Personalized science and technology knowledge question answering method and system

Abstract

The invention discloses a personalized scientific and technological knowledge question answering method and system, and the method comprises the steps: firstly obtaining and processing scientific and technological literature data, constructing a field basic model training corpus, and carrying out the continuous pre-training, and forming a field basic model; then, on the basis of the field basic model, parameters of a sampling model are initialized, initial training is conducted through seed data, the sampling model is trained through multi-round iterative reinforcement learning, and sample data are generated; and then, on the basis of the field basic model, initializing parameters of a question and answer model, constructing a distillation training data set by using sample data generated by the sampling model, and performing supervised fine tuning training until the model converges. And finally, packaging the trained question and answer model into an online reasoning interface service, and constructing a science and technology knowledge question and answer system. According to the method, scientific and technological knowledge question-answering service aiming at personalized requirements of the user is realized, and the question answering efficiency and accuracy are remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, specifically to the fields of natural language processing and deep learning, and in particular to a personalized scientific and technological knowledge question-answering method and system. Background Art

[0002] With the continuous advancement of science and technology in my country, the demand for scientific and technological knowledge is increasing. Using natural language processing technology to retrieve scientific and technological texts is playing an increasingly important role in the scientific research process of professionals in the field. Accurately and efficiently understanding scientific and technological intelligence information has become a key task in the field of science and technology.

[0003] The highly specialized nature of science and technology, along with complex text structures and rich semantic content, makes it difficult for users to access scientific literature. Therefore, building a scientific knowledge question-answering system based on scientific and technological text corpora, such as professional books, journals, papers, and patents, integrating scientific and technological knowledge, targeting the acquisition preferences of different user groups, and facilitating the retrieval of domain-specific scientific and technological information has become a key research topic.

[0004] Common personalized science and technology knowledge question-and-answer systems are generally implemented based on recommendation methods. By establishing a relationship between users and knowledge, answers are delivered to users in an active recommendation manner.

[0005] In the public-facing science and technology knowledge question system, there are many professional fields involved and a huge number of users. Traditional knowledge question-answering systems have difficulty effectively matching domain knowledge with user characteristics and search purpose preferences. In the process of science and technology knowledge retrieval, on the one hand, due to the large variety and quantity of science and technology knowledge, high professionalism, and reading difficulties, on the other hand, users have specific professional backgrounds and clear task goals. The search results in professional fields are usually highly relevant to the technical fields and research interests of the users. The disadvantage of poor adaptability of general search results to user interests leads to the overall low relevance and accuracy of professional text understanding and search results. Therefore, existing recommendation-based retrieval systems find it difficult to find and obtain science and technology knowledge that meets user needs. They lack effective means of obtaining science and technology knowledge, and the interactive process often requires a lot of time and energy, which cannot meet the needs of users to effectively obtain science and technology knowledge in the face of new problems, cross-domain, and cutting-edge technology scenarios. Summary of the Invention

[0006] This application provides a personalized scientific and technological knowledge question-answering method and system. Based on a basic model trained on scientific and technological literature data, the system uses search history as seed data and trains a sampling model through reinforcement learning in a multi-round iterative mode. Sample data is then generated through sampling, and the large question-answering model is fine-tuned through data distillation. During use, user characteristics and conversation history are fully input into the question-answering model through small sample prompts, and personalized answers are dynamically generated through thought chain reasoning, which specifically improves the efficiency and accuracy of question answering. This application overcomes the difficulty of low retrieval efficiency in existing recommendation-based knowledge question-answering systems and solves the technical problem of low user satisfaction.

[0007] First, a personalized scientific and technological knowledge question-answering method includes four steps: domain foundation model construction, sampling model training, question-answering model training, and question-answering system construction. Specifically, the method includes:

[0008] Construction of domain-based models: Acquire and process scientific literature data to construct domain-based model training corpus. Based on the open-source pre-trained model, use the training corpus to continue pre-training and construct the domain-based model.

[0009] Sampling model training: Based on the domain basic model, the parameters of the sampling model are initialized, and the sampling model is initially trained using seed data. The sampling model is trained through multiple rounds of iterative reinforcement learning to generate sample data.

[0010] Question-answering model training: Based on the domain basic model, the parameters of the question-answering model are initialized. The sample data generated by the sampling model is used to construct a distillation training dataset. The question-answering model is then fine-tuned in a supervised manner until convergence.

[0011] Question-and-answer system construction: Encapsulate the trained question-and-answer model into an online reasoning interface service to build a scientific and technological knowledge question-and-answer system. During the use of the question-and-answer system, user characteristics and conversation history are input into the question-and-answer model to dynamically generate personalized answers.

[0012] Optionally, in the domain basic model construction step, processing scientific literature data includes deduplicating the data, deleting HTML / XML tags and non-text content, and splitting the text data into title, abstract, background, problem, technical solution, and effect parts.

[0013] Optionally, in the sampling model training step, the seed data includes user information and user historical retrieval data, wherein the user information is the user name, age, occupation, unit, and industry information data registered when the user opens an account, and the historical retrieval data includes user queries, returned results, and user feedback information data through clicks or browsing.

[0014] Optionally, in the sampling model training step, the sampling model is trained through multiple rounds of iterative reinforcement learning, including: initializing the domain basic model parameters at the beginning of each iteration, using the current sampling model to generate multiple candidate answers to user questions, scoring the candidate answers in combination with rule verification and large model evaluation, and using the online GRPO reinforcement learning algorithm to update the parameters of the sampling model.

[0015] Optionally, the parameters of the sampling model are updated using an online GRPO reinforcement learning algorithm, specifically by the formula:

[0016]

[0017] Training is performed, where θ is the sampling model parameter; q is the query input, which obeys the P(Q) distribution; J GRPO (θ) is the optimization target; G is the number of sample groups, which is 16; π θ New strategies for sampling models; is the old strategy of the sampling model; π ref is the reference model strategy, which is the domain basic model at the beginning and the previous round sampling model in the subsequent iterations; clip is the clipping function; o i is the output of the i-th sample; o i，t is the output of the i-th sample at time t; o i，＜t is the output of the i-th sample before time t; ∈ is the shear range hyperparameter; β is the KL divergence weight coefficient; D KL is the KL divergence distance:

[0018]

[0019] Calculated by the relative reward value within the group: is the normalized reward, r i is the reward of the i-th sample, r represents the reward variable on all samples, mean(r) is the mean of the variable r, and std(r) is the standard deviation of the variable r.

[0020] Optionally, in the question-answering model training step, constructing a distillation training dataset using sample data generated by a sampling model includes: performing rejection sampling filtering on the sample data to remove samples with incorrect formats or scores below a threshold, combining multiple historical conversation processes of the same user into synthetic sample data in chronological order of sampling time, and splicing it with the current question to generate prompt words as model input.

[0021] In the second aspect, a personalized scientific and technological knowledge question-answering system includes a domain basic model construction module, a sampling model training module, a question-answering model training module, and a question-answering system construction module, which specifically includes:

[0022] The domain-based model construction module is used to acquire and process scientific and technological literature data, construct domain-based model training corpus, and use the training corpus to continue pre-training based on the open source pre-training model to build the domain-based model;

[0023] The sampling model training module is used to initialize the parameters of the sampling model based on the domain basic model, perform initial training on the sampling model using seed data, and train the sampling model through multiple rounds of iterative reinforcement learning to generate sample data;

[0024] The question-answering model training module is used to initialize the parameters of the question-answering model based on the domain basic model, construct a distillation training dataset using the sample data generated by the sampling model, and perform supervised fine-tuning training on the question-answering model until convergence;

[0025] The question-answering system construction module is used to encapsulate the trained question-answering model into an online reasoning interface service to build a scientific and technological knowledge question-answering system. During the use of the question-answering system, user characteristics and conversation history are input into the question-answering model to dynamically generate personalized answers.

[0026] In a third aspect, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the personalized scientific and technological knowledge question and answer method described in any one of the first aspects is implemented.

[0027] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the personalized scientific and technological knowledge question and answer method described in any one of the first aspects is implemented.

[0028] In a fifth aspect, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the personalized scientific and technological knowledge question-and-answer method described in any one of the first aspects above.

[0029] The present invention provides a personalized knowledge question-and-answer system based on a large model, which is implemented by comprehensively applying a variety of training data sets and model training methods. The domain basic model is trained based on scientific and technological literature data, so that the model has the ability to recognize domain knowledge. The retrieval history is used as seed data to solve the cold start problem of the sampling model due to the lack of labeled data. After the sampling model is trained through reinforcement learning in a multi-round iterative mode, sample data is generated, which can generate a large amount of training data for training deep thinking models, so that the sampling model has domain logic reasoning capabilities. The sampling model avoids the model degradation problem caused by overfitting of the value function through reinforcement learning of the online GRPO algorithm, and can efficiently learn discrete reward signals in single-round interaction data, focusing on the current user's current intention recognition model optimization scenario. The large question-answering model is fine-tuned based on data distillation. During use, user characteristics and purposes are input into the question-answering model through a few-sample method. The static characteristics and dynamic intentions of the current user can be perceived through few-sample reasoning. The user question-answering context is dynamically generated through thought chain reasoning to generate personalized answers to questions. It can dynamically adapt to situations where user characteristics are diverse and usage intentions are changing in the question-answering system. The generalization ability of the question-answering model meets the personalized preference applications of various types of users in real usage scenarios, meets the user's personalized scientific and technological knowledge acquisition needs, and improves the efficiency and accuracy of question answering in a targeted manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0031] Figure 1 The main flow chart provided for the embodiment of this application;

[0032] Figure 2 A schematic diagram of the sampling model training data format provided in an embodiment of the present application;

[0033] Figure 3 A schematic diagram of the training data composition provided in an embodiment of the present application;

[0034] Figure 4 A schematic diagram of the dependency relationship between models provided in the embodiment of this application;

[0035] Figure 5 This is a block diagram of a personalized science and technology knowledge question-answering system;

[0036] Figure 6 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0038] In the description of the present application: the terms "including", "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that have been explicitly listed, but may also include other steps or units that are not explicitly listed but are inherent to these processes, methods, products or apparatuses, or steps or units that are added based on further optimization solutions conceived by the present invention.

[0039] In order to overcome the low retrieval efficiency and low user satisfaction of existing recommendation-based knowledge question-answering systems, the present invention provides a personalized knowledge question-answering system based on a large model. Based on the basic model of the scientific and technological literature data training field, the retrieval history is used as seed data. The sampling model is trained through reinforcement learning in a multi-round iterative mode. Sample data is then generated through sampling. The large question-answering model is fine-tuned through data distillation. During use, user characteristics and conversation history are fully input into the question-answering model through small sample prompts. Personalized answers are dynamically generated through thought chain reasoning, which specifically improves the efficiency and accuracy of question answering.

[0040] In one embodiment, Figure 1 As shown in the figure, a personalized scientific knowledge question-answering method and a large-scale model-based personalized knowledge question-answering system are provided. This system uses data distillation to construct diverse training data and implements multi-stage, multi-model cumulative training. The system consists of a domain-based model, a sampling model, a question-answering model, and a question-answering system.

[0041] The training corpus of the domain-based model is generated through public and private scientific literature texts, and is constructed through continued pre-training based on the open source large model.

[0042] The sampling model is based on the domain basic model, uses the retrieval history as seed data, and is trained through reinforcement learning in a multi-round iterative self-improvement mode.

[0043] The question-answering model is based on the domain basic model, uses samples generated by the sampling model as training data, and is fine-tuned through data distillation.

[0044] The question system encapsulates the question-answering model into an online reasoning interface service to build a scientific and technological knowledge question-answering system.

[0045] The method includes: constructing a domain-based model: acquiring and processing scientific and technological literature data, constructing a domain-based model training corpus, and continuing pre-training with the training corpus based on an open source pre-training model to construct a domain-based model;

[0046] Sampling model training: Based on the domain basic model, the parameters of the sampling model are initialized, and the sampling model is initially trained using seed data. The sampling model is trained through multiple rounds of iterative reinforcement learning to generate sample data.

[0047] Question-answering model training: Based on the domain basic model, the parameters of the question-answering model are initialized. The sample data generated by the sampling model is used to construct a distillation training dataset. The question-answering model is then fine-tuned in a supervised manner until convergence.

[0048] Question-and-answer system construction: Encapsulate the trained question-and-answer model into an online reasoning interface service to build a scientific and technological knowledge question-and-answer system. During the use of the question-and-answer system, user characteristics and conversation history are input into the question-and-answer model to dynamically generate personalized answers.

[0049] The specific process is as follows Figure 1 As shown:

[0050] The domain-based model is constructed by pre-training scientific and technological literature data based on an open-source basic model. The domain-based model training corpus is constructed by acquiring, cleaning, and splitting scientific and technological literature. Acquiring scientific and technological literature involves categorizing and organizing publicly available online professional books, journal articles, blog pages, and other data, as well as private journal articles and patent data. Cleaning involves deduplicating and removing HTML / XML tags and non-text content. Splitting further separates the original text into sections such as title, abstract, background, problem, technical solution, and effects based on content. The above text data is formatted to form the domain-based model training corpus. Based on the open-source pre-trained model, the domain model is trained on the domain corpus data through continued pre-training. The QWen2.5 series of models can be used as open-source pre-trained models.

[0051] The sampling model is trained through cyclic iterative reinforcement learning based on the domain basic model.

[0052] The sampling model is trained through multiple rounds of iterations, with the parameters of the domain-based model initialized at the beginning of each iteration. The initial training data for the sampling model is collected from online search system data as cold-start seed training data. In subsequent iterations, the sampling model uses a self-improving sampling model to generate training data for the next stage. Specifically, the sampling model generates multiple ranked candidate answers for the same user question. The system scores the candidate answers through a combination of rule verification and large-scale model evaluation. Deepseek R1 can be used to evaluate the large-scale model, accessible through an online API.

[0053] The cold-start seed data used for model initialization training is a preference training dataset constructed based on historical data from an online search system. This seed data is obtained by collecting historical user search data from existing online search systems and associating it with user information. This information includes user name, age, occupation, company, industry, and other information registered when a user opens an account. Historical search data includes user queries, returned results, and user feedback through clicks or browsing. Users enter free text in the online search system, and the system returns a list of search results with a fixed number of results per page. Users obtain answers by clicking and browsing. The system's feedback is ranked and scored based on user feedback to construct the preference training dataset.

[0054] The training data for the sampling model uses question-answer pairs from a complete user conversation. By designing prompts, the model is guided to first generate a user feature description in a thought chain manner, and then generate an answer based on the current question. Prompts include requiring the model to output content in a specified format, placing the deep thinking process and the answer in different specific tags. For example, the model can be asked to place the thinking process in <think>< / think> Inside the tag, place the answer <answer>< / answer> The generated sample data is evaluated and ranked using a scoring function. The scoring function includes rule-based output format verification and scoring the sample data using a large model.

[0055] Use the online GRPO reinforcement learning algorithm to update the parameters of the sampling model, specifically through the formula:

[0056]

[0057] Training is performed, where θ is the sampling model parameter; q is the query input, which obeys the P(Q) distribution; J GRPO (θ) is the optimization target; G is the number of sample groups, which is 16; π θ New strategies for sampling models; is the old strategy of the sampling model; π refis the reference model strategy, which is the domain basic model at the beginning and the previous round sampling model in the subsequent iterations; clip is the clipping function; o i is the output of the i-th sample; o i，t is the output of the i-th sample at time t; o i，＜t is the output of the i-th sample before time t; ∈ is the shear range hyperparameter; β is the KL divergence weight coefficient; D KL is the KL divergence distance:

[0058]

[0059] Calculated by the relative reward value within the group: is the normalized reward, r i is the reward of the i-th sample, r represents the reward variable on all samples, mean(r) is the mean of the variable r, and std(r) is the standard deviation of the variable r.

[0060] Figure 2 This is the training data format for the sampling model. The data consists of prompts, user information, questions, and answers. The prompts also include character settings and instructions. The answers consist of multiple candidate answers to the same question and their ratings, forming a group. Each group contains N answers and ratings. N can be set to 16.

[0061] The question-answering model training is based on the domain foundation model. Sample data generated by the sampling model is used to construct a distilled training dataset, which is obtained through supervised fine-tuning. The sample data generated by the sampling model is filtered through rejection sampling. Synthetic sample data is then generated by grouping multiple historical conversations with the same user in chronological order. This data is then concatenated with the current question to generate prompt words, which serve as model input. Rejection sampling utilizes rule-based format verification and result evaluation methods to discard incorrectly formatted outputs and samples with scores below a specified threshold.

[0062] The data distillation training process of the sampling model and question-answering model is shown in Algorithm 1:

[0063] Algorithm 1: Sampling model and question-answering model training process:

[0064] Input: seed training dataset {Mtrain}, observation o∈O, action a∈A, reward r∈R;

[0065] Neural network parameters of the sampled model is the model parameter of the i-th iteration;

[0066] The neural network parameters θ of the question-answering model;

[0067] The neural network parameters Ψ of the domain base model;

[0068] Reward scoring function F;

[0069] The number of training data samples N for the question-answering model.

[0070] Output: Neural network parameters θ of the question-answering model.

[0071] step:

[0072] Initialize the training dataset D as the seed training dataset {Mtrain}.

[0073] Loop i from 1 to N, iteratively perform sampling model training, and update neural network parameters Until convergence:

[0074] (1) Initialize the sampling model parameters to the domain base model parameters:

[0075] (2) Randomly sample sample data from the training data set D.

[0076] hT(i)=(o0,a0,r0,o1,a1,r1,…,oT,aT,rT)i, construct the training dataset Mitrain.

[0077] (3) Apply the reinforcement learning algorithm GRPO to train the sampling model on the training dataset Mitrain until Converge to the optimal strategy.

[0078] (4) Clear the training dataset D.

[0079] (5) Using sampling models Perform sampling, generate sample data, score the samples using the reward scoring function, and add the sample data to the dataset D.

[0080] Execute question-answering model training and update the neural network parameters θ until convergence:

[0081] (1) Initialize the question-answering model parameters to the domain-based model parameters: θ = Ψ.

[0082] (2) Iterate the training dataset D and randomly sample N sample data hT(i) according to user groups. Then sort them by time and combine them into sample data:

[0083] HT(j)=(hT(1),hT(2),…,hT(N-1),hT(N))j, stored in the training data set Mtrain.

[0084] (3) Apply the supervised fine-tuning algorithm to train the question-answering model on the training dataset Mtrain, and update the parameter θ until θ converges to the optimal value.

[0085] (4) Save the neural network parameters θ of the question-answering model.

[0086] Figure 3 This figure illustrates the data relationship between the sampling model and the question-answering model training. The sampling model uses reinforcement learning to train on a single user question-answering data set, while the question-answering model combines multiple instances of the same user for training. In the figure, a, o, and r represent the action, observation, and reward in reinforcement learning, respectively, corresponding to the token output of the larger model. The question system encapsulates the question-answering model into an online inference interface, providing users with scientific and technological knowledge Q&A services.

[0087] Figure 4 The logical relationship between the domain basic model, sampling model, question-answering model, their training data, and the question-answering system in the system is described. The domain basic model is constructed based on scientific and technological literature corpus training, and serves as the basis for the sampling model (each iteration) and the question-answering model. The model is initialized by parameter replication. The sampling model is trained in a multi-round iterative mode. During initialization (stage 1), it is trained with seed data, and the data for subsequent iterations (stage i) is obtained by sampling the model of the previous iteration (stage i-1). The training data for the large question-answering model is generated by the sampling data of the last iteration (N) of the sampling model. The question-answering system is implemented based on the inference service provided by the large question-answering model.

[0088] From the above, it can be seen that the present invention provides a personalized knowledge question-and-answer system based on a large model, which adopts multiple training data construction methods and is implemented in a multi-stage multi-model cumulative training manner. Based on the basic model of the scientific literature data training field as the subsequent sampling model and question-and-answer model, the retrieval history is used as seed data, and the sampling model is trained through reinforcement learning in a multi-round iterative mode to generate sample data. The large question-and-answer model is fine-tuned through data distillation. During use, user characteristics and purposes are fully input into the question-and-answer model through small sample prompts. The user question-and-answer context is dynamically generated by thought chain reasoning to generate personalized question answers, thus realizing a user-personalized scientific and technological knowledge question-and-answer system.

[0089] In one embodiment, Figure 5 As shown, a personalized scientific and technological knowledge question-answering system is provided. The system includes a domain basic model construction module, a sampling model training module, a question-answering model training module, and a question-answering system construction module, which specifically includes:

[0090] The domain-based model construction module is used to acquire and process scientific and technological literature data, construct domain-based model training corpus, and use the training corpus to continue pre-training based on the open source pre-training model to build the domain-based model;

[0091] The sampling model training module is used to initialize the parameters of the sampling model based on the domain basic model, perform initial training on the sampling model using seed data, and train the sampling model through multiple rounds of iterative reinforcement learning to generate sample data;

[0092] The question-answering model training module is used to initialize the parameters of the question-answering model based on the domain basic model, construct a distillation training dataset using the sample data generated by the sampling model, and perform supervised fine-tuning training on the question-answering model until convergence;

[0093] The question-answering system construction module is used to encapsulate the trained question-answering model into an online reasoning interface service to build a scientific and technological knowledge question-answering system. During the use of the question-answering system, user characteristics and conversation history are input into the question-answering model to dynamically generate personalized answers.

[0094] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device is used to provide computing and control capabilities, and the network interface is used to communicate with an external terminal via a network connection. The computer device implements the above-mentioned personalized science and technology knowledge question-and-answer method by loading and running a computer program.

[0095] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0096] In one embodiment, a computer-readable storage medium is further provided, on which a computer program is stored, which involves all or part of the processes in the above-mentioned embodiment method.

[0097] In one embodiment, a computer program product is also provided, including a computer program / instruction, which involves all or part of the process in the above embodiment method.

[0098] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in M ​​forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (SyMchliMk) DRAM (SLDRAM), memory bus (RaMbus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0099] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A personalized scientific and technological knowledge question-answering method, characterized in that: The method includes four steps: domain basic model construction, sampling model training, question-answering model training, and question-answering system construction, which specifically include: Construction of domain-based models: Acquire and process scientific literature data to construct domain-based model training corpus. Based on the open-source pre-trained model, use the training corpus to continue pre-training and construct the domain-based model. Sampling model training: Based on the domain basic model, the parameters of the sampling model are initialized, and the sampling model is initially trained using seed data. The sampling model is trained through multiple rounds of iterative reinforcement learning to generate sample data. Question-answering model training: Based on the domain basic model, the parameters of the question-answering model are initialized. The sample data generated by the sampling model is used to construct a distillation training dataset. The question-answering model is then fine-tuned in a supervised manner until convergence. Question-and-answer system construction: Encapsulate the trained question-and-answer model into an online reasoning interface service to build a scientific and technological knowledge question-and-answer system. During the use of the question-and-answer system, user characteristics and conversation history are input into the question-and-answer model to dynamically generate personalized answers.

2. The method according to claim 1, characterized in that In the domain basic model construction step, processing scientific literature data includes deduplication of data, removal of HTML / XML tags and non-text content, and splitting text data into title, abstract, background, problem, technical solution, and effect parts.

3. The method according to claim 1, characterized in that In the sampling model training step, the seed data includes user information and user historical retrieval data, wherein the user information is the user name, age, occupation, unit, and industry information data registered when the user opens an account, and the historical retrieval data includes user queries, returned results, and user feedback information data through clicks or browsing.

4. The method according to claim 1, wherein In the sampling model training step, the sampling model is trained through multiple rounds of iterative reinforcement learning, including: initializing the domain basic model parameters at the beginning of each iteration, using the current sampling model to generate multiple candidate answers to user questions, scoring the candidate answers in combination with rule verification and large model evaluation, and using the online GRPO reinforcement learning algorithm to update the parameters of the sampling model.

5. The method according to claim 4, characterized in that Use the online GRPO reinforcement learning algorithm to update the parameters of the sampling model, specifically through the formula: Training is performed, where θ is the sampling model parameter; q is the query input, which obeys the P(Q) distribution; J GRPO (θ) is the optimization target; G is the number of sample groups, which is 16; π θ New strategies for sampling models; is the old strategy of the sampling model; π ref is the reference model strategy, which is the domain basic model at the beginning and the previous round sampling model in the subsequent iterations; clip is the clipping function; o i is the output of the i-th sample; o i，t is the output of the i-th sample at time t; o i，＜t is the output of the i-th sample before time t; ∈ is the shear range hyperparameter; β is the KL divergence weight coefficient; D KL is the KL divergence distance: Calculated by the relative reward value within the group: is the normalized reward, r i is the reward of the i-th sample, r represents the reward variable on all samples, mean(r) is the mean of the variable r, and std(r) is the standard deviation of the variable r.

6. The method according to claim 1, wherein In the question-answering model training step, constructing a distillation training dataset using sample data generated by a sampling model includes: performing rejection sampling filtering on the sample data to remove samples with incorrect formats or scores below a threshold, combining multiple historical conversation processes of the same user into synthetic sample data in chronological order of sampling time, and splicing the generated prompt words with the current question as model input.

7. A personalized science and technology knowledge question-answering system, characterized by: The system includes a domain basic model construction module, a sampling model training module, a question-answering model training module, and a question-answering system construction module, which specifically include: The domain-based model construction module is used to acquire and process scientific and technological literature data, construct domain-based model training corpus, and use the training corpus to continue pre-training based on the open source pre-training model to build the domain-based model; The sampling model training module is used to initialize the parameters of the sampling model based on the domain basic model, perform initial training on the sampling model using seed data, and train the sampling model through multiple rounds of iterative reinforcement learning to generate sample data; The question-answering model training module is used to initialize the parameters of the question-answering model based on the domain basic model, construct a distillation training dataset using the sample data generated by the sampling model, and perform supervised fine-tuning training on the question-answering model until convergence; The question-answering system construction module is used to encapsulate the trained question-answering model into an online reasoning interface service to build a scientific and technological knowledge question-answering system. During the use of the question-answering system, user characteristics and conversation history are input into the question-answering model to dynamically generate personalized answers.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images