Large language model multi-field domain optimization method and system
By optimizing the large language model across multiple domains and using the SimHash algorithm and RAG technology for data cleaning and matching, a multi-domain recognition training model was constructed. This solved the problem of extended recognition response time in large language models, achieving efficient and accurate cross-domain text recognition and multimodal data fusion, thus improving the user experience.
Patent Information
- Application Number
- CN202511128806.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-13
- Publication Date
- 2025-11-18
AI Technical Summary
When using large language models, due to the limited computing power of the system, a large amount of text training is required to cover different languages, cultures and professional scenarios, which leads to a longer recognition response time and affects the user experience.
By acquiring user input text, performing preprocessing and keyword search, using the SimHash algorithm for multi-threaded synchronous retrieval, cleaning and filtering related text data, constructing a multi-field recognition training model and continuously optimizing it, and combining RAG technology and industry knowledge base for data matching, multi-modal data fusion is achieved.
It improves the response efficiency and accuracy of large language models, reduces data redundancy, enhances cross-domain text recognition capabilities, supports the fusion of multi-source heterogeneous data, adapts to changes in the language environment, and improves the user experience.
Smart Images
Figure CN120975243A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of large language model optimization technology, specifically to a method and system for multi-field optimization of large language models. Background Technology
[0002] In a mature language model system, to provide users with a better experience, various models are deployed together and dynamically called according to the task, taking advantage of the strengths of each model to make the multi-turn, multi-talk effect better.
[0003] In existing technologies, large language models require extensive text training to ensure they can cover different languages, cultures, and professional scenarios, and to guarantee the model has a broad knowledge base and contextual understanding capabilities. However, due to limited system computing power, even with pre-training, language recognition in text still requires continuous analysis and retrieval based on different domain environments such as language, culture, and professional scenarios. This increases the recognition response time of large language models to some extent, thus affecting the user's actual experience. In view of this, we propose a multi-domain optimization method and system for large language models. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for multi-field optimization of large language models to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-domain optimization method for a large language model, comprising the following steps:
[0006] S1. Obtain user input text: Capture user input text and preprocess the text content to obtain the retrieved text data;
[0007] S2. Text content extraction: Perform keyword search on the retrieved text data obtained in S1 to match relevant scenarios and fields;
[0008] S3. Establish search range: Perform multi-threaded synchronous search based on the scene and domain matched by the searched text data to obtain multi-domain matching data;
[0009] S4. Relevance Matching: The SimHash algorithm is used to calculate the relevance of the multi-field matching data obtained in S3, and the relevance text data with the highest similarity is matched.
[0010] S5. Data cleaning and filtering: Clean the relevant text data obtained in S4, remove data with too large field differences, and obtain matching value training data.
[0011] S6. Constructing the recognition model: Use the matching values obtained in S5 as the training data for the verification dataset, and use the retrieval text data extracted in S1 as the local dataset. Substitute the verification dataset and the local dataset into the model and establish a multi-field recognition training model.
[0012] S7. Iterative optimization: Continuously train and improve the large language model by using a multi-field recognition training model to obtain an optimized large language model.
[0013] Preferably, the preprocessing of text content in S1 includes standardizing and unifying the format, content, calculation unit, and language of the text, and the calculation unit adopts internationally accepted units to facilitate the preprocessing of user input text, ensure a unified format, and thus facilitate the subsequent model's recognition, calculation, and classification of text data.
[0014] Preferably, when performing keyword search in S2, a text retrieval tool is used to extract keywords from the retrieval text data obtained in S1, obtain relevant keywords from the user input text, and perform a full-domain search using the relevant keywords as the search object. By extracting keywords from the text data, the time spent searching the entire text data is reduced, the system's response efficiency is improved, and the number of subsequent search results obtained through keyword retrieval is reduced, which is more conducive to subsequent data cleaning and extraction.
[0015] Preferably, when matching scenarios and domains in S3, the matching order is determined sequentially by language category, technical field, and usage scenario as classification methods, and the judgment results are distinguished to obtain data of multiple categories. Each data item is equipped with a corresponding matching label, so that the language model can quickly identify the classification of text data based on these matching labels, thereby facilitating data cleaning and classification, and improving the response efficiency of the language model to these text data.
[0016] Preferably, in step S4, a list of related text data is generated during correlation matching. The data in this list are arranged from highest to lowest correlation, and text data with a correlation of less than 30% is automatically removed. The remaining text data is output as related text data. The correlation of the text data is digitally calculated using SimHash, which makes the correlation of the text data more accurate in judgment and calculation, improves the accuracy of the model's judgment of the data, removes data with very low correlation, solves the problem of data redundancy, reduces the computational burden on the model, and improves the model's running efficiency.
[0017] Preferably, in step S5, when cleaning and filtering the related text data, a multi-field recognition training model is used as the comparison target for calculation. Anomalies in the related text data are extracted and removed using the isolated forest algorithm. The remaining data is then filled using statistical imputation to remove data with excessive differences, so as to obtain text data with more balanced relevance. Through multi-dimensional differential calculation methods, it is ensured that anomalies in each dimension of the text data can be accurately extracted and deleted, avoiding the impact of anomalies on the accuracy of the model's judgment results on the text data, and improving the accuracy of the entire large language model's recognition and judgment of the text data.
[0018] Preferably, the verification dataset in S6 is stored in the database of the large language model, and the verification dataset is in a dynamic update state. When the multi-domain recognition training model is training, it extracts the verification dataset of the current time period as the collection target, and after the training is completed, it backs up the exported text data to the verification dataset for overwriting and saving. In this way, the content of the verification dataset, which serves as the reference text data for the multi-domain recognition training model, is always updated, so that the large language model that is subsequently upgraded and optimized can adapt to changes in the language environment more quickly, better recognize the text data input by the user, make more accurate judgments, and improve the overall user experience of the large language model.
[0019] A large language model multi-field optimization system is provided. The system includes a data capture module, a text conversion module, a text data retrieval module, a text calculation module, a data filtering module, and a model optimization module. The data capture module includes unified access to multimodal data such as text, audio, video, and images. The text conversion module includes heterogeneous data fusion to achieve unified fusion of multimodal data and ensure subsequent data recognition, judgment, and extraction.
[0020] Preferably, the data filtering module includes RAG technology integrated with an industry knowledge base to perform multi-domain matching of text data, and obtain standard reference objects for matching the correlation of text data, thereby enabling the entire large language model optimization system to make more accurate judgments on the matching correlation of text data and improve the use effect of the large language model optimization system.
[0021] Preferably, the model optimization module includes model building, model iteration and upgrading, and data processing. The model building is used to establish a multi-domain recognition training model and a large language optimization model. The model iteration and upgrading is used to update the large language model. The data processing includes collecting the acquired text data and the text data finally selected by the user, calculating the correlation result between the text data finally selected by the user and the text data obtained by the large language model, outputting the text data correlation result as a memory label, associating the memory label with the selected text data, and synchronously outputting it to the multi-domain recognition training model. This ensures that the entire large language model always maintains its text data judgment results during iteration and upgrading, and can track and process past judgments and selected text data to provide reference for selecting data for subsequent text data retrieval by other users, thereby improving the user's retrieval experience when using the large language model.
[0022] Compared with existing technologies, this invention provides a method and system for multi-domain optimization of large language models, which has the following beneficial effects:
[0023] 1. This large language model multi-domain optimization method and system significantly reduces the computational load of full-text retrieval by extracting keywords through text retrieval tools and avoids redundant data processing. Combined with a multi-threaded synchronous retrieval mechanism, it matches multi-domain data in parallel, shortening the response latency. The SimHash algorithm is used to perform digital correlation calculation on multi-domain data, automatically removing low-correlation data with similarity below 30% to reduce noise interference. This makes the correlation of text data more accurate in judgment and calculation, improving the model's accuracy in judging data. At the same time, it removes data with very low correlation, solves the problem of data redundancy, reduces the computational burden on the model, and improves the model's running efficiency.
[0024] 2. This large-scale language model multi-domain optimization method and system adds structured labels to the data through three-level classification: language category, technical field, and usage scenario. This enhances the model's ability to recognize cross-domain text, enabling the language model to quickly identify the classification of text data based on these matching labels. This facilitates data cleaning and classification, improves the language model's response efficiency to these text data, and verifies the dynamic updating of the dataset. After model training, old data is automatically overwritten, adapting to changes in the language environment in real time. Furthermore, it achieves a logical closed loop at the user feedback level, collecting the text data finally selected by the user, generating memory labels, and associating them with the training model to optimize subsequent recommendation results.
[0025] 3. This large language model multi-domain optimization method and system supports the fusion of multi-source heterogeneous data such as text, audio, video, and images, breaking down information barriers in multimodal environments, expanding application scenarios, and integrating RAG technology with industry knowledge bases. This provides high-precision domain matching for professional fields such as medicine and finance, ensuring that the output results conform to industry standards, and mandating the conversion of standardized units and internationally accepted units to avoid cross-domain data ambiguity. It also ensures the system's consistent understanding of the data, reduces differences in the large language model's understanding of the data, and minimizes the impact on the correct processing effect of the large language model's information data. Attached Figure Description
[0026] Figure 1 This is a schematic diagram of the multi-field optimization method of the present invention;
[0027] Figure 2 This is a schematic diagram of the multi-field optimization system of the present invention. Detailed Implementation
[0028] like Figures 1-2 As shown, this invention provides a technical solution: a multi-domain optimization method for large language models, comprising the following steps:
[0029] S1. Obtain user input text: Capture user input text and preprocess the text content to obtain the retrieved text data;
[0030] S2. Text content extraction: Perform keyword search on the retrieved text data obtained in S1 to match relevant scenarios and fields;
[0031] S3. Establish search range: Perform multi-threaded synchronous search based on the scene and domain matched by the searched text data to obtain multi-domain matching data;
[0032] S4. Relevance Matching: The SimHash algorithm is used to calculate the relevance of the multi-field matching data obtained in S3, and the relevance text data with the highest similarity is matched.
[0033] S5. Data cleaning and filtering: Clean the relevant text data obtained in S4, remove data with too large field differences, and obtain matching value training data.
[0034] S6. Constructing the recognition model: Use the matching values obtained in S5 as the training data for the verification dataset, and use the retrieval text data extracted in S1 as the local dataset. Substitute the verification dataset and the local dataset into the model and establish a multi-field recognition training model.
[0035] S7. Iterative optimization: Continuously train and improve the large language model by using a multi-field recognition training model to obtain an optimized large language model.
[0036] In one embodiment of the present invention, the preprocessing of text content in S1 includes standardizing and unifying the format, content, calculation unit, and language of the text. The calculation unit adopts internationally accepted units to facilitate the preprocessing of user input text and ensure a unified format, thereby facilitating the subsequent model's recognition, calculation, and classification of text data. Meanwhile, during keyword search in S2, a text retrieval tool is used to extract keywords from the retrieved text data obtained in S1, obtaining relevant keywords from the user input text. A full-domain search is then performed using these relevant keywords as the search object. By extracting keywords from the text data, the time spent on the entire text data retrieval is reduced, improving the system's response efficiency. Furthermore, keyword retrieval results in fewer subsequent search results, which is more conducive to subsequent data cleaning and extraction.
[0037] Furthermore, in S3, when matching scenarios and domains, the matching order is determined sequentially by language category, technical field, and usage scenario. The results are then differentiated to obtain data in multiple categories, each with a corresponding matching label. This allows the language model to quickly identify the classification of text data based on these labels, facilitating data cleaning and categorization, and improving the language model's response efficiency. Specifically, in S4, when performing correlation matching, a list of correlated text data is generated. The data in this list are arranged from highest to lowest correlation, and text data with a correlation of less than 30% is automatically removed. The remaining text data is output as correlated text data. SimHash is used to digitally calculate the correlation of the text data, making the correlation of the text data more accurate in judgment and calculation, improving the model's accuracy in judging the data. At the same time, data with very low correlation is removed, solving the problem of data redundancy, reducing the computational burden on the model, and improving the model's operating efficiency.
[0038] Furthermore, in S5, when cleaning and filtering related text data, the multi-domain recognition training model is used as the comparison target for calculation. Anomalies in the related text data are extracted and removed using the Isolation Forest algorithm. The remaining data is then statistically imputed to remove excessively differentiated data, resulting in more balanced text data. Through multi-dimensional differential calculation methods, it is ensured that anomalies in each dimension of the text data can be accurately extracted and deleted, preventing them from affecting the accuracy of the model's judgment of the text data. This improves the overall accuracy of the large language model's text recognition. Meanwhile, in S6, the validation dataset is stored in the large language model's database and is dynamically updated. During training, the multi-domain recognition training model extracts the validation dataset of the current time period as the collection target, and after training, the exported text data is backed up to the validation dataset for overwriting. This ensures that the validation dataset, serving as the reference text data for the multi-domain recognition training model, is constantly updated, allowing the large language model to adapt more quickly to changes in the language environment, better recognize user-input text data, and make more accurate judgments, thus improving the overall user experience of the large language model.
[0039] The large language model multi-field optimization system includes a data capture module, a text conversion module, a text data retrieval module, a text calculation module, a data filtering module, and a model optimization module. The data capture module includes unified access to multimodal data such as text, audio, video, and images. The text conversion module includes heterogeneous data fusion to achieve unified fusion of multimodal data, ensuring the subsequent recognition, judgment, and extraction of data.
[0040] In this invention, the data filtering module includes RAG technology that integrates industry knowledge base to perform multi-domain matching on text data, and obtains standard reference objects for matching the correlation of text data. This enables the entire large language model optimization system to make more accurate judgments on the matching correlation of text data, thereby improving the effectiveness of the large language model optimization system.
[0041] It is worth noting that the model optimization module includes model building and iterative upgrades, as well as data processing. Model building is used to establish a multi-domain recognition training model and a large language optimization model. Model iterative upgrades are used to update the large language model. Data processing includes collecting the acquired text data and the text data finally selected by the user, calculating the correlation between the text data finally selected by the user and the text data obtained by the large language model, outputting the text data correlation result as a memory label, associating the memory label with the selected text data, and synchronously outputting it to the multi-domain recognition training model. This ensures that the entire large language model always maintains its text data judgment results during iterative upgrades, and can track and process past judgments and selected text data to provide reference for selecting data for subsequent text data retrieval by other users, thereby improving the user's retrieval experience when using the large language model.
[0042] The present invention has been described in detail above. However, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, any modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.
Claims
1. A multi-domain optimization method for large language models, characterized in that: Includes the following steps: S1. Obtain user input text: Capture user input text and preprocess the text content to obtain the retrieved text data; S2. Text content extraction: Perform keyword search on the retrieved text data obtained in S1 to match relevant scenarios and fields; S3. Establish search range: Perform multi-threaded synchronous search based on the scene and domain matched by the searched text data to obtain multi-domain matching data; S4. Relevance Matching: The SimHash algorithm is used to calculate the relevance of the multi-field matching data obtained in S3, and the relevance text data with the highest similarity is matched. S5. Data cleaning and filtering: Clean the relevant text data obtained in S4, remove data with too large field differences, and obtain matching value training data. S6. Constructing the recognition model: Use the matching values obtained in S5 as the training data for the verification dataset, and use the retrieval text data extracted in S1 as the local dataset. Substitute the verification dataset and the local dataset into the model and establish a multi-field recognition training model. S7. Iterative optimization: Continuously train and improve the large language model by using a multi-field recognition training model to obtain an optimized large language model.
2. The method for multi-field optimization of a large language model according to claim 1, characterized in that: The preprocessing of text content in S1 includes standardizing and unifying the text format, content, unit of calculation, and language, and the unit of calculation adopts internationally accepted units.
3. The method for multi-field optimization of a large language model according to claim 1, characterized in that: When performing keyword search in S2, a text retrieval tool is used to extract keywords from the retrieval text data obtained in S1, obtain relevant keywords from the user input text, and perform a full-domain search using the relevant keywords as the retrieval object.
4. The method for multi-field optimization of a large language model according to claim 1, characterized in that: When matching scenarios and domains in S3, the matching order is judged sequentially according to language category, technical field, and usage scenario as classification and differentiation methods, and the judgment results are distinguished to obtain data of multiple categories, and each data item is equipped with a corresponding matching label.
5. The method for multi-field optimization of a large language model according to claim 1, characterized in that: In step S4, when performing correlation matching, a list of correlated text data is generated. The correlation of the data in the correlated text data list is arranged from high to low, and text data with a correlation of less than 30% is automatically removed. The remaining text data is output as correlated text data.
6. The method for multi-field optimization of a large language model according to claim 1, characterized in that: In step S5, when cleaning and filtering related text data, a multi-field recognition training model is used as the comparison target for calculation. Abnormal data in the related text data is extracted and removed using the isolated forest algorithm. The remaining data is then filled using statistical imputation to remove data with excessive differences, so as to obtain text data with more balanced relevance.
7. The method for multi-field optimization of a large language model according to claim 1, characterized in that: The validation dataset in S6 is stored in the database of the large language model and is in a dynamic update state. The multi-field recognition training model extracts the validation dataset of the current time period as the collection target during training, and the exported text data is backed up to the validation dataset for overwriting after training.
8. A multi-field optimization system for a large language model, characterized in that: The system includes a data capture module, a text conversion module, a text data retrieval module, a text calculation module, a data filtering module, and a model optimization module. The data capture module includes unified access to multimodal data such as text, audio, video, and images. The text conversion module includes heterogeneous data fusion.
9. The large language model multi-field optimization system according to claim 8, characterized in that: The data filtering module includes RAG technology that integrates industry knowledge base to perform multi-domain matching on text data and obtain standard reference objects for matching the correlation of text data.
10. The large language model multi-field optimization system according to claim 8, characterized in that: The model optimization module includes model building, model iteration and upgrading, and data processing. The model building is used to establish a multi-domain recognition training model and a large language optimization model. The model iteration and upgrading is used to update the large language model. The data processing includes collecting the acquired text data and the text data finally selected by the user, calculating the correlation result between the text data finally selected by the user and the text data obtained by the large language model, outputting the text data correlation result as a memory label, associating the memory label with the selected text data, and synchronously outputting it to the multi-domain recognition training model.