News recommendation method and system based on big language model ensemble learning
Through the news recommendation method based on the integrated learning of large language models, using three large language models with different parameter scales to collaborate recommendations, the problem of difficult to achieve balance of news recommendation accuracy and efficiency in the existing technology is solved, and more efficient and accurate news recommendation results are achieved.
Patent Information
- Application Number
- CN202510313309.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-17
- Publication Date
- 2025-06-13
AI Technical Summary
The accuracy of existing news recommendation using a single large language model needs to be improved, and the balance between recommendation efficiency and accuracy is difficult to achieve.
Using a news recommendation method based on integrated learning of large language models, we use three large language models with different parameter scales to collaborate recommendations by collecting user data and constructing prompt instructions for natural language representation. The specific steps include: two large models with smaller parameter scales are deployed in parallel for preliminary recommendations, and when the recommendation results are inconsistent, the final decision is made by the third model with larger parameter scales.
Through a two-step recommendation strategy, the accuracy of news recommendations is improved, and the recommendation efficiency is improved while ensuring accuracy, achieving resource conservation and efficiency improvement. At the same time, the advantages of large models are used to improve the overall performance of the recommendation system.
Smart Images
Figure CN120144873A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and particularly relates to a news recommendation method and system based on the integrated learning of large language models. Background Art
[0002] With the growth of personalized needs and the exacerbation of the problem of information overload, the importance of news recommendation systems (NRS) has become increasingly prominent. It not only helps users screen out interesting news content from a vast amount of information, but also, for news platforms, is a key technology to enhance the user experience and increase user stickiness. These systems analyze users' behaviors and preferences to provide personalized news content, thereby enhancing the user experience and satisfaction. The research and application of news recommendation systems play a crucial role and have made remarkable progress.
[0003] The main task of news recommendation systems is to screen out articles that users may be interested in from a vast number of candidate news. Currently, the research on news recommendation systems mainly includes: 1) content-based recommendation: This method matches according to users' historical browsing records and the content features of news, and recommends similar or relevant news to users; 2) collaborative filtering-based recommendation: Analyze users' historical behaviors to discover users' interest preferences, and recommend news that similar users are interested in to users based on the similarity between users; 3) deep learning-based recommendation: Use deep learning models to model users' historical behaviors and the content features of news to improve the recommendation effect. The research and application of news recommendation systems are constantly progressing, but also face new challenges and opportunities.
[0004] The performance of news recommendation depends to a large extent on the understanding of news content and the modeling of users' interests. Large language models have shown the ability to handle various natural language tasks. The latest news recommendation systems explore how to combine the latest artificial intelligence technologies and use large language model technologies to improve the performance of news recommendation systems and provide more accurate and personalized news recommendation services. For example, combining the deep semantic understanding and extensive pre-trained knowledge of large language models, the ONCE method for enhancing content-based recommendation is proposed; the Prompt4NR framework including discrete templates, continuous templates, and hybrid templates and other prompt templates; the PBNR model that regards personalized news recommendation as a text-to-text language task and designs personalized prompts; the research on large language model news recommendation technology based on hybrid prompt engineering, etc. Although the application of large language models in news recommendation has made certain progress, there are still some problems and challenges, such as the balance between the recommendation accuracy and efficiency of large models and the recommendation accuracy, which may limit them in specific tasks. Summary of the Invention
[0005] To this end, the present invention provides a news recommendation method and system based on the integrated learning of large language models, which solves the problems such as the need to improve the accuracy in existing news recommendation using a single large language model.
[0006] According to the design scheme provided by the present invention, on the one hand, a news recommendation method based on the integrated learning of large language models is provided, including:
[0007] Collect data of users in the news system, and screen out a number of feature data from the data, where the feature data includes: user feature data and news feature data that are most relevant to the news browsed by users;
[0008] Use the feature data to construct a prompt instruction represented in natural language;
[0009] Input the prompt instruction into the first news recommendation model and the second news recommendation model, and obtain the news recommendation results output by both the first news recommendation model and the second news recommendation model. Among them, the first news recommendation model is obtained by fine-tuning the pre-trained first large language model, and the second news recommendation model is obtained by fine-tuning the pre-trained second large language model;
[0010] If the news recommendation results output by both the first news recommendation model and the second news recommendation model are the same, recommend news to the user according to the news recommendation results output by both of them. If the news recommendation results output by both the first news recommendation model and the second news recommendation model are different, input the prompt instruction and the news recommendation results output by both of them into the third news recommendation model, so as to recommend news to the user according to the news recommendation results output by the third news recommendation model. The third news recommendation model is a pre-trained third large language model.
[0011] As the news recommendation method based on the integrated learning of large language models of the present invention, further, the first large language model and the second large language model adopt heterogeneous large language models, and the parameter scale of the third large language model is respectively larger than the parameter scale of the first large language model and the parameter scale of the second large language model.
[0012] As the news recommendation method based on the integrated learning of large language models of the present invention, further, the process of fine-tuning the pre-trained first large language model or the pre-trained second large language model includes:
[0013] In the pre-training stage, collect text data in related fields and preprocess the data, and use the preprocessed text data to pre-train the first large language model or the second large language model, so that the first large language model or the second large language model learns the text data rules, and the text data rules include: grammatical structure and semantic relationship;
[0014] In the fine-tuning stage, sample data for fine-tuning training is constructed based on the public news recommendation dataset, and the pre-trained first large language model or second large language model is fine-tuned using the sample data for fine-tuning training. The sample data for fine-tuning training includes: positive sample data and negative sample data. Among them, both the positive sample data and the negative sample data are composed of news features for describing news text features and user behavior features for describing user behavior log features, and the positive sample data and the negative sample data are distinguished according to user behavior features.
[0015] As the news recommendation method based on large language model ensemble learning of the present invention, further, fine-tuning the pre-trained first large language model or second large language model using the sample data for fine-tuning training includes:
[0016] Construct prompt instructions represented in natural language based on the positive sample data and the negative sample data respectively, and obtain an instruction fine-tuning dataset;
[0017] Use the instruction fine-tuning dataset to fine-tune and train the pre-trained first large language model or second large language model.
[0018] As the news recommendation method based on large language model ensemble learning of the present invention, further, fine-tuning and training the pre-trained first large language model or second large language model using the instruction fine-tuning dataset includes:
[0019] Use the low-rank decomposition matrix in the LoRA fine-tuning method to fine-tune and train the pre-trained first large language model or second large language model.
[0020] As the news recommendation method based on large language model ensemble learning of the present invention, further, pre-training the third large language model includes:
[0021] Collect text data in related fields and preprocess the data, and use the preprocessed text data to pre-train the third large language model so that the third large language model learns the text data rules. The text data rules include: grammar structure and semantic relationship. As the news recommendation method based on large language model ensemble learning of the present invention, further, the third news recommendation model is obtained by fine-tuning the pre-trained third large language model, and its fine-tuning process includes:
[0022] Construct sample data for fine-tuning training based on the public news recommendation dataset. The sample data for fine-tuning training includes: positive sample data and negative sample data. Among them, both the positive sample data and the negative sample data are composed of news text content, user historical behavior logs, news features for describing news text content features, and user behavior features for describing user historical behavior log features, and the positive sample data and the negative sample data are distinguished according to user behavior features;
[0023] Construct prompt instructions in natural language based on positive sample data and negative sample data respectively, and obtain an instruction fine-tuning dataset;
[0024] Based on the low-rank decomposition matrix in the LoRA fine-tuning method and using the instruction fine-tuning dataset to fine-tune and train the pre-trained third large language model, so that the third large language model makes news recommendations based on user historical behavior characteristics, news text context information, user characteristics and news characteristics.
[0025] On the other hand, the present invention also provides a news recommendation system based on large language model ensemble learning, including: a data collection module, an instruction construction module, an initial recommendation module and a recommendation output module, wherein,
[0026] The data collection module is used to collect data of users in the news system and screen out several user feature data from the data, and the user feature data is the feature data most relevant to the news browsed by the user;
[0027] The instruction construction module is used to construct prompt instructions in natural language by using user feature data;
[0028] The initial recommendation module is used to input the prompt instructions into the first news recommendation model and the second news recommendation model, and obtain the news recommendation results output by both the first news recommendation model and the second news recommendation model. Among them, the first news recommendation model is obtained by fine-tuning the pre-trained first large language model, and the second news recommendation model is obtained by fine-tuning the pre-trained second large language model;
[0029] The recommendation output module is used to recommend news to the user according to the news recommendation results output by both when the news recommendation results output by both the first news recommendation model and the second news recommendation model are consistent, and when the news recommendation results output by both the first news recommendation model and the second news recommendation model are inconsistent, input the prompt instructions into the third news recommendation model to recommend news to the user according to the news recommendation results output by the third news recommendation model, and the third news recommendation model is the pre-trained third large language model.
[0030] The beneficial effects of the present invention:
[0031] The present invention selects two large models with relatively small parameter scales and fine-tunes them using recommendation data, and then deploys these two fine-tuned large models in parallel for preliminary recommendations. When the recommendation results are inconsistent, a larger model with a larger parameter scale makes the final decision, improving the accuracy of news recommendations through a two-step recommendation strategy, saving the resources of the news recommendation system and improving efficiency while ensuring the recommendation accuracy. The experimental results show that the solution of this case can effectively save computing resources and significantly improve the response speed of the system by using two large models with relatively small parameter scales to process recommendation tasks in parallel. It not only reduces resource consumption but also improves the user experience, and is applicable to news recommendation systems that need to process a large number of user requests; when the recommendation results of the two small models are inconsistent, a larger model with more parameters is used to make the final decision, which can improve the recommendation accuracy and utilize the advantages of large models in processing complex language and user behavior data to improve the overall performance of the recommendation system. Description of the Drawings
[0032] Figure 1 Schematic diagram of the news recommendation process based on large language model ensemble learning in the embodiment;
[0033] Figure 2 Schematic diagram of the principle of the news recommendation algorithm of large language model ensemble learning in the embodiment;
[0034] Figure 3 Schematic diagram of the data format for fine-tuning the large model in the embodiment. Detailed Embodiment
[0035] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the drawings and technical solutions.
[0036] Although individual large models perform well in news recommendations, they may have limitations in specific tasks. The solution of this case attempts to fine-tune the models and explore the cooperation strategies between large models, with the expectation of giving full play to the advantages of each model through cooperation and overcoming their deficiencies at the same time. Refer to Figure 1 As shown, an embodiment of the present invention provides a news recommendation method based on large language model ensemble learning, including:
[0037] S101. Collect data of users in the news system and screen out several feature data from the data, where the feature data includes: user feature data and news feature data most relevant to the news browsed by users;
[0038] S102. Use the feature data to construct a prompt instruction represented in natural language;
[0039] S103. Input the prompt instruction into the first news recommendation model and the second news recommendation model, and obtain the news recommendation results output by both the first news recommendation model and the second news recommendation model. Among them, the first news recommendation model is obtained by fine-tuning a pre-trained first large language model, and the second news recommendation model is obtained by fine-tuning a pre-trained second large language model;
[0040] S104. If the news recommendation results output by both the first news recommendation model and the second news recommendation model are the same, recommend news to the user based on the news recommendation results output by both of them. If the news recommendation results output by both the first news recommendation model and the second news recommendation model are different, input the prompt instruction and the news recommendation results output by both of them into the third news recommendation model, so as to recommend news to the user based on the news recommendation results output by the third news recommendation model. The third news recommendation model is a pre-trained third large language model.
[0041] As an important branch of machine learning, ensemble learning improves the performance and generalization ability of the model by combining multiple different learners. The core idea of this method is "two heads are better than one", that is, by combining multiple weak models to overcome the task limitation problems existing in a single strong model. Ensemble learning can effectively reduce the bias and variance of a single model by combining the prediction results of multiple models, thereby improving the prediction accuracy; it performs more stably when dealing with noise and outliers, improving the system robustness; ensemble learning can better handle unseen data through the combination of different base learners and has better generalization ability. In the embodiments of this case, as Figure 2 shown, on the basis of fine-tuning the large model, the ensemble learning technology is used to improve the accuracy and efficiency of news recommendation.
[0042] Among them, the first large language model and the second large language model adopt heterogeneous large language models, and the parameter scale of the third large language model is respectively larger than the parameter scale of the first large language model and the parameter scale of the second large language model.
[0043] Specifically, models with relatively small parameter scales such as the DeepSeek-R1-Distill-Llama-8B model and the Meta-Llama-3-8B-Instruct model can be selected as the first large language model and the second large language model respectively, and a model with a larger parameter scale such as DeepSeek-V3 based on the mixture of experts architecture can be selected as the third large language model, so as to improve the news recommendation accuracy and robustness through a reasonable model integration strategy.
[0044] Among them, the process of fine-tuning the pre-trained first large language model or the pre-trained second large language model can be designed to include:
[0045] In the pre-training stage, relevant domain text data is collected and preprocessed. The preprocessed text data is used to pre-train the first large language model or the second large language model, so that the first large language model or the second large language model learns the text data patterns, and the text data patterns include: grammatical structures and semantic relationships;
[0046] In the fine-tuning stage, sample data for fine-tuning training is constructed based on the publicly available news recommendation dataset. The pre-trained first large language model or second large language model is fine-tuned using the sample data for fine-tuning training. The sample data for fine-tuning training includes: positive sample data and negative sample data. Among them, both the positive sample data and the negative sample data are composed of news features for describing news text features and user behavior features for describing user behavior log features, and the positive sample data and the negative sample data are distinguished according to the user behavior features.
[0047] Among them, prompt instructions represented in natural language are constructed based on the positive sample data and the negative sample data respectively, and an instruction fine-tuning dataset is obtained; the pre-trained first large language model or second large language model is fine-tuned using the instruction fine-tuning dataset. The low-rank decomposition matrix in the LoRA fine-tuning method can be used to fine-tune the pre-trained first large language model or second large language model.
[0048] The pre-trained third large language model can use an open-source large-scale pre-trained large language model.
[0049] Alternatively, pre-train the third large language model using domain-related data. By collecting relevant domain text data and preprocessing the data, use the preprocessed text data to pre-train the third large language model so that the third large language model learns the text data patterns, where the text data patterns include: grammatical structures and semantic relationships. Among them, to improve the model accuracy and recommendation effect, the third news recommendation model can also be obtained by fine-tuning the pre-trained third large language model. The fine-tuning process includes: constructing sample data for fine-tuning training based on a public news recommendation data set. The sample data for fine-tuning training includes: positive sample data and negative sample data. Among them, both the positive sample data and the negative sample data are composed of news text content, user historical behavior logs, news features for describing the characteristics of news text content, and user behavior features for describing the characteristics of user historical behavior logs. The positive sample data and the negative sample data are distinguished according to user behavior features; construct prompt instructions represented in natural language based on the positive sample data and the negative sample data respectively, and obtain an instruction fine-tuning data set; based on the low-rank decomposition matrix in the LoRA fine-tuning method and use the instruction fine-tuning data set to fine-tune the pre-trained third large language model, so that the third large language model performs news recommendation based on user historical behavior features, news text context information, user features, and news features. Based on the low-rank decomposition matrix in the LoRA fine-tuning method and use the instruction fine-tuning data set to fine-tune the pre-trained third large language model, so that the third large language model performs news recommendation based on user historical behavior features, news text context information, user features, and news features. Make the fine-tuned third large language model not only consider user features and news features, but also incorporate more context information and long-term user behavior data to provide more accurate recommendations.
[0050] As Figure 2 shown, two models with relatively small parameters simultaneously receive user features and news features as inputs and independently output recommendation results. For the recommendation results of the two small models, use a consistency judgment mechanism. If the recommendation results of the two models are exactly the same, that is, the recommended news articles are the same, then directly use this result as the final recommendation output. In most cases, the consensus of the two models can provide relatively reliable recommendations. When the recommendation results of the two small models are inconsistent, input these results into a larger model with more parameters for the final decision. Through large language model ensemble learning, it is possible to combine the parallel processing ability and final decision-making ability of the large model to achieve resource conservation and efficiency improvement in the news recommendation system, while ensuring the accuracy of the recommendation results.
[0051] Furthermore, based on the above method, an embodiment of the present invention also provides a news recommendation system based on large language model ensemble learning, including: a data collection module, an instruction construction module, an initial recommendation module, and a recommendation output module, where,
[0052] A data collection module for collecting data of users in a news system and filtering out several pieces of user feature data from the data, where the user feature data are the feature data most relevant to users' news browsing.
[0053] An instruction construction module for constructing a prompt instruction represented in natural language by using the user feature data.
[0054] An initial recommendation module for inputting the prompt instruction into a first news recommendation model and a second news recommendation model and obtaining news recommendation results output by both the first news recommendation model and the second news recommendation model. Among them, the first news recommendation model is obtained by fine-tuning a pre-trained first large language model, and the second news recommendation model is obtained by fine-tuning a pre-trained second large language model.
[0055] A recommendation output module for recommending news to users according to the news recommendation results output by both when the news recommendation results output by both the first news recommendation model and the second news recommendation model are consistent, and inputting the prompt instruction into a third news recommendation model to recommend news to users according to the news recommendation results output by the third news recommendation model when the news recommendation results output by both the first news recommendation model and the second news recommendation model are inconsistent. The third news recommendation model is a pre-trained third large language model.
[0056] To verify the effectiveness of the solution of this case, the solution of this case will be further explained below in combination with experimental data:
[0057] The MIND dataset (Microsoft News Dataset) is a large-scale news recommendation dataset constructed from the user click logs of public news. As a benchmark dataset in the field of news recommendation, the MIND dataset provides rich data resources for news recommendation research and promotes research in the fields of news recommendation and recommendation systems.
[0058] This dataset contains approximately 160,000 English news articles and over 15 million exposure logs generated by 1 million users. Each news article contains rich text content, such as title, abstract, body, category, and entity, and the information is stored in the news.tsv file. Each exposure log records the news articles shown to users when they visit the news website homepage at a specific time, as well as the click behavior of users on these news articles, and the information is stored in the behaviors.tsv file.
[0059] Read the two files news.tsv and behaviors.tsv in the MINDsmall_train data. The news.tsv file contains news ids and corresponding news titles, while the behaviors.tsv file has an Impression News field that contains the news items that the user clicked on and did not click on to construct the training data: the fields before Impression News are used as the user's past behavior preferences, and the last field is used to predict whether the user likes the news. That is, convert the user behavior data and news data into the format as shown in Figure 3 to generate the train.json dataset for fine-tuning the large model. Process the data in MINDsmall_dev in the same way to generate the test.json dataset for evaluating the large model.
[0060] Select the DeepSeek-R1-Distill-Llama-8B model with a relatively small parameter scale and use the LoRA fine-tuning method for model fine-tuning. The fine-tuned DeepSeek-R1-Distill-Llama-8B model is called Model A. Fine-tune the Meta-Llama-3-8B-Instruct model in the same way, which is called Model B. Use Model A and Model B as the preliminary recommendation models, and use a larger parameter-scale model DeepSeek-V3 (Model C) as the final decision-making model.
[0061] The experimental results show that using two large models with relatively small parameter scales to process the recommendation task in parallel can effectively save computing resources and significantly improve the system's response speed, not only reducing resource consumption but also enhancing the user experience, which is particularly important for news recommendation systems that need to handle a large number of user requests. When the recommendation results of the two small models are inconsistent, using the larger model with more parameters for the final decision can improve the recommendation accuracy, leveraging the advantages of large models in processing complex language and user behavior data, thereby improving the overall performance of the recommendation system. By combining the prediction results of different models, the recommendation quality can be improved while maintaining efficiency; by optimizing the model integration strategy, the news platform can provide more accurate and personalized recommendations, thereby increasing user stickiness and satisfaction.
[0062] Unless otherwise specifically stated, the relative steps, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0063] In the present specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments, and for the same or similar parts among the various embodiments, reference can be made to each other. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and for the relevant parts, reference can be made to the description in the method section.
[0064] The units and method steps of the various examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of the various examples have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art can use different methods to implement the described functions for each specific application, but such implementation is not considered to exceed the scope of the present invention.
[0065] A person of ordinary skill in the art can understand that all or part of the steps in the above methods can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a magnetic disk, or an optical disc, etc. Optionally, all or part of the steps of the above embodiments can also be implemented using one or more integrated circuits. Correspondingly, the various modules / units in the above embodiments can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of the combination of hardware and software.
[0066] Finally, it should be noted that the above-described embodiments are only specific implementation manners of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, a person of ordinary skill in the art should understand that any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions recorded in the foregoing embodiments, or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes, or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.
Claims
1. A news recommendation method based on large language model ensemble learning, characterized in that: Include: Collecting user data in the news system, and filtering out a number of characteristic data from the data, wherein the characteristic data includes: user characteristic data and news characteristic data most relevant to the user browsing news; Using feature data to construct prompt instructions expressed in natural language; Inputting a prompt instruction into the first news recommendation model and the second news recommendation model, and obtaining news recommendation results output by both the first news recommendation model and the second news recommendation model, wherein the first news recommendation model is obtained by fine-tuning the pre-trained first language model, and the second news recommendation model is obtained by fine-tuning the pre-trained second language model; If the news recommendation results output by the first news recommendation model and the second news recommendation model are consistent, news is recommended to the user based on the news recommendation results output by the two news recommendation models; if the news recommendation results output by the first news recommendation model and the second news recommendation model are inconsistent, the prompt instruction and the news recommendation results output by the two news recommendation models are input into the third news recommendation model to recommend news to the user based on the news recommendation results output by the third news recommendation model, wherein the third news recommendation model is a pre-trained third largest language model.
2. The news recommendation method based on large language model ensemble learning according to claim 1 is characterized in that: The first largest language model and the second largest language model adopt heterogeneous large language models, and the parameter scale of the third largest language model is respectively larger than the parameter scale of the first largest language model and the parameter scale of the second largest language model.
3. The news recommendation method based on large language model ensemble learning according to claim 1 is characterized in that: The process of fine-tuning the pre-trained first language model or the pre-trained second language model includes: In the pre-training stage, text data in related fields are collected and pre-processed, and the pre-processed text data are used to pre-train the first language model or the second language model, so that the first language model or the second language model learns the regularity of the text data, and the regularity of the text data includes: grammatical structure and semantic relationship; In the fine-tuning stage, sample data for fine-tuning training is constructed based on a public news recommendation data set, and the pre-trained first language model or the second language model is fine-tuned using the sample data for fine-tuning training. The sample data for fine-tuning training includes: positive sample data and negative sample data, wherein both the positive sample data and the negative sample data are composed of news features for describing news text features and user behavior features for describing user behavior log features, and the positive sample data and the negative sample data are distinguished according to the user behavior features.
4. The news recommendation method based on large language model ensemble learning according to claim 3 is characterized in that: Fine-tune the pre-trained first language model or the second language model using the fine-tuning training sample data, including: Based on the positive sample data and the negative sample data, prompt instructions expressed in natural language are constructed respectively, and an instruction fine-tuning dataset is obtained; The instruction fine-tuning dataset is used to fine-tune the pre-trained first language model or the second language model.
5. The news recommendation method based on large language model ensemble learning according to claim 4 is characterized in that: Use the instruction fine-tuning dataset to fine-tune the pre-trained first language model or the second language model, including: The low-rank decomposition matrix in the LoRA fine-tuning method is used to fine-tune the pre-trained first language model or the second language model.
6. The news recommendation method based on large language model ensemble learning according to claim 1 is characterized in that: Pre-train the third language model, including: Collect text data in related fields and pre-process the data, and use the pre-processed text data to pre-train the third language model so that the third language model can learn the rules of the text data, wherein the rules of the text data include: grammatical structure and semantic relationship.
7. The news recommendation method based on large language model ensemble learning according to claim 1 or 6, characterized in that: The third news recommendation model is obtained by fine-tuning the pre-trained third language model, and the fine-tuning process includes: Constructing sample data for fine-tuning training based on a public news recommendation data set, wherein the sample data for fine-tuning training includes: positive sample data and negative sample data, wherein both the positive sample data and the negative sample data are composed of news text content, user historical behavior logs, news features for describing news text content features, and user behavior features for describing user historical behavior log features, and the positive sample data and the negative sample data are distinguished according to the user behavior features; Based on the positive sample data and the negative sample data, prompt instructions expressed in natural language are constructed respectively, and an instruction fine-tuning dataset is obtained; Based on the low-rank decomposition matrix in the LoRA fine-tuning method and using the instruction fine-tuning dataset, the pre-trained third language model is fine-tuned so that the third language model can recommend news based on user historical behavior characteristics, news text context information, user characteristics and news characteristics.
8. A news recommendation system based on large language model ensemble learning, characterized in that: It includes: data collection module, instruction construction module, initial recommendation module and recommendation output module, among which, A data collection module is used to collect user data in the news system and filter out a number of user feature data from the data, wherein the user feature data is the feature data most relevant to the user browsing news; An instruction construction module, used for constructing prompt instructions expressed in natural language by using user feature data; an initial recommendation module, used to input the prompt instruction into the first news recommendation model and the second news recommendation model, and obtain the news recommendation results output by both the first news recommendation model and the second news recommendation model, wherein the first news recommendation model is obtained by fine-tuning the pre-trained first language model, and the second news recommendation model is obtained by fine-tuning the pre-trained second language model; The recommendation output module is used to recommend news to users based on the news recommendation results output by the first news recommendation model and the second news recommendation model when the news recommendation results output by the two models are consistent, and to input prompt instructions to the third news recommendation model when the news recommendation results output by the first news recommendation model and the second news recommendation model are inconsistent, so as to recommend news to users based on the news recommendation results output by the third news recommendation model, wherein the third news recommendation model is a pre-trained third language model.
9. An electronic device, characterized in that: include: at least one processor, and a memory coupled to the at least one processor; The memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 7 can be implemented.