Distributed intelligent question answering method based on novel federated knowledge learning model
By employing a novel federated knowledge learning model, the client and server collaboratively fine-tune the large model and perform task decomposition and knowledge aggregation, thus solving the problems of training efficiency and privacy protection for large-scale language models. This achieves efficient knowledge sharing and privacy protection, and improves the performance and security of the distributed question-answering system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE CHINESE UNIV OF HONG KONG (SHENZHEN)
- Filing Date
- 2026-04-14
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies face challenges in achieving efficient training mechanisms and meeting the demands of large-scale language models. Furthermore, directly applying federated learning carries the risk of privacy breaches due to model reverse engineering, making it difficult to effectively protect privacy data in distributed intelligent question-answering scenarios.
A novel federated knowledge learning model is adopted, which allows the client to fine-tune the large model of private downstream tasks locally and the server to fine-tune the global model on a public dataset. Combined with task decomposition, knowledge aggregation and privacy de-identification modules, it transmits task-related knowledge fragments, rather than the complete model, to achieve efficient knowledge sharing and privacy protection.
It achieves efficient knowledge sharing and privacy protection, significantly improves model performance and privacy security, reduces communication overhead, and enhances practicality and scalability in distributed question-answering systems.
Smart Images
Figure CN122021664A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing, and in particular to a distributed intelligent question answering method based on a novel federated knowledge learning model. Background Technology
[0002] Large-scale language models (LMs) have achieved remarkable success in natural language processing (NLP) tasks in recent years, becoming the standard for handling various complex tasks in this field. However, as the size of these models gradually increases, although their performance advantages become increasingly apparent, several new challenges have also arisen. Among the most significant issues are the need for efficient training mechanisms and the requirement for large amounts of data.
[0003] To address these challenges, Federated Learning (FL) has garnered significant attention in LLMs. The FL paradigm significantly reduces the burden on centralized training servers by distributing training tasks across multiple client devices and leveraging decentralized data from each device. Compared to traditional centralized methods, FL effectively distributes computational load and improves data privacy. However, in practice, the sheer number of parameters in large-scale language models limits the direct efficiency of FL; this is because aggregating and broadcasting the complete model in each round of federated learning incurs excessive computational and communication costs. In specific distributed intelligent question-answering scenarios (such as cross-institutional medical record question answering and cross-financial institution risk control text analysis), high-quality text datasets are often highly sensitive privacy data; directly applying traditional federated learning frameworks to distributed question-answering systems still carries the risk of text privacy leakage due to model reverse engineering. Therefore, an improved architecture for intelligent question-answering scenarios is needed to transform the exchange of massive model parameters into lightweight, de-identified knowledge fragment exchange, addressing the legal compliance and data silo issues faced by large models in practical applications.
[0004] To alleviate this problem, existing technologies have introduced efficient parameter fine-tuning techniques into the FL framework. For example, the LoR (Local Rank Adaptation) method improves the overall efficiency of computation and communication in the FL process by minimizing the number of parameters that need fine-tuning, successfully preserving most of the model's performance without comprehensively adjusting all model parameters. Although numerous studies have attempted to integrate FL with LLMs, this approach still faces inherent limitations. Directly applying traditional federated learning frameworks to LLMs presents significant challenges because traditional FL frameworks cannot fully meet the specific needs of LLMs, thus increasing system complexity. Traditional FL requires alignment of multiple language models, checkpoints, and tokenizers to achieve effective model aggregation. This requirement significantly limits practicality and scalability unless techniques such as knowledge distillation are introduced. Although some research has explored the potential of FL in promoting multi-domain knowledge fusion in LLMs, the resulting aggregated models often lack the domain-specific specialization of the original models, resulting in weakened specialization. Traditional FL must handle the parameters of various specialized LLMs while integrating heterogeneous data knowledge from multiple domains. Balancing the need for synchronization and heterogeneity not only increases time requirements but also exacerbates communication and computational complexity. In the context of LLMs, FL's privacy protection mechanisms are particularly vulnerable, posing a significant risk of privacy leaks. Especially when the server loads fine-tuned parameters from the client, context-sensitive models may gain access to private data. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a distributed intelligent question answering method based on a novel federated knowledge learning model, thereby solving the deficiencies of the existing technology.
[0006] The objective of this invention is achieved through the following technical solution: a distributed intelligent question-answering method based on a novel federated knowledge learning model, the method comprising:
[0007] S1. Each client uses local private data to fine-tune its private downstream task model, while the server uses public datasets to fine-tune the global model to enhance its knowledge understanding capabilities.
[0008] S2. The central server breaks down the task flow into the required requirements and sends these requirements to the client.
[0009] S3. The client generates the corresponding knowledge fragments based on the received requirements and uploads them to the server;
[0010] S4. The server aggregates the fragmented knowledge received from the client to form complete task-aiding knowledge.
[0011] S5 uses a server-wide model combined with task-aiding knowledge to generate task responses.
[0012] S1 specifically includes the following:
[0013] S101, Model Selection: The client selects a base model tailored to its private downstream tasks; Model Fine-tuning: The client utilizes its held private dataset. Fine-tuning the base model and maintaining the downstream task model ;
[0014] S102. Server builds a general knowledge understanding model: The server uses a public dataset. Fine-tune the local large model to develop a global model that excels in knowledge reasoning and logical understanding. .
[0015] S2 specifically includes the following:
[0016] Task flow requirements for large server models It can be dynamically added, removed, or modified according to the server's task requirements. The server first utilizes the requirement decomposition module. Each original query in the task flow Decompose into domain-specific knowledge requirements Meanwhile, the server uses a privacy de-identification module. After removing privacy-sensitive entries, the server then distributes the requests to each client.
[0017] The demand decomposition module For a summary and feature extraction language model T5, the server uses T5's feature extraction capabilities to decompose the task flow before demand allocation. The specific decomposition process includes:
[0018] The server pre-builds prompts containing task breakdown instructions;
[0019] The original intelligent question-answering task flow domain prompts are concatenated into text to construct the text-to-text input format required by the T5 model and then input into the T5 model.
[0020] Leveraging the natural language understanding and generation capabilities of the T5 model, core semantic features are extracted from the task flow, and structured output text containing multiple sub-tasks is directly generated.
[0021] The server segments the structured output text into several independent domain-specific knowledge requirements based on a preset delimiter, which are then distributed to the client.
[0022] S3 specifically includes the following:
[0023] Each client Generate a client knowledge set based on the received requirements. This includes domain-specific knowledge fragments, and each client requires a de-identification module. To ensure that the generated domain knowledge contains no sensitive information, n is the number of elements in the client's knowledge set;
[0024] Identification module The specific implementation mechanism includes: a de-identification module. With pre-configured personalized privacy de-identification prompts, the client concatenates the received knowledge requirements with hard prompts to form a contextual instruction with privacy constraints. Subsequently, the client calls the locally fine-tuned downstream task model to generate constrained text based on the contextual instruction, thereby directly outputting de-identified knowledge fragments that do not contain sensitive entities at the source.
[0025] S4 specifically includes the following:
[0026] Once the knowledge is generated, the client uploads the de-identified relevant knowledge fragments to the server. The server then obtains the complete set of task-related knowledge contributed by each client's domain model. The server utilizes the demand decomposition module To perform knowledge aggregation, clean up multiple knowledge fragments and integrate them into a high-quality task-aided knowledge entry, based on the original query. Select relevant knowledge items and integrate them into a unified set of supplementary knowledge items. N is the number of elements in the entire knowledge set.
[0027] S5 specifically includes the following:
[0028] The server's large model leverages its knowledge reasoning and logical understanding capabilities to perform original queries. and knowledge supplementary entries Derive the final task response ;
[0029] For task query set Each original query in the task flow The server generates the corresponding task response based on the large model. , forming a response set .
[0030] The method further includes:
[0031] S6. The server maintains the obtained triplet dataset and uses this dataset to fine-tune the global model and integrate the knowledge from various clients.
[0032] S6 specifically includes the following:
[0033] As the number of communication rounds increases, the server will accumulate an increasingly comprehensive dataset of historical task-driven records. It consists of knowledge, task, and response triples. For these triple data, the server uses instructions to fine-tune and update downstream knowledge across domains to further enhance its knowledge reasoning capabilities.
[0034] This invention offers the following advantages: a distributed intelligent question-answering method based on a novel federated knowledge learning (FKL) model. Unlike traditional methods, FKL does not exchange the complete models trained by individual nodes; instead, it transmits task-related knowledge fragments. This framework introduces task decomposition, knowledge aggregation, and de-identification modules, ensuring not only high performance but also robust privacy protection. Experimental results on public and private datasets demonstrate that FKL more effectively meets the needs of the LLM ecosystem and exhibits greater practicality in real-world deployments. Attached Figure Description
[0035] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0036] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application provided below with reference to the accompanying drawings is not intended to limit the scope of protection of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. The present invention will be further described below with reference to the accompanying drawings.
[0037] This invention specifically relates to a distributed intelligent question answering method based on a novel federated knowledge learning model. The method uses a model design based on knowledge learning generated by retrieval enhancement, which effectively solves the problems of bandwidth, computational requirements and heterogeneous text data faced by large model agents when processing cross-domain natural language question answering. It can provide an efficient and high-performance federated learning process while protecting the privacy of source data.
[0038] The purpose of this invention is to effectively improve overall model performance while also effectively protecting user data privacy. Designed specifically for LLMs from an architectural perspective, this invention offers high flexibility, efficiency, lossless knowledge aggregation capabilities, and privacy protection. In FKL, the server and client abandon traditional model sharing methods and adopt a unique knowledge sharing strategy. By leveraging task-related knowledge sharing, the server ensures accurate knowledge aggregation, thereby driving the creation of a versatile and universal LLM.
[0039] In a specific embodiment of the present invention, the federated knowledge learning architecture is mainly applied to a distributed intelligent question answering and natural language processing system. For ease of understanding, the 'dataset' in this embodiment can specifically be the private electronic text or medical record documents locally owned by each client (such as medical institutions or enterprise nodes); the 'task big model' is the downstream intelligent question answering model that performs natural language understanding and generation; and the 'task flow' is the natural language intelligent question answering query request input by the user.
[0040] like Figure 1 As shown, its overall execution logic includes the architecture construction phase (S1), the intelligent question answering execution phase (S2-S5), and the architecture evolution phase (S6), specifically including the following steps:
[0041] S1. Each client uses local private data to fine-tune its private downstream task model, while the server uses public datasets to fine-tune the global model to enhance its knowledge understanding capabilities.
[0042] Furthermore, S1 specifically includes the following:
[0043] S101, Model Selection: The client selects a base model tailored to its private downstream tasks; Model Fine-tuning: The client utilizes its held private dataset. Fine-tuning the base model and maintaining the downstream task model ;
[0044] S102. Server builds a general knowledge understanding model: The server uses a public dataset. Fine-tune the local large model to develop a global model that excels in knowledge reasoning and logical understanding. .
[0045] S2. The central server breaks down the task flow into the required requirements and sends these requirements to the client.
[0046] Task flow requirements for large server models It can be dynamically added, removed, or modified according to the server's task requirements. The server first utilizes the requirement decomposition module. Each raw query in the task flow Decompose into domain-specific knowledge requirements Meanwhile, the server uses a privacy de-identification module. Remove privacy-sensitive items; the requirement decomposition process is as follows:
[0047] ,
[0048] The server then distributes the requests to each client.
[0049] Among them, the requirement decomposition module For a summary and feature extraction language model T5, the server uses T5's feature extraction capabilities to decompose the task flow before demand allocation. The specific decomposition process includes:
[0050] First, the server pre-constructs prompts containing task breakdown instructions. Then, it concatenates the original intelligent question-answering task flow prompts to construct the "text-to-text" input format required by the T5 model and inputs it into the T5 model. Next, it utilizes the T5 model's natural language understanding and generation capabilities to extract core semantic features from the task flow and directly generate structured output text containing multiple sub-tasks. Finally, the server segments the structured output text according to preset delimiters to obtain several independent domain-specific knowledge requirements for distribution to clients.
[0051] Privacy De-identification Module Personalized hard cues are employed as cross-device privacy safeguards within the system. These cues are embedded into the original context of the target language model to regulate the output of its domain-specific knowledge.
[0052] S3. The client generates the corresponding knowledge fragments based on the received requirements and uploads them to the server;
[0053] Each client Generate a client knowledge set based on the received requirements. This includes domain-specific knowledge fragments, and each client requires a de-identification module. To ensure that the generated domain knowledge contains no sensitive information, n is the number of elements in the client's knowledge set;
[0054] Identification module The specific implementation mechanisms include:
[0055] De-identification module The client, pre-configured with personalized privacy-de-identification prompts, concatenates the received knowledge requirements with hard prompts to form a contextual instruction with privacy constraints. Subsequently, the client invokes a locally fine-tuned downstream task model to generate constrained text based on the contextual instruction, thus directly outputting de-identified knowledge fragments without sensitive entities at the source. This generation process can be represented by the following formula:
[0056] ,
[0057] Where n is a hyperparameter representing the number of relevant knowledge fragments generated by a domain-specific model.
[0058] S4. The server aggregates the fragmented knowledge received from the client to form complete task-aiding knowledge.
[0059] Once the knowledge is generated, the client uploads the de-identified relevant knowledge fragments to the server. The server then obtains the complete set of task-related knowledge contributed by each client's domain model. The server utilizes the demand decomposition module This involves knowledge aggregation, cleaning up multiple knowledge fragments and integrating them into a high-quality task-aided knowledge entry, based on the original queries in the task flow. Select relevant knowledge items and integrate them into unified knowledge auxiliary items. N is the number of elements in the entire knowledge set;
[0060] .
[0061] S5 uses a server-wide model combined with task-aiding knowledge to generate task responses.
[0062] The server's large model leverages its knowledge reasoning and logical understanding capabilities to perform original queries. and knowledge supplementary entries Derive the final task response :
[0063] .
[0064] For task query set Each task in The server generates the corresponding task response based on the large model. , forming a response set .
[0065] S6. The server maintains the (task, knowledge, response) triple dataset obtained in the above process, and uses this dataset to fine-tune the global model and integrate the knowledge of various clients.
[0066] As the number of communication rounds increases, the server will accumulate an increasingly comprehensive dataset of historical task-driven records. It consists of a triplet of knowledge, task, and response:
[0067] ;
[0068] For these triple data, the server uses instruction fine-tuning to update downstream knowledge across domains, further enhancing its knowledge reasoning capabilities:
[0069] ,
[0070] It is the fine-tuned global model. This indicates the training process involving fine-tuning instructions.
[0071] In the embodiments of this invention, compared to traditional solutions, the performance of this invention significantly surpasses existing solutions. In numerous experiments, this invention achieves significant improvements in several key metrics, and the method of this invention is denoted as FKL. Specifically, FKL improves tool invocation accuracy by approximately 50% on the GLUE benchmark compared to existing popular federated learning algorithms FedPETuning, FFA, FedSA, and FedAvg, and achieves a significant improvement of approximately 0.7 times in privacy task response quality. Furthermore, in a privacy task based on OPT-1.3B, the baseline method transmits approximately 96MB of LoRA parameters per round, while FKL transmits only 4KB of token-level requests and knowledge fragments per round, achieving a communication overhead reduction of approximately 2.46 × 10^4 times.
[0072] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and improvements, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
Claims
1. A distributed intelligent question-answering method based on a novel federated knowledge learning model, characterized in that: The method includes: S1. Each client uses local private data to fine-tune its private downstream task model, while the server uses public datasets to fine-tune the global model to enhance its knowledge understanding capabilities. S2. The central server breaks down the task flow into the required requirements and sends these requirements to the client. S3. The client generates the corresponding knowledge fragments based on the received requirements and uploads them to the server; S4. The server aggregates the fragmented knowledge received from the client to form complete task-aiding knowledge. S5 uses a server-wide model combined with task-aiding knowledge to generate task responses.
2. The distributed intelligent question answering method based on a novel federated knowledge learning model according to claim 1, characterized in that: S1 specifically includes the following: S101, Model Selection: The client selects the base model based on the private downstream task; Model fine-tuning: The client utilizes the private dataset it holds. Fine-tuning the base model and maintaining the downstream task model ; S102. Server builds a general knowledge understanding model: The server uses a public dataset. Fine-tune the local large model to develop a global model that excels in knowledge reasoning and logical understanding. .
3. The distributed intelligent question answering method based on a novel federated knowledge learning model according to claim 1, characterized in that: S2 specifically includes the following: Task flow requirements for large server models It can be dynamically added, removed, or modified according to the server's task requirements. The server first utilizes the requirement decomposition module. Each original query in the task flow Decompose into domain-specific knowledge requirements Meanwhile, the server uses a privacy de-identification module. After removing privacy-sensitive entries, the server then distributes the requests to each client.
4. The distributed intelligent question answering method based on a novel federated knowledge learning model according to claim 3, characterized in that: The demand decomposition module For a summary and feature extraction language model T5, the server uses T5's feature extraction capabilities to decompose the task flow before demand allocation. The specific decomposition process includes: The server pre-builds prompts containing task breakdown instructions; The original intelligent question-answering task flow domain prompts are concatenated into text to construct the text-to-text input format required by the T5 model and then input into the T5 model. Leveraging the natural language understanding and generation capabilities of the T5 model, core semantic features are extracted from the task flow, and structured output text containing multiple sub-tasks is directly generated. The server segments the structured output text into several independent domain-specific knowledge requirements based on a preset delimiter, which are then distributed to the client.
5. A distributed intelligent question-answering method based on a novel federated knowledge learning model according to claim 1, characterized in that: S3 specifically includes the following: Each client Generate a client knowledge set based on the received requirements. This includes domain-specific knowledge fragments, and each client requires a de-identification module. This ensures that the generated domain knowledge contains no sensitive information, where n is the number of elements in the client's knowledge set; Identification module The specific implementation mechanism includes: a de-identification module. With pre-configured personalized privacy de-identification prompts, the client concatenates the received knowledge requirements with hard prompts to form a contextual instruction with privacy constraints. Subsequently, the client calls the locally fine-tuned downstream task model to generate constrained text based on the contextual instruction, thereby directly outputting de-identified knowledge fragments that do not contain sensitive entities at the source.
6. The distributed intelligent question answering method based on a novel federated knowledge learning model according to claim 1, characterized in that: S4 specifically includes the following: Once the knowledge is generated, the client uploads the de-identified relevant knowledge fragments to the server. The server then obtains the complete set of task-related knowledge contributed by each client's domain model. The server utilizes the demand decomposition module To perform knowledge aggregation, clean up multiple knowledge fragments and integrate them into a high-quality task-aided knowledge entry, based on the original query. Select relevant knowledge items and integrate them into a unified set of supplementary knowledge items. , where N is the number of elements in the entire knowledge set.
7. A distributed intelligent question-answering method based on a novel federated knowledge learning model according to claim 6, characterized in that: S5 specifically includes the following: The server's large model leverages its knowledge reasoning and logical understanding capabilities to perform original queries. and knowledge supplementary entries Derive the final task response ; For task query set Each original query in the task flow The server generates the corresponding task response based on the large model. , forming a response set .
8. A distributed intelligent question answering method based on a novel federated knowledge learning model according to any one of claims 1-7, characterized in that: The method further includes: S6. The server maintains the obtained triplet dataset and uses this dataset to fine-tune the global model and integrate the knowledge from various clients.
9. A distributed intelligent question-answering method based on a novel federated knowledge learning model according to claim 8, characterized in that: S6 specifically includes the following: As the number of communication rounds increases, the server will accumulate an increasingly comprehensive dataset of historical task-driven records. It consists of knowledge, task, and response triples. For these triple data, the server uses instructions to fine-tune and update downstream knowledge across domains to further enhance its knowledge reasoning capabilities.