set for training;

[0095] 2. One-Pass course: start training from the simplest example to the most difficult example, and will not move to the next batch of data before the model converges;

[0096] 3. Baby-Steps course: similar to One-Pass course, the difference is that it can gradually increase the complexity of the training data, rather than discarding simpler examples.

[0097] 4. Reverse course: a reverse course of One-Pass course, which can start training from the most difficult example to the simplest example.

[0098] 5. Rolling Window course: it combines the advantages of One-Pass course and Baby-Steps course, and can sample through a rolling, sliding window to train samples, so as to more effectively train the model and avoid the risk of overfitting.

[0099] In the embodiment, the difficulty scores of the sample document sets are calculated by using the multi-document difficulty scoring function, and then the plurality of sample document sets are sorted according to the difficulty scores of the sample document sets, and then the pre-set backbone model is trained in a rolling window course manner based on the plurality of sample document sets after sorting, so that the multi-document summary course learning enhancement framework can be used to efficiently generate concise summary content.

[0100] In one exemplary embodiment, the following steps can also be included:

[0101] The multi-document difficulty scoring function is determined based on the number of documents in the document set, the total length of the documents, and the document length distribution.

[0102] In a specific implementation, the difficulty score plays an important role in course learning, which can be used to distinguish simple examples and difficult examples. Traditional difficulty indicators proposed for text examples, such as n-gram frequency, specificity, word rarity, and sentence length, cannot be directly used for multi-document summarization tasks. Since the difficulty of multi-document summarization is affected by multiple factors, the technical solution of the embodiment proposes a heuristic method to measure the difficulty score of each document set, such as considering three heuristic methods, including the number of documents, the total length of the documents, and the document length distribution.

[0103] For example, the dataset can be represented as a set of M document sets, where each set is a sequence of N i documents about the same topic . Accordingly, each document set D i has a sequence of document lengths . The following gives the document features of interest, including the number of documents, the total length of the documents, and the document length distribution, from which the multi-document difficulty scoring function can be determined.

[0104] For the number of documents, different multi-document sets have different numbers of documents, and the more documents there are, the more difficult it is to summarize. The difficulty of the multi-document set D i is defined as follows:

[0105]

[0106] where Diff Di represents the difficulty score of the multi-document set D i , and N i is the number of documents in the multi-document set D i .

[0107] For the total length of the documents, the total length of the documents can reflect the difficulty of the multi-document set D ithe total length difficulty. The abstract total length of a plurality of documents is longer, and the difficulty is greater because it contains more information. The document total length difficulty is defined as follows:

[0108]

[0109] wherein L i is the total length of the plurality of documents D i .

[0110] For the document length distribution, the document length distribution can reflect the relative proportion of the length of each document in the plurality of documents. The difficulty score of the more uneven document length distribution is lower than that of the uniform distribution. For example, if a certain document length occupies a higher proportion of the total length, its difficulty score is lower, which can be regarded as a pseudo single document. The difficulty heuristic for calculating the document length distribution of the plurality of documents D i is as follows:

[0111]

[0112] wherein, is the maximum document length in the plurality of documents D i .

[0113] Based on the above heuristic, a plurality of document summary task difficulty scoring function (i.e., a plurality of document difficulty scoring function) can be proposed, which can be represented in the following formalization:

[0114]

[0115] wherein the hyperparameter = 512, represents the plurality of document difficulty scoring function.

[0116] In this embodiment, by determining the plurality of document difficulty scoring function based on the number of documents in the document set, the total length of the document, and the document length distribution, the difficulty score of each document set can be effectively measured.

[0117] In an exemplary embodiment, as shown in Figure 3 , another flowchart of a power report generation method based on a large model and a plurality of document curriculum learning framework is provided. In this embodiment, the method includes the following steps:

[0118] In step 301, based on the power designated project query task, sample power designated project query texts of different project types and sample query key information corresponding to each sample power designated project query text are obtained as training data; the sample query key information has a labeled label. In step 302, based on the power designated project query task, an execution instruction for the basic large language model, power project associated information, and input text indication information are designed as training prompts. In step 303, a basic large language model used for model training is obtained, and the model parameters of the basic large language model are adjusted using the training data and the training prompts until a training end condition is met, obtaining a preset large language model. In step 304, a course learning strategy is applied to the multi-document summarization task, and a multi-document difficulty scoring function is combined to train the preset backbone model, and a multi-document course learning framework is constructed. In step 305, in response to the question and answer operation of the power customer service platform, the power designated project query text of the question and answer operation is obtained, the power designated project query text is understood by the preset large language model, and the query key information is extracted. In step 306, according to the target query range and the target query object in the query key information, a plurality of power designated project associated files are obtained. In step 307, the multi-document course learning framework is used to integrate the plurality of power designated project associated files to generate concise file information, and according to the concise file information and the target report type, a power designated project report is generated as the query result of the power designated project query text. It should be noted that the specific definition of the above steps can refer to the specific definition of the power report generation method based on the large model and the multi-document course learning framework described above, and will not be repeated here.

[0119] It should be understood that, although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or steps or stages in other steps.

[0120] Based on the same inventive concept, the embodiments of the present application also provide a large model and multi-document curriculum learning framework-based power report generation device for implementing the large model and multi-document curriculum learning framework-based power report generation method described above. The implementation scheme for solving problems provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more large model and multi-document curriculum learning framework-based power report generation device embodiments provided below can be referred to the limitations of the large model and multi-document curriculum learning framework-based power report generation method in the above, which will not be repeated here.

[0121] In one exemplary embodiment, as shown in Figure 4 a large model and multi-document curriculum learning framework-based power report generation device is provided, comprising:

[0122] The query text acquisition module 401 is configured to acquire power designated item query text of a question and answer operation of a power customer service platform in response to the question and answer operation.

[0123] The key information extraction module 402 is configured to perform semantic understanding on the power designated item query text by using a preset large language model to extract query key information. The query key information is used to indicate a target query range, a target query object, and a target report type.

[0124] The retrieval file obtaining module 403 is configured to retrieve a plurality of power designated item associated files according to the target query range and the target query object in the query key information.

[0125] The refined information integration module 404 is configured to integrate the plurality of power designated item associated files by using a multi-document curriculum learning framework to generate refined file information.

[0126] The power report generation module 405 is configured to generate a power designated item report as a query result of the power designated item query text according to the refined file information and the target report type.

[0127] In one embodiment, the device further comprises:

[0128] The training information construction module is configured to construct training data and training prompt words based on a power designated item query task.

[0129] The model training module is configured to acquire a basic large language model for model training, adjust model parameters of the basic large language model by using the training data and the training prompt words until a training end condition is met, and obtain the preset large language model.

[0130] In one embodiment, the model training module comprises:

[0131] The training data obtaining submodule is configured to obtain, based on the power designated project query task, sample power designated project query texts of different project types and sample query key information corresponding to each of the sample power designated project query texts as the training data; the sample query key information has a labeled label.

[0132] The training prompt word obtaining submodule is configured to design, based on the power designated project query task, an execution instruction for the basic large language model, power project associated information, and input text indication information as the training prompt word.

[0133] In one embodiment, the apparatus further comprises:

[0134] The multi-document curriculum learning framework modeling module is configured to model the multi-document curriculum learning framework by applying a curriculum learning strategy to a multi-document summarization task and combining a multi-document difficulty scoring function to model train a preset backbone model.

[0135] In one embodiment, the multi-document curriculum learning framework modeling module comprises:

[0136] The difficulty score calculation submodule is configured to obtain a plurality of sample document sets and calculate difficulty scores of each of the sample document sets using the multi-document difficulty scoring function.

[0137] The sorting submodule is configured to sort the plurality of sample document sets according to the difficulty scores of each of the sample document sets.

[0138] The backbone model training submodule is configured to train the preset backbone model in a rolling window curriculum manner based on the sorted plurality of sample document sets.

[0139] In one embodiment, the apparatus further comprises:

[0140] The difficulty score function defining module is configured to determine the multi-document difficulty scoring function based on a number of documents in a document set, a total length of the documents, and a document length distribution.

[0141] Each of the above modules in the power report generation apparatus based on the large model and the multi-document curriculum learning framework can be realized by software, hardware, and a combination thereof, in whole or in part. Each of the above modules can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so as to be called and executed by a processor to perform the operations corresponding to each of the above modules.

[0142] In one exemplary embodiment, a computer device, which can be a terminal, is provided, and an internal structure diagram of the computer device can be as follows:Figure 5 The computer device shown in the figure includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be realized through WIFI, mobile cellular network, Near Field Communication (NFC) or other technologies. The computer program is executed by the processor to implement a power report generation method based on a large model and a multi-document curriculum learning framework.

[0143] Those skilled in the art can understand that, Figure 5 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0144] In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the following steps:

[0145] In response to a question and answer operation on a power customer service platform, a power designated project query text of the question and answer operation is obtained;

[0146] A preset large language model is used to perform semantic understanding on the power designated project query text, and query key information is extracted. The query key information is used to indicate a target query range, a target query object, and a target report type;

[0147] According to the target query range and the target query object in the query key information, a plurality of power designated project associated files are retrieved;

[0148] A multi-document curriculum learning framework is used to integrate the plurality of power designated project associated files to generate concise file information;

[0149] According to the simplified file information and the target report type, a power designated item report is generated as a query result of the power designated item query text.

[0150] In one embodiment, the processor, when executing the computer program, also implements the steps of the power report generation method based on the large model and the multi-document curriculum learning framework in the other embodiments described above.

[0151] In one embodiment, a computer-readable storage medium is provided, and the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the following steps are implemented:

[0152] In response to a question and answer operation on a power customer service platform, a power designated item query text of the question and answer operation is obtained;

[0153] A preset large language model is used for semantic understanding of the power designated item query text, and query key information is extracted. The query key information is used to indicate a target query range, a target query object, and a target report type;

[0154] According to the target query range and the target query object in the query key information, a plurality of power designated item associated files are retrieved;

[0155] A multi-document curriculum learning framework is used to integrate the plurality of power designated item associated files, and simplified file information is generated;

[0156] According to the simplified file information and the target report type, a power designated item report is generated as a query result of the power designated item query text.

[0157] In one embodiment, the computer program, when executed by a processor, also implements the steps of the power report generation method based on the large model and the multi-document curriculum learning framework in the other embodiments described above.

[0158] In one embodiment, a computer program product is provided, and the computer program product includes a computer program. When the computer program is executed by a processor, the following steps are implemented:

[0159] In response to a question and answer operation on a power customer service platform, a power designated item query text of the question and answer operation is obtained;

[0160] A preset large language model is used for semantic understanding of the power designated item query text, and query key information is extracted. The query key information is used to indicate a target query range, a target query object, and a target report type;

[0161] According to the target query range and the target query object in the query key information, a plurality of power designated item associated files are retrieved;

[0162] The multiple power designated project associated files are integrated by using a multi-document course learning framework to generate a simplified file information;

[0163] According to the simplified file information and the target report type, a power designated project report is generated as a query result of the power designated project query text.

[0164] In one embodiment, the computer program, when executed by a processor, also implements the steps of the power report generation method based on a large model and a multi-document course learning framework in the other embodiments described above.

[0165] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0166] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. The volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., without being limited thereto.

[0167] The technical features of the above embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0168] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A power report generation method based on a large model and a multi-document curriculum learning framework, characterized by, The method comprises: In response to a question and answer operation of a power customer service platform, an electric power designated project query text of the question and answer operation is acquired; A preset large language model is used for semantic understanding of the electric power designated project query text, and query key information is obtained by classifying and arranging a plurality of keywords extracted for different report types; the query key information is used to indicate a target query range, a target query object, and a target report type; According to the target query range and the target query object in the query key information, a plurality of electric power designated project associated files are retrieved; A multi-document course learning framework is used to integrate the plurality of electric power designated project associated files to generate simplified file information; According to the simplified file information and the target report type, an electric power designated project report is generated as a query result of the electric power designated project query text; Further comprising: A few-shot fine-tuning is performed through prompting to construct training data and training prompts based on an electric power designated project query task; the training data and the training prompts are used to train the preset large language model; The method further comprises: A course learning strategy is applied to a multi-document summarization task, and a multi-document difficulty scoring function is combined to train a preset backbone model in a rolling window course manner to construct the multi-document course learning framework; Further comprising: Based on the number of documents, the total length of documents, and the length distribution of documents in a document set, the multi-document difficulty scoring function is determined; the multi-document difficulty scoring function is represented by the following formula: where D i is a multi-document collection, N i is the number of documents in the multi-document collection, L i is the total length of the documents in the multi-document collection, is the document length distribution of the multi-document collection, is the maximum document length in the multi-document collection, and the hyperparameter is 512.

2. The method of claim 1, wherein, The method further comprises: A basic large language model for model training is acquired, and the training data and the training prompts are used to adjust the model parameters of the basic large language model until a training end condition is met to obtain the preset large language model.

3. The method of claim 1, wherein, The construction of the training data and the training prompts based on the electric power designated project query task comprises: Based on the electric power designated project query task, sample electric power designated project query texts of different project types and sample query key information corresponding to each sample electric power designated project query text are acquired as the training data; the sample query key information has labeled labels; Based on the electric power designated project query task, execution instructions for a basic large language model, electric power project associated information, and input text indication information are designed as the training prompts.

4. The method of claim 1, wherein, The model training of the preset backbone model by applying the course learning strategy to the multi-document summarization task and combining the multi-document difficulty scoring function comprises: A plurality of sample document sets are acquired, and the difficulty scores of each sample document set are calculated using the multi-document difficulty scoring function; According to the difficulty scores of each sample document set, the plurality of sample document sets are sorted; Based on the sorted plurality of sample document sets, the preset backbone model is trained in a rolling window course manner. 5.A power report generation apparatus based on a large model and a multi-document curriculum learning framework, characterized by, The device comprises: A query text acquisition module is configured to acquire an electric power designated project query text of a question and answer operation of a power customer service platform in response to the question and answer operation; The key information extraction module is configured to perform semantic understanding on the power designated project query text by using a preset large language model, and obtain query key information by classifying and organizing a plurality of keywords extracted for different report types; the query key information is used to indicate a target query range, a target query object, and a target report type; The retrieval file obtaining module is configured to perform retrieval according to the target query range and the target query object in the query key information, and obtain a plurality of power designated project associated files; The simplified information integration module is configured to integrate the plurality of power designated project associated files by using a multi-document curriculum learning framework, and generate simplified file information; The power report generation module is configured to generate a power designated project report according to the simplified file information and the target report type, as a query result of the power designated project query text; The training information construction module is configured to perform few-shot fine-tuning through a prompt engineering, and construct training data and training prompt words based on a power designated project query task; the training data and the training prompt words are used to train the preset large language model; The device further comprises: The multi-document curriculum learning framework construction module is configured to apply a curriculum learning strategy to a multi-document summarization task, combine a multi-document difficulty scoring function, and construct the multi-document curriculum learning framework by performing model training on a preset backbone model in a rolling window curriculum manner. The difficulty scoring function definition module is configured to determine the multi-document difficulty scoring function based on a number of documents in a document set, a total length of the documents, and a document length distribution; the multi-document difficulty scoring function is expressed by the following formula: The processor executes the computer program to implement the steps of the method of any one of claims 1 to 4. The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 4. where D i is a multi-document collection, N i is the number of documents in the multi-document collection, L i is the total length of the documents in the multi-document collection, is the document length distribution of the multi-document collection, is the maximum document length in the multi-document collection, and the hyperparameter is 512. 6.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-5 when the computer program is executed by the processor. The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 4.

7. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 4.

8. A computer program product comprising a computer program, characterized in that, ​

Images