A method and system for searching question-answering results based on a large model.

Abstract

This application relates to a question-answering result search method and system based on a large model. The method first trains N first question-answering models using first sample data composed of question-answer pairs from various reading platforms for ebooks. Each first question-answering model fully learns the question-answering patterns of users on a specific reading platform. Then, it extracts the derivation process vector representation of a pre-set mature large language model and uses this derivation process vector representation and the first sample data to form second sample data. This second sample data contains knowledge of how the mature large language model thinks. Finally, it uses the second sample data, which possesses the thinking ability of the mature large language model, to enhance the training of the second question-answering model based on the combination of N first question-answering models and a fusion model, thereby improving the accuracy of the second question-answering model in searching for the current answer text of the current question text.

Description

Technical Field

[0001] This application relates to the field of large model technology, and in particular to a method and system for searching question-answering results based on large models. Background Technology

[0002] With the rapid development of Artificial Intelligence (AI) technology, large-scale language models (such as the GPT series, LLaMA, and ChatGLM) have demonstrated outstanding performance in natural language processing tasks such as intelligent question answering, text summarization, and dialogue generation. These models, pre-trained on massive amounts of text data, have acquired powerful language understanding and generation capabilities. Especially in the field of e-book reading, users often need to ask questions about the book's content; an efficient question-answering system can greatly enhance the reading and learning experience.

[0003] Currently, there are two main ways to implement question-answering systems based on large models:

[0004] 1) Directly using mature LLMs (such as ChatGPT) to answer user questions has the advantage of strong model versatility and broad knowledge coverage, but for questions and answers in specific domains (such as the detailed content of an e-book), the answers may not be accurate enough and may easily lead to illusions.

[0005] 2) Fine-tuning a small or basic LLM for domain-specific data (such as question-answer pairs from a single ebook on a single reading platform) to obtain a dedicated question-answering model may perform well on that specific dataset, but it is prone to overfitting and has poor generalization ability. When users come from different platforms, even if they ask questions about the same ebook, their question expression habits and answer preferences may differ, leading to unstable performance of a single model. Summary of the Invention

[0006] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0007] The main objective of this disclosure is to propose a question-and-answer result search method and system based on a large model, which can improve the accuracy of searching for the current answer text of the current question text.

[0008] The first aspect of this application proposes a question-answering result search method based on a large model, the method comprising:

[0009] Multiple first data samples of the same e-book from N reading platforms are obtained, along with N initial large language models with identical structures. The first data samples from each reading platform are used to train each initial large language model separately to obtain N first question-answering models. Here, N is an integer greater than 1, and the first data samples include a set of historical question texts and historical answer texts.

[0010] Based on multiple first data samples from N reading platforms, the derivation process vector representation of the mature large language model is extracted from a preset mature large language model, and a second data sample is formed based on each first data sample and the corresponding derivation process vector representation to obtain multiple second data samples; wherein, the mature large language model is an open-source large language model;

[0011] A second question-answering model is constructed based on N first question-answering models and a fusion model, and the second question-answering model is enhanced and trained based on multiple second data samples to obtain the enhanced second question-answering model; the fusion model is used to select one prediction result from the N prediction results corresponding to the N first question-answering models as the prediction result of the second question-answering model;

[0012] The current question text of the e-book is input into the second question-answering model that has been enhanced and trained, so that the prediction result output by the second question-answering model after enhancement training is used as the current answer text of the current question text.

[0013] The question-answering result search method based on a large model provided in this embodiment has at least the following beneficial effects:

[0014] This method first trains N first question-answering models using first sample data composed of question-answer pairs from various reading platforms for e-books. Each first question-answering model fully learns the question-answering patterns of users on a specific reading platform. Then, it extracts the derivation process vector representation of the mature large language model from a pre-set mature large language model, and combines the derivation process vector representation with the first sample data to form second sample data. This second sample data contains knowledge of how the mature large language model thinks. Finally, it uses the second sample data with the thinking ability of the mature large language model to enhance the training of the second question-answering model based on the combination of N first question-answering models and the fusion model, so as to improve the accuracy of the second question-answering model in searching for the current answer text of the current question text.

[0015] A second aspect of this application proposes a question-answering result search system based on a large model, the system comprising:

[0016] The first training module is used to acquire multiple first data samples of the same e-book from N reading platforms, as well as N initial large language models with the same structure, and to train each initial large language model separately using the first data sample of each reading platform to obtain N first question-answering models; where N is an integer greater than 1, and the first data sample includes a set of historical question texts and historical answer texts;

[0017] The data selection module is used to extract the derivation process vector representation of the mature large language model from a preset mature large language model based on multiple first data samples from N reading platforms, and to form a second data sample based on each first data sample and the corresponding derivation process vector representation, so as to obtain multiple second data samples; wherein, the mature large language model is an open source large language model;

[0018] The second training module is used to construct a second question-answering model based on N first question-answering models and a fusion model, and to perform augmentation training on the second question-answering model based on multiple second data samples to obtain the augmented second question-answering model; the fusion model is used to select one prediction result from the N prediction results corresponding to the N first question-answering models as the prediction result of the second question-answering model;

[0019] The question-answering prediction module is used to input the current question text of the e-book into the second question-answering model that has been enhanced and trained, so that the prediction result output by the second question-answering model after enhancement training is used as the current answer text of the current question text.

[0020] A third aspect of this application provides an electronic device including at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform a large-model-based question-answering result search method as described in the first aspect above.

[0021] A fourth aspect of this application provides a readable storage medium storing computer-executable instructions for causing a computer to perform a large-model-based question-answering result search method as described in the first aspect above.

[0022] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a question-answering result search method based on a large model provided in an embodiment of this application;

[0025] Figure 2 This is a schematic diagram of the structure of a question-answering result search system based on a large model provided in an embodiment of this application;

[0026] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0029] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., indicating the orientation or positional relationship according to the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed or function in a specific orientation, and therefore should not be construed as a limitation of this application.

[0030] With the rapid development of Artificial Intelligence (AI) technology, large-scale language models (such as the GPT series, LLaMA, and DeepSeek series) have demonstrated outstanding performance in natural language processing tasks such as intelligent question answering, text summarization, and dialogue generation. These models, pre-trained on massive amounts of text data, have acquired powerful language understanding and generation capabilities. Especially in the field of e-book reading, users often need to ask questions about the book's content; an efficient question-answering system can greatly enhance the reading and learning experience.

[0031] Currently, there are two main ways to implement question-answering systems based on large models:

[0032] 1) Directly using mature LLMs (such as ChatGPT and DeepSeek-R1) to answer user questions has the advantage of strong model versatility and broad knowledge coverage, but for questions and answers in specific domains (such as the detailed content of an e-book), the answers may not be accurate enough and may easily lead to illusions.

[0033] 2) Fine-tuning a small or basic LLM for domain-specific data (such as question-answer pairs from a single ebook on a single reading platform) to obtain a dedicated question-answering model may perform well on that specific dataset, but it is prone to overfitting and has poor generalization ability. When users come from different platforms, even if they ask questions about the same ebook, their question expression habits and answer preferences may differ, leading to unstable performance of a single model.

[0034] like Figure 1 One embodiment of this application provides a question-answering result search method based on a large model, the method including the following steps S110 to S140;

[0035] Step S110: Obtain multiple first data samples of the same e-book from N reading platforms, and N initial large language models with the same structure. Train each initial large language model separately using the first data samples from each reading platform to obtain N first question-answering models. Here, N is an integer greater than 1.

[0036] In this embodiment, the reading platform includes, but is not limited to: QQ Reading, Qimao, Qidian, etc.

[0037] E-books refer to electronic books published on various reading platforms, such as "The Three-Body Problem" on JD.com's reading platform.

[0038] The first data sample includes a set of historical question texts and their corresponding historical answer texts. Both the historical question texts and historical answer texts are data collected from various platforms at historical moments, for example:

[0039] The historical question from a user on the First Reading platform was: "Is the 'Dark Forest' theory really valid?"; the historical answer was: Its validity depends on two axioms and two concepts (a chain of suspicion and technological explosion). The key is that the 'chain of suspicion' cannot be broken.

[0040] The historical question text from users of the second reading platform was: "Is the 'Dark Forest' theory really valid?"; the historical answer text was: "Within the framework set by the novel, the 'Dark Forest' theory is logically consistent and impeccable; it is an extremely powerful thought experiment."

[0041] It is evident that different reading platforms provide different answer texts for the same question.

[0042] Each reading platform offers a large number of question-and-answer pairs about the same ebook, but these questions and answers may have different language styles and emphases depending on the platform's user base.

[0043] This step aims to train N "first-response models," each of which fully learns the question-and-answer patterns (such as language style and emphasis) of users on a specific reading platform. The initial large language model here can be an untrained large language model (such as an LLM model).

[0044] Step S120: Based on multiple first data samples from N reading platforms, extract the derivation process vector representation of the mature large language model from the preset mature large language model, and form a second data sample based on each first data sample and the corresponding derivation process vector representation to obtain multiple second data samples.

[0045] Among them, mature large language models are open-source large language models, such as ChatGPT, LLaMA, and DeepSeek. This step leverages the powerful inference capabilities of mature large models to process historical questions in the first data sample. It not only requires them to derive the answers to the questions, but more importantly, to extract the "inference process vector representation" within the mature large model. This inference process vector representation is a high-dimensional vector that characterizes the model's reasoning path and thought process, containing the process of how the mature large model thinks and derives its conclusions. The following describes a process for extracting the inference process vector representation:

[0046] The mature large language model used is DeepSeek-R1. It can utilize the grouping relative policy optimization within the large model, that is, first sampling multiple answers (i.e., the responses given by the model) from the policy model, then calculating the advantage value based on the relative quality of the answers, and finally generating an inference vector with verifiable rewards. It should be noted that, due to the different mature large models used, their internal processes for generating inference vector representations also differ. However, the process of extracting these vectors is common knowledge in the field and is not the focus of this embodiment; therefore, it will not be described in detail here.

[0047] This embodiment extracts the reasoning path and thought process of the model by analyzing responses to historical question texts of the first data sample across different reading platforms. This process vector is then combined with the original first data sample to form a higher-quality second data sample. These new samples contain knowledge about how to think.

[0048] Step S130: Construct a second question-answering model based on N first question-answering models and a fusion model, and perform augmented training on the second question-answering model based on multiple second data samples to obtain the augmented training second question-answering model.

[0049] The fusion model is a decision-making module whose function is to select the optimal prediction from the N predictions (i.e., answers to the question) provided by the N first question-answering models as the final prediction. After generating high-quality second data samples in the above steps, the second question-answering model, which integrates the N first question-answering models and the fusion model, undergoes augmented training. During training, historical question texts and newly introduced "derivation process vector representations" are input to each first question-answering model. Then, based on the fusion model, it learns how to select the optimal answer, and the parameters of the entire second question-answering model are optimized through backpropagation of the loss function.

[0050] Step S140: Input the current question text of the e-book into the second question-answering model that has been augmented and trained, so that the prediction result output by the second question-answering model that has been augmented and trained is used as the current answer text of the current question text.

[0051] During the deployment phase, when a user asks a question about an ebook (regardless of the platform), the question is simply input into the pre-trained second question-answering model. This model internally invokes the various first question-answering models to make predictions, and the fusion model makes the final decision, outputting an accurate and reliable answer.

[0052] Further, in step S120, the derivation process vector representation of the mature large language model is extracted from the preset mature large language model, and a second data sample is formed based on each first data sample and the corresponding derivation process vector representation, specifically including the following steps S1210 to S1240:

[0053] Step S1210: Extract historical question text from the first data sample.

[0054] Step S1220: Input the historical question text into the mature large language model to obtain the first answer text and the derivation process vector representation output by the mature large language model based on the derivation mode, and the second answer text output based on the non-derivation mode.

[0055] Taking DeepSeek-R1 as an example, it can use the inference mode (i.e., thinking mode) to require the model to output a vector representation of its internal inference chain or attention weights along with the answer; the non-inference mode is the standard direct answer generation mode.

[0056] Step S1230: If the similarity between the first answer text and the second answer text is greater than a preset first threshold, determine the historical question text and the corresponding derivation process vector representation.

[0057] Only when the answers in the two models are inconsistent can it be proven that the input first data sample can reflect the effect of the derivation mode of the mature large language model. When the answers in the two models are consistent, it is proven that the derivation mode of the mature large language model has no value for searching the answer text corresponding to the historical question text in the first data sample. Therefore, in this embodiment, only the first data sample with a similarity greater than a preset first threshold between the first answer text and the second answer text is selected, and the first data sample is combined with the corresponding derivation process vector representation to form the second data sample. The purpose of this is to retain the valuable derivation process vector representation and filter out the worthless derivation process vector representation.

[0058] The preferred similarity metric here is text cosine similarity.

[0059] Step S1240: The historical question text, the corresponding derivation process vector representation, and the corresponding historical answer text are combined to form the second data sample.

[0060] The structure of the new sample is: <historical question text, derivation process vector representation, historical answer text>.

[0061] Further, in step S130, the second question-answering model is enhanced and trained based on multiple second data samples to obtain the enhanced second question-answering model, specifically including the following steps S1310 to S1330:

[0062] Step S1310: Input the historical question text and derivation process vector representation in each second data sample into N first question-answering models respectively, so that the N first question-answering models output the corresponding N predicted answer texts respectively.

[0063] Step S1320: Determine the similarity between the N predicted answer texts and their corresponding historical answer texts, and select the predicted answer text with the highest similarity as the prediction result of the second question answering model.

[0064] This step simulates the decision-making process of the fusion model.

[0065] In step S1330, the loss value is calculated based on the predicted answer text with the highest similarity and the prediction result, and the loss is used for backpropagation until the second question-answering model with enhanced training is obtained.

[0066] Through continuous iterative training, the fusion model learns how to better select the most suitable answer from the first question-answering model, thereby minimizing the overall loss.

[0067] In some embodiments, the fusion model can be a neural network module based on an attention mechanism. The fusion model takes as input the answer embeddings from multiple first-level question-answering models and the derivation process vector V generated by a mature large-scale model. By calculating attention weights, it performs a weighted summation of the answers from multiple first-level question-answering models. The resulting weighted summation vector is then decoded into the final answer text. This approach allows for a more nuanced fusion of the intelligence from multiple first-level question-answering models, further improving the accuracy of the answers.

[0068] Furthermore, before step S140, step S150 is also included:

[0069] Step S150: Preprocess the current problem text.

[0070] Preprocessing includes at least basic cleaning (such as removing irrelevant characters, HTML tags, and normalized encoding) and text normalization (such as word segmentation, conversion to lowercase, and removal of stop words) to ensure that the data input to the model is clean and consistent.

[0071] This embodiment has at least the following beneficial effects:

[0072] This method first trains N first question-answering models using first sample data composed of question-answer pairs from various reading platforms for e-books. Each first question-answering model fully learns the question-answering patterns of users on a specific reading platform. Then, it extracts the derivation process vector representation of the mature large language model from a pre-set mature large language model, and combines the derivation process vector representation with the first sample data to form second sample data. This second sample data contains knowledge of how the mature large language model thinks. Finally, it uses the second sample data with the thinking ability of the mature large language model to enhance the training of the second question-answering model based on the combination of N first question-answering models and the fusion model, so as to improve the accuracy of the second question-answering model in searching for the current answer text of the current question text.

[0073] This embodiment uses user question and answer data of "The Three-Body Problem" on three different reading platforms (let's say platform A, platform B, and platform C) as an example to explain in detail the implementation of this method.

[0074] Step S910: Train the first question-answering model;

[0075] Collect all historical question-and-answer pairs (historical question texts and their corresponding historical answer texts) about "The Three-Body Problem" from reading platforms A, B, and C. Each question-and-answer pair is a first data sample. Assume that platforms A, B, and C each have 4800 pairs.

[0076] Prepare three initial large language models with the same structure (e.g., all of which are open source LLaMA-7B models).

[0077] Independent training was conducted: the first LLaMA model was fine-tuned using 4800 data samples from platform A to obtain the "first question-answering model"; the second LLaMA model was fine-tuned using 4800 data samples from platform B to obtain the "first question-answering model"; and the third LLaMA model was fine-tuned using 4800 data samples from platform C. Thus, N=3 first question-answering models were obtained.

[0078] Step S920: Construct the second data sample;

[0079] Selecting a mature large language model: In this example, ChatGPT (gpt-3.5-turbo) is selected as the model.

[0080] Iterate through all the first data samples (14400). For each sample, the historical question Q-old is:

[0081] Step S921: Input it into ChatGPT and request it to output the response A1 and the derivation process vector representation V in "derivation mode" via a specific API call or PromptEngineering (e.g., extract the average of its last few hidden states as a vector representation).

[0082] In step S922, Q is input into ChatGPT, which is then asked to output the answer A2 directly in "standard mode".

[0083] Step S923: Calculate the text cosine similarity between answer A1 and answer A2. Set a threshold. =0.85. If the similarity is greater than 0.85, then ChatGPT's reasoning on this problem is considered stable and reliable, and its derivation variable V is of high quality.

[0084] Step S924: Reassemble this high-quality data into new samples (Q-new, V, A-new) as a second data sample. Assume that 12,000 high-quality second data samples are ultimately selected.

[0085] Step S930: Construct and enhance the training of the second question-answering model;

[0086] The second question-answering model is an ensemble model. It consists of two parts:

[0087] 1) The first question-answering model that has already been trained (with fixed parameters, not participating in training).

[0088] 2) A trainable fusion model. This fusion model can be a similarity-based selector or a complex neural network (such as an attention network). If it is a similarity-based selector, its input is the answer features and derivation process vector representation V output by the three first question-answering models, and its output is the answer selected from one of the first question-answering models. If it is a complex neural network (such as an attention network), its input is the answer features and derivation process vector representation V output by the three first question-answering models, and its output is the combined result of selecting the question and answer from the three first question-answering models.

[0089] Enhanced training:

[0090] Take a second data sample (Q-new, V, A-new).

[0091] Inputting Q-new and V into the three first question-answering models respectively yields three predicted answers: Pred-A, Pred-B, and Pred-C.

[0092] The fusion model selects one answer as the final prediction Pred-Final based on V and information from the three predicted answers Pred-A, Pred-B, and Pred-C (e.g., selecting the predicted answer with the highest similarity to A-new).

[0093] Calculate the loss: Calculate the cross-entropy loss between the predicted answer Pred-Final and the actual answer A-new.

[0094] By using the backpropagation algorithm, only the parameters of the fusion model are updated, making the loss smaller and smaller. In other words, the fusion model is trained to learn how to make the best choice.

[0095] Repeat the above process until all 12,000 second data samples have been trained multiple times and the model converges.

[0096] Step S940: Deployment and Inference;

[0097] Now, a new user from platform B asked: "What is Luo Ji's Wallfacer Project in The Three-Body Problem?"

[0098] The text in question is preprocessed (basic cleaning and text normalization).

[0099] The preprocessed question Q is input into the second question-answering model after augmentation training.

[0100] Internal workings of the model:

[0101] a. The three first-response question-answering models (A, B, C) each generate their own answers based on Q.

[0102] b. The fusion model weighs the reliability of the three answers based on the features learned from Q (for example, it may learn that the expert model of platform B answers the questions about the plot of "The Three-Body Problem" best), and finally selects the answer generated by the first question-answering model of B as the output.

[0103] The final selected answer is returned to the user: "Luo Ji's Wall-Facing Plan is...".

[0104] like Figure 2 As shown in one embodiment of this application, a question-answering result search system based on a large model is provided. The system includes:

[0105] The first training module 1100 is used to acquire multiple first data samples of the same e-book from N reading platforms, as well as N initial large language models with the same structure, and to train each initial large language model separately with the first data samples of each reading platform to obtain N first question-answering models; where N is an integer greater than 1, and the first data samples include a set of historical question texts and historical answer texts;

[0106] The data selection module 1200 is used to extract the derivation process vector representation of the mature large language model from the preset mature large language model based on multiple first data samples from N reading platforms, and to form a second data sample based on each first data sample and the corresponding derivation process vector representation, so as to obtain multiple second data samples; wherein, the mature large language model is an open source large language model.

[0107] The second training module 1300 is used to construct a second question answering model based on N first question answering models and a fusion model, and to perform augmented training on the second question answering model based on multiple second data samples to obtain the augmented training second question answering model; the fusion model is used to select one prediction result from the N prediction results corresponding to the N first question answering models as the prediction result of the second question answering model;

[0108] The question-answering prediction module 1400 is used to input the current question text of the e-book into the second question-answering model that has been enhanced and trained, so that the prediction result output by the second question-answering model after enhancement training is used as the current answer text of the current question text.

[0109] It should be noted that the question-answering result search system based on a large model in this application and the aforementioned question-answering result search method based on a large model are based on the same inventive concept. Therefore, the question-answering result search system based on a large model in this application corresponds to the aforementioned question-answering result search method based on a large model, and will not be described again here.

[0110] like Figure 3 This application also provides an electronic device, which includes:

[0111] At least one memory;

[0112] At least one processor;

[0113] At least one program;

[0114] The program is stored in memory, and the processor executes at least one program to implement the above-described question-answering result search method based on a large model, as described in this disclosure.

[0115] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0116] The electronic devices according to embodiments of this application will now be described in detail.

[0117] Electronic devices include:

[0118] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0119] The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to perform a question-answering result search method based on a large model according to an embodiment of this disclosure.

[0120] The input / output interface 1800 is used to implement information input and output.

[0121] The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0122] Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900);

[0123] The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 communicate with each other within the device via bus 2000.

[0124] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the above-described question-answering result search method based on a large model.

[0125] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, which can be linked to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0126] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0127] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0129] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0130] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0131] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication links shown or discussed may be through some interfaces; the indirect coupling or communication links between apparatuses or units may be electrical, mechanical, or other forms.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0135] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A question-answering result search method based on a large model, characterized in that, The method includes: Multiple first data samples of the same e-book from N reading platforms are obtained, along with N initial large language models with identical structures. The first data samples from each reading platform are used to train each initial large language model separately to obtain N first question-answering models. Here, N is an integer greater than 1, and the first data samples include a set of historical question texts and historical answer texts. Based on multiple first data samples from N reading platforms, the derivation process vector representation of the mature large language model is extracted from a preset mature large language model, and a second data sample is formed based on each first data sample and the corresponding derivation process vector representation to obtain multiple second data samples; wherein, the mature large language model is an open-source large language model; A second question-answering model is constructed based on N first question-answering models and a fusion model, and the second question-answering model is enhanced and trained based on multiple second data samples to obtain the enhanced second question-answering model; the fusion model is used to select one prediction result from the N prediction results corresponding to the N first question-answering models as the prediction result of the second question-answering model; The current question text of the e-book is input into the second question-answering model that has been enhanced and trained, so that the prediction result output by the second question-answering model after enhancement training is used as the current answer text of the current question text.

2. The question-answering result search method based on a large model according to claim 1, characterized in that, The step of extracting the derivation process vector representation of the mature large language model from the preset mature large language model, and forming a second data sample based on each first data sample and the corresponding derivation process vector representation, includes: Extract historical question text from the first data sample; The historical question text is input into the mature large language model to obtain the first answer text and the derivation process vector representation output by the mature large language model based on the derivation mode, and the second answer text output based on the non-derivation mode. If the similarity between the first answer text and the second answer text is greater than a preset first threshold, the historical question text and the corresponding derivation process vector representation are determined. The historical question text, the corresponding derivation process vector representation, and the corresponding historical answer text are combined to form a second data sample.

3. The question-answering result search method based on a large model according to claim 2, characterized in that, The step of enhancing the training of the second question-answering model using multiple second data samples to obtain the enhanced second question-answering model includes: The historical question text and the derivation process vector representation in each of the second data samples are respectively input into N first question-answering models to obtain N predicted answer texts output by N first question-answering models respectively. Determine the similarity between N predicted answer texts and their corresponding historical answer texts, and select the predicted answer text with the highest similarity as the prediction result of the second question answering model; The loss value is calculated based on the predicted answer text with the highest similarity and the prediction result, and backpropagation is performed based on the loss value until the second question-answering model with enhanced training is obtained.

4. The question-answering result search method based on a large model according to claim 1, characterized in that, Before inputting the current question text of the e-book into the second question-answering model that has undergone enhanced training, the method further includes: The current problem text is preprocessed.

5. The question-answering result search method based on a large model according to claim 4, characterized in that, The preprocessing includes at least basic cleaning and text normalization.

6. The question-answering result search method based on a large model according to claim 1, characterized in that, The mature large language model mentioned is ChatGPT.

7. The question-answering result search method based on a large model according to claim 2, characterized in that, The similarity mentioned is text cosine similarity.

8. A question-answering result search system based on a large model, characterized in that, The system includes: The first training module is used to acquire multiple first data samples of the same e-book from N reading platforms, as well as N initial large language models with the same structure, and to train each initial large language model separately using the first data sample of each reading platform to obtain N first question-answering models; where N is an integer greater than 1, and the first data sample includes a set of historical question texts and historical answer texts; The data selection module is used to extract the derivation process vector representation of the mature large language model from a preset mature large language model based on multiple first data samples from N reading platforms, and to form a second data sample based on each first data sample and the corresponding derivation process vector representation, so as to obtain multiple second data samples; wherein, the mature large language model is an open source large language model; The second training module is used to construct a second question-answering model based on N first question-answering models and a fusion model, and to perform augmentation training on the second question-answering model based on multiple second data samples to obtain the augmented second question-answering model; the fusion model is used to select one prediction result from the N prediction results corresponding to the N first question-answering models as the prediction result of the second question-answering model; The question-answering prediction module is used to input the current question text of the e-book into the second question-answering model that has been enhanced and trained, so that the prediction result output by the second question-answering model after enhancement training is used as the current answer text of the current question text.

9. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform a large-model-based question-answering result search method as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium stores computer-executable instructions for causing a computer to perform a large-model-based question-answering result search method as described in any one of claims 1 to 7.

Images