A training method for cultural relics question answering model based on large language model
Through the cultural relics question-and-answer model training method based on the large language model, the problem of difficulty in obtaining historical relics information is solved, and fast and accurate information acquisition and question-and-answer model optimization is achieved.
Patent Information
- Application Number
- CN202411828263.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-04-04
AI Technical Summary
The prior art is difficult to provide information related to historical relics quickly and effectively, making it difficult for enthusiasts in related fields to quickly obtain the required information.
The cultural relics question-and-answer model training method based on the large language model, by collecting, processing and classifying cultural relics information, building a training model for pre-training, extracting keywords, establishing a pre-set question bank, performing model dialogue correction and optimizing databases, and forming standardized answers.
The ability to quickly obtain historical cultural relics information is realized, and the optimization of the question-and-answer model is accelerated through user feedback, and the accuracy of output information is improved.
Smart Images
Figure CN119760075A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the technical field of question-answering model training, and relates to a cultural relic question-answering model training method based on a large language model. Background Art
[0002] Large language models refer to machine learning models trained with large amounts of text data. They can generate and understand natural language. Currently, the most widely used large language models include GPT, BERT, XLNet, etc. For ease of use, developers usually perform specific training based on large language models to build dialogue models that suit their industry.
[0003] my country is a country with a long history and a large number of historical relics. Enthusiasts in related fields learn the historical information contained in cultural relics and discover their cultural value by interpreting them. Due to the huge amount of cultural relic data, it is necessary to establish a dialogue model in the direction of cultural relic questions and answers so that enthusiasts in related fields can obtain the information they want to know more quickly.
[0004] Application Contents
[0005] In response to the above problems, this application proposes a cultural relic question-answering model training method based on a large language model, which effectively solves the problems in the prior art.
[0006] In order to achieve the above purpose, the technical solution adopted in this application is as follows:
[0007] A method for training a cultural relic question-answering model based on a large language model, comprising:
[0008] Collect cultural relics information and establish a cultural relics information database;
[0009] Cultural relics information processing: removing duplication and junk information from cultural relics information obtained from multiple channels, and filtering out duplicate information and junk information;
[0010] Classification of cultural relics information: Classification of cultural relics information into various categories based on the database;
[0011] Construct a training model, connect the processed cultural relic information database to the existing large language model, pre-train the cultural relic information data, and obtain a pre-training model;
[0012] Keyword extraction: extracting keywords for cultural relic classification from cultural relic information classification to form a pre-search formula for cultural relic search;
[0013] Preset question library: create preset questions based on pre-search keyword information and build a preset question library;
[0014] Model dialogue correction: ask the training model questions through preset questions to correct and optimize the answers output by the training model;
[0015] Establish an optimized database, organize and collect the corrected and optimized answers as standardized answers to relevant questions.
[0016] Optionally, also include,
[0017] Feedback information collection: establish a feedback information collection database to collect feedback information from users during use;
[0018] Feedback information is filtered to extract information that is helpful for database update.
[0019] Feedback information distribution: after classifying user feedback information, distribute it to relevant professionals for answering and processing;
[0020] Positive information is stored in the database, and the positive information in the feedback information is summarized and updated to the cultural relics information database.
[0021] Optionally, feedback information collection includes dialogue feedback information and module feedback information.
[0022] Dialogue feedback information, that is, the dialogue fragments used to raise questions during the dialogue between the user and the model are intercepted;
[0023] Module feedback information refers to the information collected in the module dedicated to problem feedback established in the model.
[0024] Optional, high-frequency information warning, extract keywords from the information collected in the feedback information collection, prioritize them according to the frequency of keyword occurrence, issue warnings for the high-frequency parts, and process them first.
[0025] Optionally, the collection of cultural relics information includes text information and picture information; it also includes,
[0026] Graphic information analysis, text analysis of cultural relics picture information;
[0027] Graphic information reading, key parts of the pictures provided by users are intercepted and analyzed, and the pictures are fitted with the graphic information analysis content in the cultural relics information database, and the answers with a high degree of fit are output.
[0028] Optionally, also include,
[0029] User confirmation: ask the user to confirm the accuracy of the output answer. After receiving negative feedback, the next priority answer is output;
[0030] Retrieve the image archives, and after getting positive feedback from users, parse the images used in the user search and archive them as a control group for retrieval.
[0031] Optionally, a database update co-creation module is established to allow users to upload cultural relics information after the database update deadline through the co-creation module, and the information is updated to the database after screening.
[0032] Optionally, a points mechanism can be established to give basic points to users who participate in updating the co-creation module database, to reward points to users who provide valid information, to deduct points from users who provide false information, and to stipulate corresponding rewards and penalties.
[0033] Compared with the prior art, this application has the following beneficial effects:
[0034] 1. Based on the large language model and training mechanism in the existing technology, by accessing the database of cultural relics information, a question-answering model in the field of historical cultural relics is established, so that enthusiasts in related fields can obtain the information they want to know more quickly;
[0035] 2. In the process of improving the question-answering model, a large amount of data correction and optimization is required. By establishing feedback, users can participate in the training process of the question-answering model during use, thus accelerating the optimization process of the question-answering model;
[0036] 3. When a keyword appears frequently in the feedback information, it means that the part is highly controversial or has more interested users. Prioritizing the part corresponding to the keyword will help improve the accuracy of the information output by the question-answering model. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 It is a schematic diagram of the process structure of an embodiment of the present application. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0039] See also Figure 1 , which is a cultural relic question-answering model training method based on a large language model disclosed in an embodiment of the present application, including collecting cultural relic information and establishing a cultural relic information database;
[0040] Cultural relics information processing: removing duplication and junk information from cultural relics information obtained from multiple channels, and filtering out duplicate information and junk information;
[0041] Classification of cultural relics information: Classification of cultural relics information into various categories based on the database;
[0042] Construct a training model, connect the processed cultural relic information database to the existing large language model, pre-train the cultural relic information data, and obtain a pre-training model;
[0043] Keyword extraction: extracting keywords for cultural relic classification from cultural relic information classification to form a pre-search formula for cultural relic search;
[0044] Preset question library: create preset questions based on pre-search keyword information and build a preset question library;
[0045] Model dialogue correction: ask the training model questions through preset questions to correct and optimize the answers output by the training model;
[0046] Establish an optimized database, organize and collect the corrected and optimized answers as standardized answers to relevant questions.
[0047] In this way, based on the large language model and training mechanism in the existing technology, by accessing the database of cultural relics information, a question-answering model in the field of historical cultural relics is established, so that enthusiasts in related fields can obtain the information they want to know more quickly.
[0048] Specifically, the collection of cultural relic information includes but is not limited to the Internet, documents and museum databases. The repeated parts of the cultural relic information obtained from multiple channels are merged and removed to reduce the amount of data. The erroneous information that is in the guessing stage or has been corrected is removed to ensure the authority of the interpretation of cultural relic information relative to current historical materials, and the collected information is classified in various ways.
[0049] The classification methods include but are not limited to historical era, cultural relic category, excavation time and region, and collection location, etc., and cultural relic information should be classified into multiple categories based on the central classification method, and the classified categories should be extracted as keywords to form a keyword search formula belonging to the cultural relic. During the dialogue process, the keywords provided by the user and the keywords of the cultural relic are used to search word by word, thereby reducing the amount of search data and improving the search speed and accuracy;
[0050] Preset questions are created based on keyword information, and the cultural relics question-and-answer model is pre-trained with the help of a large language model through the preset questions. During the pre-training process, the answers output by the pre-trained model are corrected and optimized, and an optimized database is established. When user questions are close to preset questions, the optimized database is called to improve the speed and accuracy of information extraction.
[0051] As a specific implementation method of a cultural relic question-answering model training method based on a large language model provided in the application, it also includes feedback information collection, establishing a feedback information collection database, and collecting feedback information from users during use;
[0052] Feedback information is filtered to extract information that is helpful for database update.
[0053] Feedback information distribution: after classifying user feedback information, distribute it to relevant professionals for answering and processing;
[0054] Positive information is stored in the database, and the positive information in the feedback information is summarized and updated to the cultural relics information database.
[0055] In general, the process of improving the question-answering model requires a large amount of data correction and optimization. By establishing feedback, users can participate in the training process of the question-answering model during use, thereby accelerating the optimization process of the question-answering model.
[0056] Specifically, user feedback information is multi-faceted and may include issues such as historical knowledge corrections or model operation. Feedback related to cultural relics is screened, classified by keywords, and summarized and submitted to relevant professional personnel for answers and updates.
[0057] Furthermore, feedback information collection includes dialogue feedback information and module feedback information.
[0058] Dialogue feedback information, that is, the dialogue fragments used to raise questions during the dialogue between the user and the model are intercepted;
[0059] Module feedback information refers to the information collected in the module dedicated to problem feedback established in the model.
[0060] Depending on the specific usage scenario, user feedback information may be provided in a specially designed feedback module or may appear in a conversation with the question-and-answer model. By intercepting some negative and questioning keywords in the conversation and intercepting conversation segments involving the keywords, more feedback information can be collected to accelerate the process of question-and-answer model training and optimization.
[0061] Furthermore, high-frequency information warning is used to extract keywords from the information collected in the feedback information collection, and priority is sorted according to the frequency of occurrence of keywords. The parts that appear frequently are warned and processed first.
[0062] In combination with specific usage scenarios, when a keyword appears frequently in the feedback information, it means that the part is more controversial or has more interested users. Prioritizing the part corresponding to the keyword can help improve the accuracy of the information output by the question-answering model.
[0063] As a specific implementation method of a cultural relic question-answering model training method based on a large language model provided in the application, the cultural relic information collected includes text information, picture information, and also includes
[0064] Graphic information analysis, text analysis of cultural relics pictures;
[0065] Graphic information reading, key parts of the pictures provided by users are intercepted and analyzed, and the pictures are fitted with the graphic information analysis content in the cultural relics information database, and the answers with a high degree of fit are output.
[0066] It should be understood that in the cultural relic question-and-answer model, there is not only the retrieval of text questions, but also the retrieval of pictures of certain historical cultural relics. By supplementing picture information with the collection of cultural relic information, it is easier to analyze and communicate with the pictures.
[0067] Specifically, there are a variety of information in the cultural relics picture information, including but not limited to shape, material, pattern and text, etc. By reading and fitting the information, the cultural relics involved in the keywords in the database are matched, and the cultural relics information with a high degree of similarity to the read information is selected for output.
[0068] In some feasible ways, collecting cultural relic information may also include three-dimensional modeling information of the cultural relic.
[0069] Furthermore, it also includes user confirmation, asking the user to confirm the accuracy of the output answer, and outputting the next priority answer after receiving negative feedback;
[0070] Retrieve the image archives, and after getting positive feedback from users, parse the images used in the user search and archive them as a control group for retrieval.
[0071] It should be understood that due to the huge amount of cultural relics data and the clarity of the pictures, there may be certain errors when reading information. By setting multiple possibilities and establishing priorities according to the degree of matching, users are allowed to participate in confirmation through output one by one. After receiving positive feedback, the retrieved pictures are archived. After being confirmed as valid information by professionals, they become a reference picture together with the model pictures in the original database, so that other users can output more accurate answers when searching.
[0072] As a specific implementation method of a cultural relic question-and-answer model training method based on a large language model provided in the application, a database update co-creation module is established, allowing users to upload cultural relic information after the database update deadline through the co-creation module, and update the information to the database after screening.
[0073] It should be understood that the interpretation of cultural relic information will change as more historical materials are added, and the question-answering model is limited by the database update time and cannot update the data in real time. After the co-creation module is established, targeted updates can be made through the historical materials provided by users to increase the speed of information updating. At the same time, a temporary database can also be established through the co-creation module. When the user retrieves information beyond the database update deadline, inaccurate information can be output through the temporary database of the co-creation module and inaccurate information prompts can be given.
[0074] Furthermore, a points mechanism is established to give basic points to users who participate in the update of the co-creation module database, to reward points to users who provide valid information, to deduct points from users who provide false information, and to stipulate corresponding rewards and punishments.
[0075] It should be understood that by setting up a point reward mechanism, the provision of false information can be reduced to a certain extent, the time for information screening can be reduced, and at the same time, rewards can be given to users who actively participate in information updates, thereby increasing their enthusiasm for participation and speeding up model training and optimization updates.
[0076] Although the present application has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.
Claims
1. A method for training a cultural relic question-answering model based on a large language model, characterized in that: include: Collect cultural relics information and establish a cultural relics information database; Cultural relics information processing: removing duplication and junk information from cultural relics information obtained from multiple channels, and filtering out duplicate information and junk information; Classification of cultural relics information: Classification of cultural relics information into various categories based on the database; Construct a training model, connect the processed cultural relic information database to the existing large language model, pre-train the cultural relic information data, and obtain a pre-training model; Keyword extraction: extracting keywords for cultural relic classification from cultural relic information classification to form a pre-search formula for cultural relic search; Preset question library: create preset questions based on pre-search keyword information and build a preset question library; Model dialogue correction: ask the training model questions through preset questions to correct and optimize the answers output by the training model; Establish an optimized database, organize and collect the corrected and optimized answers as standardized answers to relevant questions.
2. The method for training a cultural relic question-answering model based on a large language model according to claim 1, characterized in that: Also includes Feedback information collection: establish a feedback information collection database to collect feedback information from users during use; Feedback information is filtered to extract information that is helpful for database update. Feedback information distribution: after classifying user feedback information, distribute it to relevant professionals for answering and processing; Positive information is stored in the database, and the positive information in the feedback information is summarized and updated to the cultural relics information database.
3. The method for training a cultural relic question-answering model based on a large language model according to claim 2, characterized in that: Feedback information collection includes dialogue feedback information and module feedback information. Dialogue feedback information, that is, the dialogue fragments used to raise questions during the dialogue between the user and the model are intercepted; Module feedback information refers to the information collected in the module dedicated to problem feedback established in the model.
4. The method for training a cultural relic question-answering model based on a large language model according to claim 3, characterized in that: High-frequency information warning: extract keywords from the information collected in the feedback information collection, prioritize them according to the frequency of keyword occurrence, and issue warnings for the high-frequency parts, giving priority to processing them.
5. The method for training a cultural relic question-answering model based on a large language model according to claim 1, characterized in that: Collect cultural relics information including text information and picture information; include, Graphic information analysis, text analysis of cultural relics picture information; Graphic information reading, key parts of the pictures provided by users are intercepted and analyzed, and the pictures are fitted with the graphic information analysis content in the cultural relics information database, and the answers with a high degree of fit are output.
6. The method for training a cultural relic question-answering model based on a large language model according to claim 5, characterized in that: Also includes User confirmation: ask the user to confirm the accuracy of the output answer. After receiving negative feedback, the next priority answer is output; Retrieve the image archives, and after getting positive feedback from users, parse the images used in the user search and archive them as a control group for retrieval.
7. The method for training a cultural relic question-answering model based on a large language model according to claim 1, characterized in that: Establish a database update co-creation module to enable users to upload cultural relics information after the database update deadline through the co-creation module, and update the information to the database after screening.
8. The method for training a cultural relic question-answering model based on a large language model according to claim 7, characterized in that: A points mechanism will be established to give basic points to users who participate in updating the co-creation module database, to reward points to users who provide valid information, to deduct points from users who provide false information, and to stipulate corresponding rewards and punishments.