Model training methods, information processing methods, and related equipment

By introducing superiority/inferiority information and argumentation information labels during the training process of the neural network model, the problem that the neural network model cannot accurately judge the superiority/inferiority of the answer is solved, and more efficient answer quality judgment and improved accuracy of the generated results are achieved.

CN123070291APending Publication Date: 2026-11-10TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510592522.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-08
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing neural network models cannot accurately capture the relationship between multiple answers to the same question, resulting in low accuracy of the generated results after training.

Method used

By acquiring sample questions, multiple sample answers, superiority/inferiority information labels, and argumentation information labels from the training sample set, and utilizing the differences between predicted superiority/inferiority information and its labels, as well as the differences between predicted argumentation information and its labels, the prediction loss is determined, and the information processing model is trained to improve the accuracy of answer superiority/inferiority judgment.

Benefits of technology

It improves the accuracy of the information processing model in judging the quality of answers, optimizes the model's training effect and generalization ability, enabling it to more accurately capture the key elements of the answer and improve the accuracy of the generated results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070291A_ABST
    Figure CN123070291A_ABST
Patent Text Reader

Abstract

This application discloses a model training method, an information processing method, and related equipment. The application obtains an information processing model to be trained and a training sample set. Using the information processing model to be trained, the sample questions and sample answers are processed to obtain predicted superiority / inferiority information and predicted argumentation information. The predicted superiority / inferiority information includes predicted superiority / inferiority relationships corresponding to the relation labels. Based on the differences between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the differences between the predicted argumentation information and the argumentation information labels, a prediction loss is determined. The prediction loss is used to train the information processing model to be trained, resulting in a trained information processing model. This application enables the trained information processing model to more accurately judge the quality of answers, thereby improving the accuracy of the trained information processing model in predicting the superiority / inferiority of answers.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a model training method, an information processing method, and related equipment. Background Technology

[0002] In the application of neural network models based on question-answering information, situations often arise where multiple answers address the same question. In existing technologies, preference learning is typically introduced during model training to help the neural network model identify the quality of multiple answers. Through preference learning, the neural network model can analyze the quality of these answers, thereby determining which answers are more appropriate and accurate.

[0003] However, during model training, preference learning typically relies on rating differences or preference rankings between different answers to enable the neural network model to learn to identify the merits of answers. Because the information contained in rating differences or preference rankings is relatively simple, the neural network model cannot accurately capture the relationships between answers when analyzing their merits. This makes it difficult for the trained model to accurately judge the merits of answers, resulting in low accuracy in the generated results. Summary of the Invention

[0004] This application provides a model training method, an information processing method, and related equipment, which enable the trained information processing model to more accurately judge the quality of answers, thereby improving the accuracy of the trained information processing model in predicting the quality of answers.

[0005] This application provides a model training method, comprising: acquiring an information processing model to be trained and a training sample set, wherein the training sample set includes a sample question, multiple sample answers to the sample question, superiority / inferiority information labels, and argumentation information labels, wherein the superiority / inferiority information labels include relation labels representing the superiority / inferiority relationship between the sample answers, and the argumentation information labels are information labels that argue the superiority / inferiority relationship; processing the sample question and the sample answers through the information processing model to be trained to obtain predicted superiority / inferiority information and predicted argumentation information, wherein the predicted superiority / inferiority information includes the predicted superiority / inferiority relationship corresponding to the relation labels; determining a prediction loss based on the difference between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the difference between the predicted argumentation information and the argumentation information labels; and training the information processing model to be trained using the prediction loss to obtain a trained information processing model.

[0006] This application provides an information processing method, comprising: acquiring a trained information processing model, a problem to be processed, and multiple answers to the problem to be processed, wherein the trained information processing model is obtained by any model training method provided in this application; and processing the problem to be processed and the answers to be processed through the trained information processing model to obtain target superiority information, wherein the target superiority information includes a target superiority relationship used to represent the superiority relationship between the answers to be processed.

[0007] This application embodiment also provides a model training apparatus, comprising: a training acquisition unit, configured to acquire an information processing model to be trained and a training sample set, the training sample set including sample questions, multiple sample answers to the sample questions, superiority / inferiority information labels, and argumentation information labels, the superiority / inferiority information labels including relation labels representing the superiority / inferiority relationships between the sample answers, and the argumentation information labels being information labels that argue the superiority / inferiority relationships; a sample processing unit, configured to process the sample questions and sample answers through the information processing model to be trained to obtain predicted superiority / inferiority information and predicted argumentation information, the predicted superiority / inferiority information including the predicted superiority / inferiority relationships corresponding to the relation labels; a loss determination unit, configured to determine a prediction loss based on the difference between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the difference between the predicted argumentation information and the argumentation information labels; and a training unit, configured to train the information processing model to be trained using the prediction loss to obtain a trained information processing model.

[0008] In some embodiments, the sample processing unit includes a first feature processing subunit and a second feature processing subunit, comprising: the first feature processing subunit being used to perform first feature processing on the sample to be processed through the information processing model to be trained, to obtain a first original score of the sample to be processed, wherein the sample to be processed includes the sample question and the sample answer, and each sample to be processed corresponds to one sample answer; by comparing the first original scores, the prediction superiority relationship between the sample answers in the sample to be processed is obtained; the second feature processing subunit being used to perform second feature processing on the sample to be processed through the information processing model to be trained, to obtain a second original score of the sample to be processed; and the prediction argument information is determined by the second original score.

[0009] In some implementations, obtaining the predicted superiority relationship between the sample answers in the samples to be processed by comparing the first original scores includes: for any two samples to be processed, determining a positive sample and a negative sample from the two samples to be processed based on the relationship label; calculating the difference between the first original score of the positive sample and the first original score of the negative sample to obtain the predicted superiority relationship between the sample answers in the two samples to be processed.

[0010] In some implementations, the superiority / inferiority information label further includes a superiority / inferiority category label, and the predicted superiority / inferiority information further includes the category prediction probability corresponding to the superiority / inferiority category label. The first feature processing subunit is further configured to: perform probability transformation processing on the first original score to obtain the category prediction probability.

[0011] In some implementations, the predicted loss includes superiority / inferiority information loss and argumentation information loss. The loss determination unit includes a first loss determination subunit, a second loss determination subunit, and a third loss determination subunit, comprising: the first loss determination subunit, configured to determine the superiority / inferiority information loss based on the difference between the predicted superiority / inferiority information and the superiority / inferiority information label; the second loss determination subunit, configured to determine the argumentation information loss based on the difference between the predicted argumentation information and the argumentation information label; and the third loss determination subunit, configured to combine the superiority / inferiority information loss and the argumentation information loss to obtain the predicted loss.

[0012] In some implementations, the predicted argument information and the argument information label are text information, and the second loss determination subunit is specifically used to: determine the word difference between the predicted word and the target word in the predicted argument information, wherein the target word is the word corresponding to the predicted word in the argument information label; and determine the argument information loss based on the word difference.

[0013] In some embodiments, the model training apparatus further includes a threshold determination unit, which is configured to: calculate the standard deviation of the sample to be processed using the first original score; determine the target parameter value corresponding to the sample to be processed based on the standard deviation, wherein the target parameter value is used to adjust the prediction quality information corresponding to the sample to be processed, so as to use the adjusted prediction quality information to determine the prediction loss.

[0014] In some embodiments, the model training apparatus further includes a reversal unit, which is used to: determine the cumulative number of deviations corresponding to the sample answer, the cumulative number of deviations representing the cumulative number of deviations between the predicted superiority / inferiority information corresponding to the sample answer and the corresponding superiority / inferiority information label; when the cumulative number of deviations corresponding to the sample answer exceeds the preset number, reversing the superiority / inferiority information label corresponding to the sample answer to obtain a reversed label, the reversed label being used to determine the prediction loss.

[0015] In some embodiments, the model training apparatus further includes a testing unit, which is configured to: acquire a test sample set, the test sample set including test questions, multiple test answers to the test questions, and test superiority / inferiority information labels, the test superiority / inferiority information labels including test relationship labels representing the superiority / inferiority relationships between the test answers; process the test questions and test answers through the trained information processing model to obtain test predicted superiority / inferiority information, the test predicted superiority / inferiority information including the test predicted superiority / inferiority relationships corresponding to the test relationship labels; and determine an evaluation index score for the trained information processing model based on the difference between the test predicted superiority / inferiority information and the test superiority / inferiority information labels, the evaluation index score being used to evaluate the performance of the trained information processing model.

[0016] This application embodiment also provides an information processing apparatus, comprising: acquiring a post-trained information processing model, a problem to be processed, and multiple unprocessed answers to the problem to be processed, wherein the post-trained information processing model is obtained by any model training method provided in this application embodiment; and processing the problem to be processed and the unprocessed answers through the post-trained information processing model to obtain target superiority information, wherein the target superiority information includes a target superiority relationship used to represent the superiority relationship between the unprocessed answers.

[0017] This application also provides an electronic device, including a processor and a memory, wherein the memory stores multiple instructions; the processor loads instructions from the memory to execute steps in any of the model training methods or information processing methods provided in this application.

[0018] This application also provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the model training methods or information processing methods provided in this application.

[0019] This application also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps in any of the model training methods or information processing methods provided in this application.

[0020] This application embodiment can obtain an information processing model to be trained and a training sample set. The training sample set includes a sample question, multiple sample answers to the sample question, superiority / inferiority information labels, and argumentation information labels. The superiority / inferiority information labels include relation labels representing the superiority / inferiority relationship between the sample answers, and the argumentation information labels are information labels that demonstrate the superiority / inferiority relationship. The information processing model to be trained processes the sample question and the sample answers to obtain predicted superiority / inferiority information and predicted argumentation information. The predicted superiority / inferiority information includes the predicted superiority / inferiority relationship corresponding to the relation labels. Based on the difference between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the difference between the predicted argumentation information and the argumentation information labels, a prediction loss is determined. The prediction loss is used to train the information processing model to be trained to obtain a trained information processing model.

[0021] In this application, the predictive loss of the information processing model to be trained is constructed by combining the differences between predicted superiority / inferiority information and labels, as well as the differences between predicted argumentation information and labels. This ensures that the training process is not limited to focusing solely on the superiority / inferiority information of the final answer, such as superiority / inferiority relationships, but also considers the argumentation logic of the answer's superiority / inferiority information based on the argumentation information. Specifically, this argumentation information provides the model with additional feature information, enabling the model to more accurately capture the key elements that need to be focused on when comparing the superiority / inferiority of answers. This optimizes the model's understanding and processing of answer quality, allowing the model to learn more efficiently and improve its overall performance, thereby enhancing the model's training effect and generalization ability. Consequently, the trained information processing model can more accurately judge the superiority / inferiority of answers, thus improving the accuracy of the trained information processing model in predicting the superiority / inferiority of answers. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1a This is a schematic diagram of a scenario illustrating the model training method and information processing method provided in the embodiments of this application;

[0024] Figure 1b This is a flowchart illustrating the model training method provided in the embodiments of this application;

[0025] Figure 1c This is a schematic diagram of the sample processing procedure provided in the embodiments of this application;

[0026] Figure 1d This is a schematic diagram of the model training process provided in the embodiments of this application;

[0027] Figure 1e This is a schematic diagram illustrating the processing steps of the classification and sorting tasks provided in the embodiments of this application;

[0028] Figure 2a This is a flowchart illustrating the information processing method provided in an embodiment of this application;

[0029] Figure 2b This is a schematic diagram of the information processing process provided in the embodiments of this application;

[0030] Figure 3a This is a flowchart illustrating the model training method provided in the embodiments of this application;

[0031] Figure 3b This is a schematic diagram of the comparison table of trained model performance provided in the embodiments of this application;

[0032] Figure 4 This is a schematic diagram of the structure of the model training device provided in the embodiments of this application;

[0033] Figure 5 This is a schematic diagram of the structure of the information processing device provided in the embodiments of this application;

[0034] Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] This application provides a model training method, an information processing method, and related equipment.

[0037] Specifically, the model training device or information processing device can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, or personal computer (PC); the server can be a single server or a server cluster consisting of multiple servers.

[0038] In some embodiments, the model training device or information processing device may also be integrated into multiple electronic devices. For example, the model training device or information processing device may be integrated into multiple servers, and the model training method or information processing method of this application may be implemented by multiple servers.

[0039] In some embodiments, the terminal can also be used as a server to implement some or all of the functions of a server.

[0040] For example, refer to Figure 1a Taking the joint execution of the model training method and information processing method by a terminal and a server as an example, the model training and information processing system provided in this application includes a terminal and a server, etc.; the terminal and the server are connected through a network, such as through a wired or wireless network.

[0041] The terminal can be used to: acquire a user-input question and multiple answers to that question, send the acquired question and answers to the server for processing, and receive target quality information returned by the server. The terminal can include mobile phones, vehicle-mounted terminals, aircraft, tablets, laptops, or personal computers (PCs), etc. A client can also be installed on the terminal, which can be an application client or a browser client, etc.

[0042] The server can be used to: acquire the information processing model to be trained and the training sample set, which includes sample questions, multiple sample answers to the sample questions, superiority / inferiority information labels, and argumentation information labels. The superiority / inferiority information labels include relation labels representing the superiority / inferiority relationships between sample answers, and the argumentation information labels are information labels representing the superiority / inferiority relationships of arguments. The information processing model to be trained processes the sample questions and sample answers to obtain predicted superiority / inferiority information and predicted argumentation information. The predicted superiority / inferiority information includes predicted superiority / inferiority relationships corresponding to the relation labels. Based on the differences between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the differences between the predicted argumentation information and the argumentation information labels, the prediction loss is determined. The prediction loss is used to train the information processing model to be trained, resulting in the trained information processing model.

[0043] The server can also be used to: acquire a trained information processing model, a problem to be processed, and multiple answers to the problem to be processed; process the problem to be processed and the answers to be processed using the trained information processing model to obtain target superiority information, which includes target superiority relationships used to represent the superiority relationships between the answers to be processed. The server can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0044] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0045] The following sections provide detailed descriptions. It should be noted that the order of the following embodiments is not intended to limit the preferred order of the embodiments. It is understood that in the specific embodiments of this application, user-related data is involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0046] In this embodiment, a model training method is provided, such as Figure 1b As shown, the specific process of this model training method can be as follows:

[0047] 110. Obtain the information processing model to be trained and the training sample set. The training sample set includes sample questions, multiple sample answers to the sample questions, good and bad information labels, and argument information labels.

[0048] The information processing model to be trained refers to the information processing model that needs to be trained. This information processing model is used to process the input question and multiple answers to the question in order to predict the quality of the multiple answers and / or the corresponding argumentation information.

[0049] The training sample set refers to the collection of samples and their labels used to train the information processing model to be trained. The training sample set may include at least one question (i.e., a sample question) and multiple answers to each sample question (i.e., sample answers). The sample labels in the training sample set may include labels indicating merit or demerit and labels indicating argumentation. In this embodiment, the sample questions and sample answers may be presented in one or more combined forms such as text, images, audio, or video.

[0050] The superiority / inferiority information label includes relationship labels used to indicate the superiority / inferiority relationship between sample answers. Superiority / inferiority information refers to information used to indicate the superiority / inferiority of sample answers. Superiority / inferiority information labels are labels indicating the superiority / inferiority of sample answers. In this embodiment, superiority / inferiority information includes superiority / inferiority relationships, which refer to the comparative relationship between different sample answers. Specifically, in a given set of sample answers, the superiority / inferiority relationship refers to the differences and relative comparisons between the various sample answers in terms of quality, accuracy, completeness, and validity. For example, for any set of sample answers y1 and y2, the superiority / inferiority relationship could be that answer y1 is superior to answer y2 in terms of quality, and / or accuracy, and / or completeness, and / or validity.

[0051] In this context, "argumentation information" refers to evidence or explanations used to support or prove the superiority or inferiority relationship between sample answers. Argumentation information can include analysis, data support, logical reasoning, and other content related to this relationship. For example, argumentation information can be the reasons and evidence that clarify the superiority or inferiority relationship. In this embodiment, the argumentation information label is a label composed of the argumentation information regarding the superiority or inferiority relationship between sample answers; that is, the argumentation information regarding the superiority or inferiority relationship between sample answers is used as a type of label in model training. For example, for any set of sample answers y1 and y2, the corresponding argumentation information label can be the argumentation information regarding that superiority or inferiority relationship, such as "The reason why answer y1 is better than answer y2 is that answer y1 is more consistent with the core requirements of the question and its expression is more accurate."

[0052] For example, the model training method of this application embodiment can be applied to various information processing scenarios. For instance, it can be used for information processing models used in question-and-answer scenarios, such as question-and-answer systems, knowledge base question-and-answer systems, automatic question-and-answer in search engines, and virtual dialogue assistants. By acquiring the information processing model to be trained and a training sample set, the information processing model is trained so that the trained information processing model can provide responses to question-and-answer scenarios with information on the quality of answers and / or corresponding argumentation information.

[0053] In this embodiment of the application, the information processing model to be trained can be a model with various structures, such as a purely discriminative encoder, an encoder-decoder, a purely generative decoder, etc., which can be selected according to actual needs or application scenarios.

[0054] In some implementations, the information processing model to be trained is a language model, such as an autoregressive language model, a masked modeling language model, a discriminative language model (such as a large language model), and so on. Specifically, it can be a generative pre-trained transformer (GTP), a bidirectional transformer model (BERT), a Roberta model (RoBERTa), a deep seeker, etc. The specific language model used by the information processing model to be trained in this application embodiment is not limited, and can be selected according to actual needs or application scenarios.

[0055] In some implementations, multiple sample answers to a sample question may come from human responses to the sample question, or from multiple sample answers generated by inputting the sample question into one or more dialogue models.

[0056] In some implementations, the relationship labels and / or argument information labels are descriptive labels in text form. The relationship labels can use symbols, words, or keywords to describe the relative merits of multiple sample answers. The argument information labels can use text to express information about the relative merits of arguments.

[0057] For example, a relation label can be expressed as y1>y2>y3, or "y1 is more popular than y2". An argument information label can be expressed as "The reason why article y1 is more popular than article y2 is that article y1 has a clearer structure, more coherent logic, and uses more vivid and descriptive language, which can better attract readers' attention. At the same time, the viewpoints presented in article y1 are more novel and unique, and have higher inspirational value".

[0058] In some implementations, the superiority / inferiority information also includes classification information for the sample answer, where the category of the classification information is related to the superiority / inferiority of the sample answer. Therefore, the superiority / inferiority information label also includes a superiority / inferiority category label for the sample answer, which indicates the category to which the sample answer belongs. The superiority / inferiority category label can be in the form of numbers (e.g., 0, 1, 2, etc.), letters (e.g., A, B, C), symbols, or text (e.g., superior, inferior). One sample answer corresponds to one superiority / inferiority category label.

[0059] It is understood that in this embodiment, the superiority / inferiority category label is an absolute label determined based on an absolute classification standard (i.e., category), while the relationship label is a relative label determined by comparing the superiority / inferiority of sample answers. By using both absolute and relative labels during model training, the model can have clear category judgments in classification tasks, and can better analyze the relative superiority / inferiority between samples when dealing with complex relationship tasks, ultimately improving the overall performance and flexibility of the model. Absolute labels are markers used to accurately label and classify sample answers. They typically categorize each sample answer explicitly into a specific category or attribute based on predetermined standards or categories. Relative labels are a way of labeling sample answers based on the superiority / inferiority relationship between them. Unlike absolute labels, relative labels do not explicitly specify a fixed category or attribute for text elements; instead, they assign labels based on the relationship (i.e., superiority / inferiority relationship) between sample answers.

[0060] In some implementations, the superiority / inferiority category labels are binary classification labels. For example, sample answers can be divided into two categories: positive (e.g., representing 1) and negative (e.g., representing 0), where the positive category represents "superior" or "good," and the negative category represents "inferior" or "bad."

[0061] In some implementations, the classification task is a binary classification task, meaning the positive and negative category labels are binary classification labels. A binary classification task involves dividing sample answers into two categories, and the binary classification labels are the labels for these two categories: positive and negative. Thus, the first raw score can represent the category score for either the positive or negative class. The category prediction probability can represent the probability of either the positive or negative class occurring.

[0062] In some implementations, a sample problem corresponds to a good or bad category label, that is, there is a one-to-one correspondence between sample problems and good or bad category labels.

[0063] In some implementations, a sample question corresponds to at least one relation label and / or at least one argument information label. Each relation label or argument information label corresponds to any two sample answers from a plurality of sample answers to the sample question; that is, any relation label represents the superiority / inferiority relationship between any two sample answers, and any argument information label is an information label that argues the superiority / inferiority relationship between any two sample answers. Thus, during the processing of the information processing model to be trained, the sample answers can be processed pairwise to compare their superiority / inferiority relationships. For example, a sample question x and its corresponding responses (i.e., sample answers) are y1, y2, ..., y... n It is possible to create separate samples for each pair of answers. Relationship tags and Each argument information label. Taking n=3 as an example, three relation labels and corresponding three argument information labels can be created for y1 and y2, y1 and y3, and y1 and y3 respectively.

[0064] In this embodiment, the superiority / inferiority information tags and the argumentation information tags can be tags based on user preferences or tags based on objective facts. These tags can be manually labeled based on user preferences or generated using automated algorithms based on data analysis results.

[0065] In some implementations, relationship labels can be preference labels, which represent the relationships between sample answers based on the user's preferences for those answers. Preference labels can help understand the user's tendencies or interests, enabling the trained information processing model to provide personalized processing results.

[0066] 120. Using the information processing model to be trained, the sample questions and sample answers are processed to obtain the predicted superiority and inferiority information and the predicted argumentation information. The predicted superiority and inferiority information includes the predicted superiority and inferiority relationships corresponding to the relation labels.

[0067] Among these, the predicted superiority / inferiority information refers to the superiority / inferiority information of the sample answers inferred by the information processing model to be trained based on the sample questions and sample answers. The predicted superiority / inferiority relationship refers to the superiority / inferiority relationship between the sample answers inferred by the model. The predicted justification information refers to the justification information of the predicted superiority / inferiority relationship inferred by the information processing model to be trained based on the sample questions and sample answers.

[0068] For example, the process by which the information processing model to be trained processes the sample questions and sample answers is the process of prediction. This prediction process can be a classification prediction based on a classification task, a ranking prediction based on a ranking task, a regression prediction based on a generation task, and so on.

[0069] In this embodiment of the application, the prediction quality information and prediction argumentation information refer to the specific results of the prediction of the sample. They can be expressed as the probability of the sample answer belonging to a certain category or the most likely category label in classification prediction, or the difference or relative order between multiple sample answers in sequential prediction, or the continuous numerical output of the model prediction in regression prediction, etc.

[0070] In some implementations, the predicted superiority or inferiority information also includes the predicted superiority or inferiority category corresponding to the superiority or inferiority category label.

[0071] In this embodiment, during the process of the information processing model to be trained processing sample questions and sample answers, the model can execute processing tasks corresponding to different labels of the samples and obtain the processing results of the corresponding tasks. For example, for the superiority / inferiority category label, the model can perform a classification task on the sample answers (i.e., perform classification prediction) to obtain the predicted superiority / inferiority category corresponding to the superiority / inferiority category label. Similarly, for the relationship label, the model can perform a sorting task on the sample answers (i.e., perform order prediction) to obtain the predicted superiority / inferiority relationship corresponding to the relationship label. Furthermore, for the argumentation information label, the model can perform a generation task on the sample answers (such as regression prediction) to obtain the predicted argumentation information corresponding to the argumentation information label.

[0072] In some implementations, a sample question and a sample answer in the training sample set can be considered as a single sample. This allows the information processing model to combine the sample question with the predicted quality information and corresponding argumentation information of the sample answer during processing. This enables a better understanding and focus on the contextual information between the question and the answer, improving the accuracy of the prediction results. Specifically, the information processing model to be trained processes the sample question and sample answer to obtain predicted quality information and predicted argumentation information, including:

[0073] The first feature processing is performed on the sample to be processed by the information processing model to be trained, and the first original score of the sample to be processed is obtained. The sample to be processed includes the sample question and the sample answer, and each sample to be processed corresponds to a sample answer.

[0074] By comparing the first original scores, the predictive merits among the sample answers in the sample to be processed are obtained;

[0075] The second feature processing is performed on the sample to be processed by the information processing model to be trained, and the second original score of the sample to be processed is obtained.

[0076] The predictive argument information is determined by the second original score.

[0077] The first feature processing refers to the process of extracting category-related features from the sample to be processed using the information processing model to be trained. The second feature processing refers to the process of generating features from the sample to be processed based on the features obtained from the information processing model to be trained. In this embodiment, the first and second feature processing steps may include, but are not limited to, one or more of the following steps: preprocessing and standardization, feature extraction, encoding, and output layer processing.

[0078] Here, the first raw score is the original output of the information processing model to be trained after the first feature processing, i.e., the original predicted value. This first raw score can also be called the category logits, i.e., the category raw score, which represents the model's preference for each category.

[0079] The second raw score is the original output of the information processing model to be trained after the second feature processing. This second raw score can also be called the generated logits, which represents the model's preference for each possible generated word or symbol. The second raw score can be represented as a predicted value in the form of a continuous numerical prediction of the argument information.

[0080] For example, the training sample set may include at least one sample question x and multiple sample answers for each sample question. For instance, the sample answers to any sample question x in the training sample set might be y1, y2, ..., y... n For a given sample question x, it can be combined with the corresponding sample answers to form an input sequence x,y1,y2,…,y n The input information is processed by the training model. The input sequence is x, y1, y2, ..., y... n In the sample question x and any sample answer y, i That is, to form a sample z to be processed. i Therefore, for the information processing model to be trained, the input sequence x,y1,y2,…,y n It contains n samples to be processed, namely z1, z2, ..., z n .

[0081] The information processing model to be trained can perform first feature processing on the n samples to be processed to obtain the first raw score for each sample. Specifically, in the first feature processing, features can be extracted from the samples to be processed through feature extraction, and then the extracted features are encoded and output to output the first raw score, i.e., the category logits. For ranking tasks, the first raw scores of the n samples to be processed can be compared to obtain the prediction superiority relationship between the sample answers in the samples to be processed, i.e., the prediction superiority relationship can be the difference in the first raw scores of the samples to be processed.

[0082] It is understandable that for any two samples to be processed, such as sample z... i and z j The difference between their first raw scores is logits. i -logits j If the difference is positive, it indicates that the sample z to be processed is positive. i Superior to z j , i.e. z i >z j That is, the corresponding sample answer yi >y j If the difference is negative, it indicates that the sample z to be processed is... j Superior to z i , i.e. z i <z j That is, the corresponding sample answer y i <y j Therefore, this difference can represent the predictive quality relationship between the answers to be processed in the sample to be processed.

[0083] For the generation task, the information processing model to be trained can perform first feature processing on the n samples to be processed to obtain the second original score of the samples. Specifically, in the second feature processing, features can be extracted from the samples to be processed through feature extraction, and then the extracted features are encoded and output to output the second original score, i.e., generate logits. The second original score can be converted into a probability between 0 and 1, i.e., the predicted generation probability, through functions such as softmax or sigmoid, and this predicted generation probability is used as the prediction argument information.

[0084] In some implementations, the argument information is textual information, and the model in the generation task predicts the probability of the next word based on the preceding word sequence. Therefore, the predicted generation probability can include the probability of each word in the generated word sequence.

[0085] It should be noted that since each sample to be processed corresponds to one sample answer, the prediction quality and argumentation information of the sample to be processed are the same as the prediction quality and argumentation information of the corresponding sample answer. Therefore, the prediction information obtained by the information processing model to be trained from the sample to be processed can also directly represent the prediction information of the corresponding sample answer.

[0086] In some implementations, in a ranking task, the first raw scores of each pair of samples to be processed can be compared to determine the predicted superiority relationship between the sample answers. Thus, the predicted superiority relationship is the superiority relationship between each pair of sample answers among multiple sample answers. For example, as... Figure 1c The sample processing procedure shown is used in a sorting task for n samples z1, z2, ..., zn. n You can select any two samples to be processed, such as z. i and z j The difference between the first original scores of the two samples is calculated to obtain the predicted superiority relationship between any two samples to be processed. This predicted superiority relationship represents the superiority relationship between the sample answers in the two samples to be processed. By analogy, the predicted superiority relationships between all pairwise combinations of sample answers in multiple sample answers can be obtained.

[0087] In some implementations, during the generation task, a second feature processing can be performed on each pair of samples to be processed to generate predictive argument information regarding the predicted superiority or inferiority relationship between the two samples. That is, the predictive argument information is the argument information corresponding to the predicted superiority or inferiority relationship between each pair of sample answers among multiple sample answers. For example, such as... Figure 1c The sample processing procedure shown in the diagram, in the generation task, involves processing n samples z1, z2, ..., z... n You can select any two samples to be processed, such as z. i and z j This generates argument information for the two samples to be processed. This predicted argument information can be directly represented as the argument information regarding the superiority or inferiority relationship between the sample answers in the two samples. Similarly, this can be used to obtain the predicted argument information for all pairwise combinations of sample answers from multiple sample answers.

[0088] In some implementations, for ranking tasks, relatively good and relatively bad samples can be determined from each pair of samples to be processed based on relational labels. The relative superiority or inferiority of each pair of samples is used to determine the predicted superiority or inferiority relationship between them, enabling accurate identification of the difference between the predicted result and the label based on the true label, thereby improving the model training effect. Specifically, by comparing the first original scores, the predicted superiority or inferiority relationship between the sample answers in the samples to be processed is obtained, including:

[0089] For any two samples to be processed, based on the relation labels, determine the positive and negative samples from the two samples to be processed;

[0090] The difference between the first original score of the positive sample and the first original score of the negative sample is calculated to obtain the predictive superiority relationship between the sample answers in any two samples to be processed.

[0091] In this context, a positive sample refers to the relatively better one of any two samples to be processed. A negative sample refers to the relatively worse one of any two samples to be processed. That is, in the superiority / inferiority relationship between the sample answers of any two samples to be processed, the sample answer in the positive sample is superior to that in the negative sample.

[0092] For example, for n samples z1, z2, ..., zn to be processed n You can select any two samples to be processed, such as z. i and z j If the relation labels corresponding to the two samples to be processed represent the superiority or inferiority relationship between them as z... i >z j Then z in these two samples i For positive samples, z j These are negative samples. Therefore, the first original score logits of the positive samples can be used.i Subtract the first original score logits of the negative samples j The difference in logits between the two is obtained, which represents the predictive superiority / inferiority relationship. i -logits j The predicted superiority / inferiority relationship represents the relative merits of the sample answers between the two samples to be processed. Similarly, the predicted superiority / inferiority relationships between all pairwise combinations of sample answers across multiple sample answers can be obtained.

[0093] In some implementations, the information processing model to be trained can also be used for classification tasks of the samples to be processed, in order to obtain the predicted class probability of the sample problem in each sample to be processed. For example, such as Figure 1c The sample processing procedure shown allows for the determination of the predicted class probability for each sample in a classification task.

[0094] Therefore, as Figure 1d The model training process shown can be performed by inputting sample questions and corresponding multiple sample answers to the information processing model to be trained, and jointly training the information processing model to be trained on classification, ranking and generation tasks. During the training process, the information processing model to be trained can simultaneously process these three tasks to obtain the category prediction probability, prediction superiority relationship and prediction argument information.

[0095] In some implementations, since both the superiority / inferiority category and the superiority / inferiority relationship are related to the superiority / inferiority of the samples, the first original score can be used simultaneously for both classification and ranking tasks. This allows the two tasks to share the same original output from the model's inference, thereby improving computational efficiency and reducing model complexity. Specifically, the superiority / inferiority information label also includes superiority / inferiority category labels, and the predicted superiority / inferiority information also includes the category prediction probability corresponding to the superiority / inferiority category labels. Through the information processing model to be trained, the sample questions and sample answers are processed to obtain predicted superiority / inferiority information and predicted argument information, which also includes:

[0096] The first original score is subjected to probability transformation to obtain the category prediction probability.

[0097] The probability transformation process refers to converting the raw score into a predicted probability. For example, the first raw score can be converted into a probability between 0 and 1, i.e., the class prediction probability, using functions such as the softmax function or the sigmoid function. This class prediction probability can represent the probability of occurrence of each class predicted by the model.

[0098] For example, such as Figure 1eThe processing steps for the classification and ranking tasks shown can be used simultaneously for both tasks, applying the initial raw score obtained from the first feature processing to determine the predicted class probabilities and prediction rankings. For the classification task, the predicted class probability of each sample can be calculated using q = σ(logit). Here, the sigmoid function σ() converts the initial raw score logit into a probability for each class, which is between 0 and 1. This yields the predicted class probability of the answer for each of the n samples to be processed.

[0099] Because information processing models need to process different parts of the input in different tasks, such as the input sequence x,y1,y2,…,y for any sample problem x. n The generation task requires pairwise feature processing of each sample, while the ranking and classification tasks require individual feature processing of each sample. In some implementations, a masking mechanism can be used to allow the model to focus only on the specific information to be processed in different tasks. For example, a mask can be used to obscure parts of the input sequence that do not need to be processed, or to obscure features that do not need to be processed during feature computation.

[0100] In this embodiment, the information processing model can be used to perform multiple tasks, such as classification, ranking, and generation tasks. Therefore, for different tasks, parts of the network in the model can be shared, or separate networks can be added for different tasks, enabling the information processing model to perform multiple tasks simultaneously on a single input. For example, in some implementations, different tasks in the information processing model to be trained can share the network used for feature extraction and encoding, but use different output heads. For instance, the information processing model to be trained may include a classification task head corresponding to a classification task and a generation task head corresponding to a generation task.

[0101] 130. Determine the prediction loss based on the difference between the predicted superior and inferior information and the superior and inferior information labels, as well as the difference between the predicted argument information and the argument information labels.

[0102] In this context, prediction loss refers to the loss that measures the difference between the predicted information obtained by the information processing model under training and the sample labels. This loss can be calculated using a loss function, which can be one or more combinations of mean squared error loss function, cross-entropy loss function, and perceptual loss function.

[0103] For example, for multiple sample answers in the training sample set, we can calculate the difference between the predicted superiority / inferiority information and the superiority / inferiority information label for the same sample answer, and the difference between the predicted argumentation information and the argumentation information label for the same sample answer. Combining these calculated differences, we obtain the overall loss of the information processing model to be trained, i.e., the prediction loss.

[0104] Therefore, in this embodiment, the predictive loss of the information processing model to be trained is constructed by combining the differences between the predicted superiority / inferiority information and the labels, as well as the differences between the predicted argumentation information and the labels. This ensures that the training process is not limited to focusing solely on the superiority / inferiority information of the final answer, such as the superiority / inferiority relationship, but also considers the argumentation logic of the answer's superiority / inferiority information based on the argumentation information. Specifically, this argumentation information provides the model with additional feature information to assist it, enabling the model to more accurately capture the key elements that need to be focused on when comparing the superiority / inferiority of answers. This optimizes the model's understanding and processing of answer quality, allowing the model to learn more efficiently and improve its overall performance, thereby enhancing the model's training effect and generalization ability. Consequently, the trained information processing model can more accurately judge the superiority / inferiority of answers, thus improving the accuracy of the trained information processing model in predicting the superiority / inferiority of answers.

[0105] In some implementations, losses can be calculated separately for merit-disadvantage information and argumentation information, and then combined to determine the predicted loss. Specifically, the predicted loss includes merit-disadvantage information loss and argumentation information loss. The predicted loss is determined based on the difference between the predicted merit-disadvantage information and the merit-disadvantage information labels, and the difference between the predicted argumentation information and the argumentation information labels, including:

[0106] The loss of good and bad information is determined based on the difference between the predicted good and bad information and the good and bad information labels;

[0107] The loss of argumentation information is determined based on the difference between the predicted argumentation information and the argumentation information labels;

[0108] By combining the loss of good and bad information and the loss of argumentative information, the predicted loss is obtained.

[0109] Among them, the superior-inferiority information loss refers to the loss that measures the difference between the predicted superior-inferiority information and the superior-inferiority information labels. The superior-inferiority information loss aims to measure the accuracy of the model in judging the superiority or inferiority of answers, so as to optimize the model's ability to distinguish the superiority or inferiority of information.

[0110] Argument information loss refers to the loss that measures the difference between the predicted argument information and the argument information label. Argument information loss aims to measure the accuracy of the model in reasoning and generating argument information, in order to optimize the model's ability to generate argument information.

[0111] For example, the loss of superior / inferiority information and the loss of argumentation information can be calculated by respectively combining the predicted superior / inferiority information and the predicted argumentation information with the corresponding labels. Then, the loss of superior / inferiority information and the loss of argumentation information are combined to obtain the predicted loss. The combination method can be one or more of the following, including but not limited to weighted summation and weighted average.

[0112] In some implementations, the predicted loss can be obtained by weighting and summing the loss of superior and inferior information and the loss of argumentation information.

[0113] In some implementations, the loss of superior-inferiority information includes the loss of superior-inferiority relationship. The loss of superior-inferiority relationship refers to the loss in measuring the difference between the predicted superior-inferiority relationship and the relationship label.

[0114] It should be noted that the first raw score of any sample to be processed can represent the model's confidence level towards that sample. Therefore, during training, for ranking tasks, if the method of calculating the predicted superiority relationship by subtracting the first raw score of the positive sample from the first raw score of the negative sample is used, ideally, for any two samples to be processed, the model should tend to consider the positive sample better than the negative sample. That is, the difference between the first raw score of the positive sample and the first raw score of the negative sample should be positive.

[0115] Therefore, if the difference between the first original score of the positive sample and the first original score of the negative sample predicted by the model is positive, it indicates that the model's predicted superiority / inferiority relationship is consistent with the relationship label, i.e., there is no difference. If the difference is negative, it indicates that the model's predicted superiority / inferiority relationship is inconsistent with the relationship label, i.e., there is a difference. Thus, during training, the predicted superiority / inferiority relationship can represent the difference between the predicted superiority / inferiority relationship and the superiority / inferiority label, and therefore, the superiority / inferiority relationship loss can be directly calculated from this predicted superiority / inferiority relationship.

[0116] For example, in some implementations, if the predicted superiority / inferiority relationship is obtained by calculating the difference between the first original score of the positive sample and the first original score of the negative sample, the superiority / inferiority relationship loss can be directly calculated from this predicted superiority / inferiority relationship. Specifically, determining the superiority / inferiority information loss based on the difference between the predicted superiority / inferiority information and the superiority / inferiority information labels includes:

[0117] The probability transformation of the predicted superiority / inferiority relationship is performed to obtain the predicted superiority / inferiority probabilities.

[0118] The loss in the superior-inferiority relationship is determined by predicting the probability of superiority or inferiority.

[0119] The prediction superiority probability is used to represent the probability that the model predicts a positive sample to be better than a negative sample.

[0120] For example, the superior-inferiority relationship loss can be obtained through -logσ(logit)w -logit l The result is obtained through calculation. Where, logit w and logit l Representing positive and negative samples respectively, logit w -logit l This represents the predictive merit relationship between pairwise positive and negative samples. The sigmoid function σ() can be used to convert this predictive merit relationship into predictive merit probabilities, which range from 0 to 1. Then, the logarithmic function log() is used to convert these probabilities to a logarithmic scale, transforming small probability values ​​(close to 0) into larger negative numbers, and probabilities close to 1 into values ​​close to 0. Therefore, when σ() is close to 1 (i.e., the model is confident that an event will occur), log(σ()) will be close to 0, meaning the model has very high confidence in this prediction and the loss is small. When σ() is close to 0 (i.e., the model is confident that an event will not occur), log(σ()) will become very negative, meaning the model performs poorly on this prediction and the loss is large. Since the loss value is generally non-negative, adding a negative sign in the above formula ensures that the merit relationship loss is a positive number.

[0121] In some implementations, the loss of superior and inferior information also includes classification loss. For example, this can be achieved through λ1l c +λ2l r +λ3l l The predicted loss is calculated, l c l r and l l These represent the classification loss, the superiority / inferiority information loss, and the argumentation information loss, respectively. λ1, λ2, and λ3 are the weights, which can be set to 0.5, 1, and 0.8, respectively.

[0122] In some implementations, classification loss can be determined based on the difference between the predicted category and the label. Specifically, the good / bad information label also includes a good / bad category label, and determining the good / bad information loss based on the difference between the predicted good / bad information and the good / bad information label also includes:

[0123] The classification loss is determined based on the difference between the predicted class probability and the superior / inferior class labels.

[0124] For example, classification loss can be achieved through... The calculation yields q. i p represents the predicted class probability of any sample to be processed (the i-th sample to be processed). iThis represents the superiority / inferiority category label corresponding to the sample answer in the sample to be processed. Therefore, the category prediction probability can be transformed to a logarithmic scale using the logarithmic function log(), and then the transformed logarithmic probability is multiplied by the corresponding superiority / inferiority category label value. All products are then summed to obtain the category information loss. Since the loss value is generally non-negative, a negative sign is added to the above formula to ensure that the superiority / inferiority relationship loss is a positive number.

[0125] In some implementations, for generation tasks, the difference between the true labels and model predictions at each time step can be calculated to ensure that the argument information loss accurately measures the difference between the true labels and the model's predicted probabilities, thereby improving the model's training performance. Specifically, the predicted argument information and argument information labels are text information. Based on the difference between the predicted argument information and the argument information labels, the argument information loss is determined, including:

[0126] Determine the word differences between predicted words and target words in the predicted argument information. The target word is the word corresponding to the predicted word in the argument information label.

[0127] The loss of argumentative information is determined by word differences.

[0128] Here, the predicted word refers to the next word in the predicted argument information at any given time step. The target word refers to the true next word at that time step during training, i.e., the correct next word. Word difference refers to the difference information between the predicted word and the target word.

[0129] For example, the loss of argumentative information can be Where, p t y is the probability of a word in the predicted argument information predicted at any time step (time step t) of the model, where T is the length of the predicted argument information, and y is the probability of a word in the predicted argument information. k To demonstrate the information labeling, the model in the generation task predicts the probability p of the next word based on the preceding word sequence. t Therefore, the probability p of the word predicted at time step t can be obtained using the logarithmic function log(). t Transform to a logarithmic scale, then calculate the transformed logarithmic probability and the true label y of the corresponding word in the argument information label. t Multiplying these results yields the word differences at time step t. The sum of the product results (i.e., word differences) across all time steps is then used to obtain the argument information loss.

[0130] During model training, samples can be processed multiple times, resulting in multiple predictions for each sample. For example, in some implementations, different random seeds can be used to train the information processing model, and / or dropout can be enabled during training. This results in multiple information processing models with different weight configurations, i.e., multiple different instances of the model. During training, these models (i.e., different instances) can be used to process the sample questions and answers on the training sample set, yielding information on the prediction quality of different instances. The core idea of ​​Dropout is to randomly discard some neuron connections during training, ensuring that the network structure is not completely identical each time it is trained. This reduces the model's dependence on certain neurons and enhances the model's generalization ability.

[0131] In some implementations, if multiple predicted superiority / inferiority information values ​​of a sample answer differ from their corresponding superiority / inferiority information labels during multiple processing steps of the sample, the labels can be reversed to obtain more accurate feedback information during model training, thereby improving the model's training performance. Specifically, after processing the sample question and sample answer using the information processing model to be trained to obtain predicted superiority / inferiority information and predicted argument information, the process further includes:

[0132] Determine the cumulative deviation number corresponding to the sample answer. The cumulative deviation number represents the cumulative number of times that the predicted superiority or inferiority information corresponding to the sample answer satisfies the preset deviation with the corresponding superiority or inferiority information label.

[0133] When the cumulative number of deviations corresponding to the sample answers exceeds the preset number, the good and bad information labels corresponding to the sample answers are reversed to obtain the reversed labels. The reversed labels are used to determine the prediction loss.

[0134] The preset deviation and preset quantity are set according to actual needs. The preset deviation is used to determine whether the predicted quality information corresponding to the sample answer is consistent with the corresponding quality information label, that is, it can be used to indicate whether the predicted quality information of the model in this prediction is correct.

[0135] Among them, the reversal process refers to adjusting the superiority and inferiority information labels so that the superiority and inferiority information of the sample answer represented by them is reversed.

[0136] For example, the cumulative deviation count represents the cumulative number of predicted superiority / inferiority information items corresponding to a sample answer that satisfy a preset deviation from their corresponding superiority / inferiority information labels. These multiple predicted superiority / inferiority information items are obtained by processing multiple different instances of the information processing model to be trained. These multiple different instances are information processing models to be trained with different weight configurations.

[0137] For classification tasks, if the number of discrepancies between the predicted class and the good / bad class labels of any sample being processed exceeds a preset number, such as the cumulative discrepancy exceeding 60%, then the original good / bad class labels can be reversed, for example, class 1 can be reversed to 0.

[0138] For a ranking task, if, among multiple predictions of superiority / inferiority from different instances, the number of times the predicted superiority / inferiority relationship between any two samples differs from the relationship labels exceeds a preset number (e.g., the cumulative deviation exceeds 60%), then the superiority / inferiority relationship in the original relationship labels can be reversed. For example, the superiority / inferiority relationship can be reversed from A>B to A. <B。

[0139] Understandably, if a sample frequently makes incorrect predictions in classification or ranking tasks across multiple different instances, it may indicate high noise levels in that sample. Therefore, corrective measures such as label reversal can be implemented to significantly improve the quality of the model training data. This helps reduce the misleading influence of erroneous data on model training, allowing the model to focus on learning high-quality, accurate data features, thereby improving the model's accuracy and reliability.

[0140] In some implementations, high-noise samples can be directly removed. Specifically, after processing the sample questions and sample answers through the information processing model to be trained to obtain prediction quality information and prediction argument information, the method further includes: removing sample answers from the training sample set when the deviation of the prediction bias corresponding to the sample answer exceeds a preset number, and not using the prediction quality information corresponding to the sample answer to determine the prediction loss.

[0141] In some implementations, for classification and / or ranking tasks, the characteristics of multiple prediction results obtained from processing the sample to be processed can be combined to determine the standard deviation of the sample to be processed, so as to dynamically adjust the prediction loss and enable the model to better handle samples of different difficulty levels during training. Specifically, the superior-inferiority information loss is determined based on the difference between the predicted superior-inferiority information and the superior-inferiority information labels, including:

[0142] The standard deviation of the sample to be processed is calculated using the first raw score;

[0143] Based on the standard deviation, the target parameter value corresponding to the sample to be processed is determined. The target parameter value is used to adjust the prediction quality information corresponding to the sample to be processed, so that the adjusted prediction quality information can be used to determine the prediction loss.

[0144] The target parameter value refers to the value used to adjust the predictive quality information in the loss calculation. The target parameter value can be used as an adjustment factor in the prediction loss to adjust the output of the original model.

[0145] For example, the standard deviation of the sample to be processed is the standard deviation of multiple prediction quality information corresponding to the sample, which is obtained by multiple different instances of the information processing model to be trained. These multiple different instances are information processing models to be trained with different weight configurations.

[0146] For a classification task, during the processing of samples by different instances, the logits values ​​of the multiple first original scores for any sample to be processed by different instances are [3.2, 2.8, 3.5, 2.9, 3.1]. The standard deviation of these logits values ​​can be calculated, denoted as ε. The larger the standard deviation, the greater the learning difficulty of the sample to be processed in the classification task. In this case, the target parameter value m = k - ε for the sample to be processed can be calculated based on preset hyperparameters, where k is a hyperparameter, which can be set to 1. This target parameter value is then used in the calculation of the classification loss for each instance. It is understood that in this embodiment, by calculating the target parameter value for each sample to be processed, the target parameter value corresponding to each sample can be added to the corresponding calculation process for that sample when calculating the loss.

[0147] For example, in a sorting task, during the processing of samples by different instances, the logits values ​​of the multiple first original scores for any given sample are [3.2, 2.8, 3.5, 2.9, 3.1]. The difference between the logits values ​​(logits) calculated from the first original scores of the positive sample and the negative sample in any two samples can be obtained. i -logits j (i.e., predicting the superiority / inferiority relationship). The standard deviation of the logits difference is calculated and denoted as ε. A larger standard deviation indicates a greater learning difficulty in determining the superiority / inferiority relationship between the two samples to be processed in the ranking task. In this case, the target parameter value m = k - ε can be calculated based on preset hyperparameters, where k is a hyperparameter. This target parameter value is then used in the calculation of the superiority / inferiority relationship loss for each instance. It is understood that in this embodiment, by calculating the target parameter value for each pair of samples to be processed, the target parameter values ​​corresponding to each pair of samples to be processed can be added to the corresponding calculation process for each pair of samples to be processed when calculating the loss.

[0148] Therefore, for classification and / or ranking tasks, the standard deviation of a sample can be dynamically determined by combining the characteristics of multiple prediction results obtained from processing each sample. A larger standard deviation indicates a potentially higher learning difficulty for that sample. Thus, a dynamic target parameter value can be introduced using this standard deviation to adjust the prediction loss, allocate more training resources and attention, and enable the model to better handle samples of varying difficulty levels during training. Furthermore, a larger sample standard deviation indicates higher noise levels in the sample. Therefore, by using correction measures such as adjusting the target parameter value, the quality of the model training data can be significantly improved, reducing the misleading influence of erroneous data on model training. This allows the model to focus on learning high-quality, accurate data features, thereby improving the model's accuracy and reliability.

[0149] In some implementations, the target parameter value is used to adjust the prediction quality information corresponding to the sample to be processed, so that the adjusted prediction quality information is used to determine the prediction loss. For example, for any sample to be processed, the target parameter value of the sample is used to adjust its corresponding prediction quality information, such as prediction quality relationship and / or category prediction probability.

[0150] In some implementations, the target parameter value can be used to adjust the predicted superiority / inferiority relationship. For example, for the superiority / inferiority relationship loss -logσ(logit) w -logit l This allows adjusting the difference between the first original score of the corresponding positive sample and the first original score of the negative sample using the target parameter value. For example, the difference can be adjusted by logging the target parameter value. w -logit l Subtract the target parameter value m. Therefore, the superior-inferiority relationship loss can be -logσ(logit) w -logit l -m). Where, logit w and logit l Representing positive and negative samples respectively, logit w -logit l This represents the predictive merit relationship between pairs of positive and negative samples, and σ() represents the sigmoid function.

[0151] In some implementations, the target parameter value can be used to adjust the class prediction probability. For example, for classification loss. The first raw score of the corresponding sample to be processed can be adjusted using the target parameter value, such as adjusting the first raw score q. i Subtract the target parameter value m, and calculate the adjusted class prediction probability from the difference. Thus, the classification loss can be... Where, q i p represents the predicted class probability of any sample to be processed (the i-th sample to be processed).i This indicates the superiority or inferiority category label corresponding to the sample answer in the sample to be processed.

[0152] 140. Use the prediction loss to train the information processing model to be trained, and obtain the trained information processing model.

[0153] For example, in this embodiment, the information processing model to be trained can be trained for multiple training cycles based on the prediction loss. A training cycle refers to the number of times the training sample set is input into the information processing model to be trained and the model parameters are updated once. In each training cycle, the prediction loss of the current training cycle can be calculated based on the prediction information and the corresponding labels, and backpropagation and model parameter update processes can be performed based on this loss until the preset training termination condition is reached (such as reaching a preset number or model convergence). The information processing model updated in the last training cycle is the post-trained information processing model.

[0154] In some implementations, test data can be used to test the trained information processing model to evaluate its performance and ensure that the final trained model has good performance and generalization ability. Specifically, after training the information processing model to be trained using prediction loss to obtain the trained information processing model, the process further includes:

[0155] Obtain a test sample set, which includes test questions, multiple test answers to the test questions, and test superiority / inferiority information labels. The test superiority / inferiority information labels include test relationship labels used to indicate the superiority / inferiority relationship between test answers.

[0156] By using the trained information processing model, the test questions and test answers are processed to obtain test prediction quality information, which includes the test prediction quality relationship corresponding to the test relationship label.

[0157] Based on the difference between the predicted superior and inferior information used in the test and the labels of the superior and inferior information used in the test, the evaluation index score of the post-trained information processing model is determined. The evaluation index score is used to evaluate the performance of the post-trained information processing model.

[0158] The test sample set refers to the combination of samples and their labels used to test the trained information processing model. The test sample set may include at least one question (i.e., the test question) and multiple answers to each sample question (i.e., the test answers). In this embodiment, the test question, test answer, and test quality information labels are the corresponding samples and their labels used to test the trained information processing model.

[0159] The evaluation index score refers to the score of the evaluation index. Evaluation indices may include, but are not limited to, one or more of the following: accuracy, recall, F1 score, and area under the ROC curve (AUC value).

[0160] For example, during testing, processing the test questions and answers using a trained information processing model constitutes a prediction process. This prediction process can be a process of predicting the quality of test answers based on the test questions and answers. This prediction process could be a process of predicting the order of samples (i.e., test answers) based on a ranking task, and / or a process of predicting the classification of samples based on a classification task. For example, for classification and ranking tasks, the quality relationship of the test predictions and / or the probability of the test category predictions can be determined based on the original scores (logits) output by the model for each answer to be processed, thus obtaining the quality relationship of the test predictions and / or the probability of the test category predictions. Based on one or more evaluation index calculation methods, scores for corresponding indicators can be calculated for the quality information of the test predictions and the labels of the quality information, to obtain one or more evaluation index scores for the trained information processing model. It should be noted that the specific implementation and principle of the classification prediction and order prediction processes of the post-trained information processing model in this application embodiment are the same as those of the classification prediction and order prediction processes in the model training method. For details, please refer to the implementation methods in the foregoing embodiments. This application embodiment will not elaborate further.

[0161] For example, the difference between the test prediction information and the test information labels can be used to determine correctly predicted samples and incorrectly predicted samples. Correctly predicted samples are those where the test prediction information and the test information labels are the same, while incorrectly predicted samples are those where the test prediction information and the test information labels are different. Scores for various evaluation metrics can be calculated based on the correctly predicted and incorrectly predicted samples. For example, the percentage of correctly predicted samples out of the total sample size can be calculated to obtain precision; the proportion of correctly predicted positive samples can be calculated to obtain precision and recall; or the F1 score (the harmonic mean of precision and recall) can be calculated from precision and recall. It should be noted that the greater the difference between the test prediction information and the test information labels, the better the evaluation metric scores will generally be. For example, higher precision, recall, and F1 scores, and an AUC value closer to 1, indicate better model performance.

[0162] It should be noted that in this embodiment, only the test sample set is used to test the trained information processing model for ranking and / or classification tasks, without testing the generation task. This simplifies the evaluation process, improves efficiency, and helps to quickly determine and improve model performance. Furthermore, since the generation task is based on the relative merits of the answers, testing only the classification and / or ranking tasks ensures that the model has high processing capabilities in the generation task.

[0163] In some implementations, after determining the evaluation index score of the post-trained information processing model based on the difference between the test prediction quality information and the test quality information label, the method further includes: if the evaluation index score does not meet the preset score, retraining the post-trained information processing model.

[0164] The preset score refers to the score conditions set according to actual needs. The preset scores may differ for different types of evaluation indicators, and can be set according to the application scenario or actual needs. This application embodiment does not impose any restrictions.

[0165] Understandably, when the evaluation metric score does not meet the preset score, it means the performance of the trained information processing model has not met expectations. Therefore, the model can be retrained until the evaluation metric score meets the preset score. When the evaluation metric score meets the preset score, it means the performance of the trained information processing model has met expectations. Therefore, the model can be discontinued and used in real-world applications. This ensures that the final trained model has good performance and generalization ability.

[0166] The model training scheme provided in this application can be applied to various model training scenarios. For example, it involves obtaining an information processing model to be trained and a training sample set. The training sample set includes sample questions, multiple sample answers to the sample questions, superiority / inferiority information labels, and argumentation information labels. The superiority / inferiority information labels include relation labels representing the superiority / inferiority relationships between sample answers, and the argumentation information labels are information labels representing the superiority / inferiority relationships of arguments. The information processing model to be trained processes the sample questions and sample answers to obtain predicted superiority / inferiority information and predicted argumentation information. The predicted superiority / inferiority information includes predicted superiority / inferiority relationships corresponding to the relation labels. Based on the differences between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the differences between the predicted argumentation information and the argumentation information labels, a prediction loss is determined. The prediction loss is then used to train the information processing model to be trained, resulting in a trained information processing model.

[0167] As can be seen from the above, in this embodiment, the prediction loss of the information processing model to be trained is constructed by combining the differences between the predicted superiority / inferiority information and the labels, as well as the differences between the predicted argumentation information and the labels. This ensures that the training process is not limited to focusing solely on the superiority / inferiority information of the final answer, such as the superiority / inferiority relationship, but also considers the argumentation logic of the answer's superiority / inferiority information based on the argumentation information. Specifically, this argumentation information provides the model with additional feature information, enabling the model to more accurately capture the key elements that need to be focused on when comparing the superiority / inferiority of answers. This optimizes the model's understanding and processing of answer quality, allowing the model to learn more efficiently and improve its overall performance, thereby enhancing the model's training effect and generalization ability. Consequently, the trained information processing model can more accurately judge the superiority / inferiority of answers, thus improving the accuracy of the trained information processing model in predicting the superiority / inferiority of answers.

[0168] In this embodiment, an information processing method is provided, such as... Figure 2a As shown, the specific process of this information processing method can be as follows:

[0169] 210. Obtain the trained information processing model, the problem to be processed, and multiple answers to the problem to be processed.

[0170] Among them, the questions to be processed and the answers to be processed refer to the questions and answers that the information processing model to be trained will process during the reasoning (inference) phase after training.

[0171] For example, in this embodiment, the trained information processing model is obtained using any of the model training methods provided in this embodiment. The problem to be processed and the answer to be processed can come from user input or be automatically generated by the model.

[0172] 220. By training the information processing model, the problem to be processed and the answer to be processed are processed to obtain the target superiority information, which includes the target superiority relationship used to represent the superiority relationship between the answers to be processed.

[0173] Among them, target quality information refers to the quality information of the answer to be processed, inferred by the information processing model after training, based on the question to be processed and the answer to be processed. Target quality relationship refers to the quality relationship between the answers to be processed, inferred by the information processing model after training, based on the question to be processed and the answer to be processed.

[0174] For example, during the inference phase, a trained information processing model can be used to predict the problem and the answer to be processed. In this prediction process, each answer and the problem can be considered as a single piece of information to be processed. Figure 2bThe information processing process shown describes how the trained information processing model can perform classification prediction based on a classification task, or order prediction based on a ranking task. This allows for the determination of the relative merits and / or predicted categories of the answers to be processed.

[0175] The information processing method of this application embodiment can be applied to various question-and-answer based information processing scenarios. For example, in a question-and-answer scenario, a user can input a question and multiple answers to the question into a question-and-answer system, so that the question-and-answer system can call a trained information processing model to process the input question and answers, obtain the relationship between the quality of the answers and the predicted category. This question-and-answer system can be a system in the fields of intelligent customer service, question-and-answer search engine, knowledge base and FAQ management system, and sentiment analysis.

[0176] Alternatively, users can input a question into the question-answering system, which will automatically generate multiple answers. Simultaneously, the system will call a trained post-processing model to process the input question and the generated answers, determining the relative merits of the answers and their predicted categories. For example, the predicted category for any answer can be either positive (e.g., represented as 1) or negative (e.g., represented as 0).

[0177] Therefore, in specific problem scenarios, question-answering systems can select one or more answers from multiple answers to a question input by the user or automatically generated by the system, based on the superiority / inferiority relationship and predicted category obtained from the model processing, to provide feedback to the user, thereby improving the accuracy of the feedback information from the question-answering system. For example, when the relationship label is a preference label, the superiority / inferiority relationship can represent the user's relative preference for multiple answers to a question, while the predicted category can represent whether each label is objectively correct (i.e., superior or inferior). Therefore, the question-answering system can select one or more answers that are most preferred by the user and are correct (i.e., predicted as superior) from multiple answers to a question based on the preference label and the predicted category, and provide feedback to the user. For example, if the superiority / inferiority relationship of four answers to any question is y1>y2>y3>y4, where the predicted category of answers y1 and y2 is superior, and the predicted category of answers y3 and y4 is inferior, the question-answering system can provide feedback to the user with answer y1 as the optimal answer, or provide feedback to the user with y1 and y2 as the optimal answers.

[0178] In some implementations, when predicting the superiority or inferiority relationship as the superiority or inferiority relationship between pairwise sample answers among multiple sample answers, the target superiority or inferiority relationship is the superiority or inferiority relationship between pairwise unprocessed answers among multiple unprocessed answers. For example, the superiority or inferiority relationship between any two answers can be "Answer 1 is more popular than Answer 2" or "Answer 2 is more popular than Answer 1".

[0179] In some implementations, the post-trained information processing model can also perform a regression prediction process based on the generated task to generate target argument information, which is information about the superiority or inferiority relationship of the target.

[0180] In some implementations, the target superiority / inferiority relationship, and / or target predicted category, and / or target argument information are in a user-readable information form, such as text. For example, the trained information processing model can perform prediction processing on the question to be processed and the answer to be processed. This prediction processing can include a classification prediction process based on a classification task, a ranking order prediction process based on a ranking task, and a regression prediction process based on a generation task, to obtain the probability of the superiority / inferiority relationship, the probability of the predicted category, and the probability of the argument information, respectively. These three probabilities can be converted into the corresponding target superiority / inferiority relationship, target predicted category, and target argument information. For classification and ranking tasks, the target superiority / inferiority relationship and the target predicted category can be determined based on the original scores (logits) output by the model for each answer to be processed. For generation tasks, the target argument information can be determined based on the original scores (logits) output by the model for each pair of answers to be processed. It should be noted that the specific implementation and principle of the classification prediction, ranking order prediction, and regression prediction processes of the trained information processing model in this application embodiment are the same as the classification prediction, ranking order prediction, and regression prediction processes in the model training method. For details, please refer to the implementation methods in the foregoing embodiments, which will not be repeated in this application embodiment.

[0181] The information processing scheme provided in this application can be applied to various information processing scenarios. For example, it can obtain a trained information processing model, a problem to be processed, and multiple answers to the problem to be processed; through the trained information processing model, the problem to be processed and the answers to be processed are processed to obtain target superiority information, which includes target superiority relationships used to represent the superiority relationships between the answers to be processed.

[0182] As can be seen from the above, in the embodiments of this application, during the training process of the information processing model, the prediction loss of the information processing model to be trained is constructed by combining the differences between the predicted superiority / inferiority information and the labels, as well as the differences between the predicted argumentation information and the labels. This ensures that the training process is not limited to focusing solely on the superiority / inferiority information of the final answer, such as the superiority / inferiority relationship, but also considers the argumentation logic of the superiority / inferiority information of the answer based on the argumentation information. Specifically, this argumentation information provides the model with additional feature information assistance, enabling the model to more accurately capture the key elements that need to be focused on when comparing the superiority / inferiority of answers, thereby optimizing the model's understanding and processing of answer quality. This allows the model to learn more efficiently and improve its overall performance, thus enhancing the training effect and generalization ability of the model. Consequently, the trained information processing model can more accurately judge the superiority / inferiority of answers, thereby improving the accuracy of the trained information processing model in predicting the superiority / inferiority information of answers.

[0183] The method described in the above embodiments will be further described in detail below.

[0184] In this embodiment, the method of this application will be described in detail.

[0185] like Figure 3a As shown, the specific process of a model training method is as follows:

[0186] 310. Obtain the information processing model to be trained and the training sample set. The training sample set includes sample questions, multiple sample answers to the sample questions, superiority and inferiority information labels, and argumentation information labels. Superiority and inferiority information labels include relational labels used to represent the superiority and inferiority relationships between sample answers, and argumentation information labels are information labels representing the superiority and inferiority relationships of arguments.

[0187] For example, one can obtain the information processing model to be trained and the training sample set, and train the information processing model so that the trained information processing model can provide responses with good or bad information for question-and-answer scenarios.

[0188] For the training sample set, we can first pull online user queries (i.e., sample questions), and then use different versions of the dialogue model and manually generated corresponding responses (i.e., sample answers) to generate the raw data. Assume the user query is x, and the corresponding generated responses are y1, y2, ..., y n .

[0189] The collected raw data is then manually labeled with classification and preference data. For example, the responses generated in the previous step can be categorized (i.e., good / bad category labels) and labeled with preferences (i.e., relationship labels). For example, preference labeling: y1>y2>y3, classification labeling: y1=1, y2=1, y3=0. Here, 0 represents bad, 1 represents good, and a>b indicates a stronger preference for a. Categorization can coexist within preferences. For example, if both sample answer A and sample answer B are high-quality samples, and the relationship label A>B, then preferences can also be constructed through classification. If sample answer A is known to be a high-quality sample and sample answer B is a low-quality sample, then the relationship label A>B can be constructed.

[0190] After manually labeling preferences in the previous step, these preferences are directly input into a more powerful LLM to generate reasons for the preferences (i.e., argument information labels). For example, if A > B, the more powerful LLM generates reasons for the preferences: A is better than B because A is more fluent, etc.

[0191] For example, labeled preference data, such as "A>B," can be input into a more pre-trained LLM (i.e., a pre-trained language model) to generate reasons for the preferences. For instance, if the input is "Article A is more popular than Article B," a stronger LLM might output, "Article A is more popular because its structure is clearer, its logic is more coherent, and its language is more vivid and engaging, better attracting readers' attention. Furthermore, Article A presents more novel and unique viewpoints, offering greater inspiration." This approach provides the model with more feature information to assist in classification and preference training, enabling it to more accurately capture the key elements to focus on when comparing preferences. This improves the model's training effectiveness and generalization ability, further optimizing its understanding and processing of different data. Consequently, in subsequent training using a combination of three types of data and dynamic thresholds, the model can learn more efficiently and improve its overall performance.

[0192] 320. Using the information processing model to be trained, the sample questions and sample answers are processed to obtain the predicted superiority and inferiority information and the predicted argumentation information. The predicted superiority and inferiority information includes the predicted superiority and inferiority relationships corresponding to the relation labels.

[0193] For example, the process by which the information processing model to be trained processes sample questions and sample answers is the process of prediction. This prediction process can be the process of predicting the quality and argumentation information of sample answers based on the sample questions and sample answers. For instance, the information processing model to be trained can perform classification prediction based on classification tasks, ranking prediction based on ranking tasks, and regression prediction based on generation tasks, etc. For classification and ranking tasks, the quality and probability of prediction can be determined based on the first raw score (class logits) output by the model for each sample. For generation tasks, the argumentation information can be determined based on the second raw score (generated logits) output by the model for each pair of samples.

[0194] 330. Determine the cumulative deviation number corresponding to the sample answer. The cumulative deviation number represents the cumulative number of predicted superiority / inferiority information and corresponding superiority / inferiority information labels that satisfy the preset deviation.

[0195] 340. When the cumulative number of deviations corresponding to the sample answers exceeds the preset number, the good and bad information labels corresponding to the sample answers are reversed to obtain the reversed labels. The reversed labels are used to determine the prediction loss.

[0196] For example, different random seeds can be used to train the information processing model to be trained, and / or dropout can be enabled during training to obtain multiple information processing models to be trained with different weight configurations, i.e., multiple different instances of the information processing model to be trained. During training, for the training sample set, these models (i.e., different instances) are used to process the sample questions and sample answers respectively to obtain the prediction quality information of different instances, and the threshold and noise are calculated based on the prediction values.

[0197] Label inversion refers to reversing the relative importance of sample answers by adjusting the labels. For classification and ranking tasks, the consistency between model predictions and labels can be determined by comparing the predicted categories (the categories corresponding to the predicted category probabilities) and predicted preference order (the predicted relative importance) of different models (i.e., different instances) with the labels. If a sample frequently makes errors in classification or preference order predictions (i.e., relative importance) across multiple inferences, the sample may contain high noise. Label inversion can address this noise by reversing the original labels; for example, a classification of 1 is reversed to 0, and a relative importance of A>B is reversed to A. <B。

[0198] 350. Determine the target parameter value corresponding to the sample to be processed. The target parameter value is used to adjust the prediction quality information corresponding to the sample to be processed, so that the adjusted prediction quality information can be used to determine the prediction loss.

[0199] For example, a dynamic threshold, also known as a decision threshold (i.e., the target parameter value), can then be obtained based on classification and preference data (superiority / disadvantage relationship). This can be achieved by statistically analyzing the stability of different models in predicting sample classification or preferences, i.e., by observing the values ​​of the model outputs and collecting the logits values ​​of different models for each sample. Suppose that for a given sample (regardless of classification or ranking), the logits values ​​given by different models are [3.2, 2.8, 3.5, 2.9, 3.1]. Calculate its standard deviation, denoted as ε. The larger the standard deviation, the greater the learning difficulty of the sample. In this case, the dynamic threshold calculated for the sample is m = k - ε, where k is a hyperparameter. For a classification task, q1 = σ(logit w ),q2=σ(logit l The original classification loss was calculated as follows: After adding a dynamic threshold m, the loss is calculated as follows: Similarly, for ranking tasks, the original loss based on superiority / inferiority relationships is calculated as -logσ(logit). w -logit l After adding a dynamic threshold, the loss is changed to -logσ(logit) w -logit l -m). Where, logit w and logit l Representing positive and negative samples respectively, logit w -logit l This represents the predictive merit relationship between pairwise positive and negative samples, where σ() represents the sigmoid function, and q i p represents the predicted class probability of any sample to be processed (the i-th sample to be processed). i This indicates the superiority or inferiority category label corresponding to the sample answer in the sample to be processed.

[0200] 360. Based on the differences between the predicted superior and inferior information and the superior and inferior information labels, as well as the differences between the predicted argument information and the argument information labels, determine the prediction loss.

[0201] 370. Use the prediction loss to train the information processing model to be trained, and obtain the trained information processing model.

[0202] For example, the model can be trained simultaneously using the losses and dynamic thresholds from classification, ranking, and generation tasks. First, noise is removed using inverted labels (either by direct removal or inversion). Then, during training, the calculation methods for the losses of multiple tasks are reorganized, and the losses from the three types of data are weighted to obtain the predicted loss. Finally, a dynamic value (i.e., the target parameter value) is introduced into the loss calculation.

[0203] Specifically, the losses calculated from the three types of data can be weighted, i.e., λ1l c +λ2l r +λ3l l , where l c Represents classification loss, l r This represents the loss of preference (i.e., the loss of superiority / inferiority relationship), l l Let λm represent the language modeling loss (i.e., the loss of argument information), and λ1, λ2, and λ3 be the corresponding weight coefficients. During training, the loss for categorical data and preference data (i.e., superiority / inferiority relationships) is adjusted using the previously calculated dynamic threshold m. For the language modeling part (i.e., the generation task), the next label prediction method is followed. Specifically, the model receives a text sequence (i.e., argument information labels) such as "A is better than B because [A's narration is more fluent and more eye-catching]", and then calculates the loss based on the content contained in "[]". This approach allows for better utilization of labeled data to automatically identify noise in the data and the learning difficulty of samples, improving the model's generalization ability and adaptability.

[0204] After obtaining the trained information processing model, the trained information processing model can be used for information processing. The specific information processing process includes the following steps:

[0205] The system acquires a post-trained information processing model, a problem to be processed, and multiple answers to the problem to be processed. The post-trained information processing model is used to process the problem to be processed and the answers to be processed to obtain target superiority information, which includes target superiority relationships used to represent the superiority relationships between the answers to be processed.

[0206] like Figure 3b The table showing the comparison of the trained model performance illustrates that, for the same information processing model to be trained, compared to Model 1 (trained only for classification tasks), Model 2 (trained only for ranking tasks), and Model 3 (trained only for classification and ranking tasks), the information processing model trained in this embodiment simultaneously on classification, ranking, and generation tasks has the highest accuracy in predicting superiority / inferiority information and the highest F1 score. Therefore, in this embodiment, by jointly training the information processing model using classification, ranking, and generation tasks, and by introducing the logic of reasoning about the superiority / inferiority of answers during training, the trained information processing model can have better processing capabilities in classification and ranking tasks. This allows the trained information processing model to more accurately judge the quality of answers, thereby improving the accuracy of its predictions regarding the superiority / inferiority of answers.

[0207] As can be seen from the above, in this embodiment, by simultaneously training language modeling, classification, and preference tasks, various types of information in the labeled data can be fully integrated, no longer limited to the utilization of a single type of data. Whether it's simple classification labeling or superior-inferiority relationship labeling, both can function within the same training framework, avoiding the previous problem of some training data being idle or unable to be effectively integrated, thereby improving the overall utilization rate of the data and providing a richer and more comprehensive information foundation for model training. Furthermore, by utilizing dynamic thresholding technology combined with the randomness introduced by repeated training and dropout, noise in the data can be effectively identified, including incorrect preference order and incorrect classification categories. For high-noise samples, not only can they be accurately detected, but corrective measures such as label reversal can also be taken, significantly improving the quality of the model training data. This helps reduce the misleading effect of erroneous data on model training, allowing the model to focus on learning high-quality, accurate data features, thereby improving the model's accuracy and reliability. Additionally, the calculation of dynamic thresholds can also reflect the difficulty level of sample learning. For samples with higher learning difficulty (larger standard deviation), the loss calculation method is adjusted by adjusting the threshold, allocating more training resources and attention, enabling the model to better cope with samples of different difficulty levels during training.

[0208] To better implement the above methods, this application also provides a model training device or information processing device, which can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers.

[0209] For example, in this embodiment, the method of this application embodiment will be described in detail by taking the model training device specifically integrated into the server as an example.

[0210] For example, such as Figure 4 As shown, the model training device may include a training acquisition unit 410, a sample processing unit 420, a loss determination unit 430, and a training unit 440, as follows:

[0211] (I) Training Acquisition Unit 410

[0212] This is used to obtain the information processing model to be trained and the training sample set. The training sample set includes sample questions, multiple sample answers to the sample questions, superiority and inferiority information labels, and argumentation information labels. The superiority and inferiority information labels include relational labels used to represent the superiority and inferiority relationships between sample answers, and the argumentation information labels are information labels representing the superiority and inferiority relationships of the arguments.

[0213] (II) Sample Processing Unit 420

[0214] It is used to process sample questions and sample answers through the information processing model to be trained, and obtain prediction quality information and prediction argument information. The prediction quality information includes the prediction quality relationship corresponding to the relationship label.

[0215] In some embodiments, the sample processing unit includes a first feature processing subunit and a second feature processing subunit, comprising: a first feature processing subunit, configured to perform first feature processing on the sample to be processed using a training information processing model to obtain a first original score for the sample to be processed, wherein the sample to be processed includes a sample question and a sample answer, and each sample to be processed corresponds to a sample answer; and to obtain the predictive superiority relationship between the sample answers in the sample to be processed by comparing the first original scores; and a second feature processing subunit, configured to perform second feature processing on the sample to be processed using a training information processing model to obtain a second original score for the sample to be processed; and to determine the predictive argument information based on the second original score.

[0216] In some implementations, the predictive superiority relationship between sample answers in the samples to be processed is obtained by comparing the first original scores, including: for any two samples to be processed, based on the relationship labels, determining positive and negative samples from any two samples to be processed; calculating the difference between the first original score of the positive sample and the first original score of the negative sample to obtain the predictive superiority relationship between sample answers in any two samples to be processed.

[0217] In some implementations, the superiority / inferiority information label also includes a superiority / inferiority category label, and the predicted superiority / inferiority information also includes the category prediction probability corresponding to the superiority / inferiority category label. The first feature processing subunit is further configured to: perform probability transformation processing on the first original score to obtain the category prediction probability.

[0218] (III) Loss Determination Unit 430

[0219] It is used to determine the prediction loss based on the difference between the predicted superior and inferior information and the superior and inferior information label, as well as the difference between the predicted argument information and the argument information label.

[0220] In some implementations, the predicted loss includes superior / inferiority information loss and argumentation information loss. The loss determination unit includes a first loss determination subunit, a second loss determination subunit, and a third loss determination subunit, comprising: a first loss determination subunit for determining superior / inferiority information loss based on the difference between predicted superior / inferiority information and superior / inferiority information labels; a second loss determination subunit for determining argumentation information loss based on the difference between predicted argumentation information and argumentation information labels; and a third loss determination subunit for combining superior / inferiority information loss and argumentation information loss to obtain the predicted loss.

[0221] In some implementations, the predicted argument information and the argument information label are text information, and the second loss determination subunit is specifically used to: determine the word difference between the predicted word and the target word in the predicted argument information, where the target word is the word corresponding to the predicted word in the argument information label; and determine the argument information loss based on the word difference.

[0222] In some implementations, the model training apparatus further includes a threshold determination unit, which is used to: calculate the standard deviation of the sample to be processed using the first original score; and determine the target parameter value corresponding to the sample to be processed based on the standard deviation. The target parameter value is used to adjust the prediction quality information corresponding to the sample to be processed, so as to use the adjusted prediction quality information to determine the prediction loss.

[0223] In some implementations, the model training apparatus further includes a reversal unit, which is used to: determine the cumulative number of deviations corresponding to the sample answer, where the cumulative number of deviations represents the cumulative number of predicted superiority / inferiority information and corresponding superiority / inferiority information labels corresponding to the sample answer satisfying a preset deviation; when the cumulative number of deviations corresponding to the sample answer exceeds the preset number, the superiority / inferiority information labels corresponding to the sample answer are reversed to obtain reversed labels, and the reversed labels are used to determine the prediction loss.

[0224] (iv) Training Unit 440

[0225] It is used to train the information processing model to be trained using predictive loss, and obtain the trained information processing model.

[0226] In some embodiments, the model training apparatus further includes a testing unit, which is used to: acquire a test sample set, the test sample set including test questions, multiple test answers to the test questions, and test superiority / inferiority information labels, the test superiority / inferiority information labels including test relationship labels used to represent the superiority / inferiority relationship between test answers; process the test questions and test answers through the trained information processing model to obtain test predicted superiority / inferiority information, the test predicted superiority / inferiority information including the test predicted superiority / inferiority relationship corresponding to the test relationship labels; and determine the evaluation index score of the trained information processing model based on the difference between the test predicted superiority / inferiority information and the test superiority / inferiority information labels, the evaluation index score being used to evaluate the performance of the trained information processing model.

[0227] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0228] As can be seen from the above, the model training device in this embodiment includes a training acquisition unit, a sample processing unit, a loss determination unit, and a training unit. The training acquisition unit acquires the information processing model to be trained and a training sample set. The training sample set includes sample questions, multiple sample answers to the sample questions, superiority / inferiority information labels, and argumentation information labels. The superiority / inferiority information labels include relational labels representing the superiority / inferiority relationships between sample answers, and the argumentation information labels are information labels representing the superiority / inferiority relationships of arguments. The sample processing unit processes the sample questions and sample answers using the information processing model to be trained to obtain predicted superiority / inferiority information and predicted argumentation information. The predicted superiority / inferiority information includes predicted superiority / inferiority relationships corresponding to the relational labels. The loss determination unit determines the prediction loss based on the differences between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the differences between the predicted argumentation information and the argumentation information labels. The training unit trains the information processing model to be trained using the prediction loss to obtain the trained information processing model.

[0229] Therefore, in this embodiment, the predictive loss of the information processing model to be trained is constructed by combining the differences between the predicted superiority / inferiority information and the labels, as well as the differences between the predicted argumentation information and the labels. This ensures that the training process is not limited to focusing solely on the superiority / inferiority information of the final answer, such as the superiority / inferiority relationship, but also considers the argumentation logic of the answer's superiority / inferiority information based on the argumentation information. Specifically, this argumentation information provides the model with additional feature information to assist it, enabling the model to more accurately capture the key elements that need to be focused on when comparing the superiority / inferiority of answers. This optimizes the model's understanding and processing of answer quality, allowing the model to learn more efficiently and improve its overall performance, thereby enhancing the model's training effect and generalization ability. Consequently, the trained information processing model can more accurately judge the superiority / inferiority of answers, thus improving the accuracy of the trained information processing model in predicting the superiority / inferiority of answers.

[0230] To better implement the above methods, this application also provides an information processing device, which can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers.

[0231] For example, in this embodiment, the method of this application embodiment will be described in detail by taking the information processing device specifically integrated into the server as an example.

[0232] For example, such as Figure 5 As shown, the information processing device may include a processing acquisition unit 510 and an information processing unit 520, as follows:

[0233] (a) Processing and Acquisition Unit 510

[0234] This is used to obtain a post-trained information processing model, a problem to be processed, and multiple unprocessed answers to the problem. The post-trained information processing model is obtained through any of the model training methods provided in the embodiments of this application.

[0235] (II) Information Processing Unit 520

[0236] This is used to process the problem and the answer to be processed through the trained information processing model, and to obtain the target superiority information, which includes the target superiority relationship used to represent the superiority relationship between the answers to be processed.

[0237] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0238] As can be seen from the above, the information processing device of this embodiment includes a processing acquisition unit and an information processing unit. The processing acquisition unit is used to acquire a trained information processing model, a problem to be processed, and multiple answers to the problem to be processed. The trained information processing model is obtained through any of the model training methods provided in this application embodiment. The information processing unit is used to process the problem to be processed and the answers to be processed through the trained information processing model to obtain target superiority information, which includes target superiority relationships used to represent the superiority relationships between the answers to be processed.

[0239] Therefore, in this embodiment, during the training of the information processing model, the prediction loss of the information processing model to be trained is constructed by combining the differences between the predicted superiority / inferiority information and the labels, as well as the differences between the predicted argumentation information and the labels. This ensures that the training process is not limited to focusing solely on the superiority / inferiority information of the final answer, such as the superiority / inferiority relationship, but also considers the argumentation logic of the superiority / inferiority information of the answer based on the argumentation information. Specifically, this argumentation information provides the model with additional feature information to assist it, enabling the model to more accurately capture the key elements that need to be focused on when comparing the superiority / inferiority of answers. This optimizes the model's understanding and processing of answer quality, allowing the model to learn more efficiently and improve its overall performance, thereby enhancing the training effect and generalization ability of the model. Consequently, the trained information processing model can more accurately judge the superiority / inferiority of answers, thus improving the accuracy of the trained information processing model in predicting the superiority / inferiority information of answers.

[0240] This application also provides an electronic device, which can be a terminal, a server, or other similar device. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers, etc.

[0241] In some embodiments, the model training device or information processing device may also be integrated into multiple electronic devices. For example, the model training device or information processing device may be integrated into multiple servers, and the model training method or information processing method of this application may be implemented by multiple servers.

[0242] In this embodiment, a server will be used as an example for detailed description. For example, ... Figure 6 As shown, it illustrates a schematic diagram of the server structure involved in an embodiment of this application. Specifically:

[0243] The server may include components such as a processor 610 with one or more processing cores, a memory 620 with one or more computer-readable storage media, a power supply 630, an input module 640, and a communication module 650. Those skilled in the art will understand that... Figure 6 The server architecture shown does not constitute a limitation on the server and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Wherein:

[0244] The processor 610 is the control center of the server, connecting various parts of the server via various interfaces and lines. It performs various server functions and processes data by running or executing software programs and / or modules stored in the memory 620, and by accessing data stored in the memory 620. In some embodiments, the processor 610 may include one or more processing cores; in some embodiments, the processor 610 may integrate an application processor and a modem processor, wherein the application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 610.

[0245] The memory 620 can be used to store software programs and modules. The processor 610 executes various functional applications and data processing by running the software programs and modules stored in the memory 620. The memory 620 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the server, etc. In addition, the memory 620 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 620 may also include a memory controller to provide the processor 610 with access to the memory 620.

[0246] The server also includes a power supply 630 that supplies power to the various components. In some embodiments, the power supply 630 can be logically connected to the processor 610 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 630 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0247] The server may also include an input module 640, which can be used to receive input numeric or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0248] The server may also include a communication module 650. In some embodiments, the communication module 650 may include a wireless module, through which the server can perform short-range wireless transmission, thereby providing users with wireless broadband internet access. For example, the communication module 650 can be used to help users send and receive emails, browse web pages, and access streaming media.

[0249] Although not shown, the server may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 610 in the server loads the executable files corresponding to the processes of one or more applications into the memory 620 according to the following instructions, and the processor 610 runs the applications stored in the memory 620 to realize various functions, as follows:

[0250] Obtain the information processing model to be trained and the training sample set. The training sample set includes sample questions, multiple sample answers to the sample questions, superiority / inferiority information labels, and argumentation information labels. The superiority / inferiority information labels include relation labels representing the superiority / inferiority relationships between sample answers, and the argumentation information labels are the information labels representing the superiority / inferiority relationships of arguments. The information processing model to be trained processes the sample questions and sample answers to obtain predicted superiority / inferiority information and predicted argumentation information. The predicted superiority / inferiority information includes the predicted superiority / inferiority relationships corresponding to the relation labels. Based on the differences between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the differences between the predicted argumentation information and the argumentation information labels, determine the prediction loss. Use the prediction loss to train the information processing model to be trained, obtaining the trained information processing model.

[0251] Alternatively, a post-trained information processing model, a problem to be processed, and multiple answers to the problem to be processed can be obtained. The post-trained information processing model is obtained through any of the model training methods provided in the embodiments of this application. The problem to be processed and the answers to be processed are processed through the post-trained information processing model to obtain target superiority information. The target superiority information includes target superiority relationships used to represent the superiority relationships between the answers to be processed.

[0252] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0253] As can be seen from the above, in this embodiment, the prediction loss of the information processing model to be trained is constructed by combining the differences between the predicted superiority / inferiority information and the labels, as well as the differences between the predicted argumentation information and the labels. This ensures that the training process is not limited to focusing solely on the superiority / inferiority information of the final answer, such as the superiority / inferiority relationship, but also considers the argumentation logic of the answer's superiority / inferiority information based on the argumentation information. Specifically, this argumentation information provides the model with additional feature information, enabling the model to more accurately capture the key elements that need to be focused on when comparing the superiority / inferiority of answers. This optimizes the model's understanding and processing of answer quality, allowing the model to learn more efficiently and improve its overall performance, thereby enhancing the model's training effect and generalization ability. Consequently, the trained information processing model can more accurately judge the superiority / inferiority of answers, thus improving the accuracy of the trained information processing model in predicting the superiority / inferiority of answers.

[0254] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0255] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the model training methods or information processing methods provided in embodiments of this application. For example, the instructions can execute the following steps:

[0256] Obtain the information processing model to be trained and the training sample set. The training sample set includes sample questions, multiple sample answers to the sample questions, superiority / inferiority information labels, and argumentation information labels. The superiority / inferiority information labels include relation labels representing the superiority / inferiority relationships between sample answers, and the argumentation information labels are the information labels representing the superiority / inferiority relationships of arguments. The information processing model to be trained processes the sample questions and sample answers to obtain predicted superiority / inferiority information and predicted argumentation information. The predicted superiority / inferiority information includes the predicted superiority / inferiority relationships corresponding to the relation labels. Based on the differences between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the differences between the predicted argumentation information and the argumentation information labels, determine the prediction loss. Use the prediction loss to train the information processing model to be trained, obtaining the trained information processing model.

[0257] Alternatively, a post-trained information processing model, a problem to be processed, and multiple answers to the problem to be processed can be obtained. The post-trained information processing model is obtained through any of the model training methods provided in the embodiments of this application. The problem to be processed and the answers to be processed are processed through the post-trained information processing model to obtain target superiority information. The target superiority information includes target superiority relationships used to represent the superiority relationships between the answers to be processed.

[0258] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0259] According to one aspect of this application, a computer program product or computer program is provided, comprising a computer program or instructions that, when executed by a processor, implement the steps of the methods provided in the various optional implementations of the above embodiments. The computer program / instructions are stored in a computer-readable storage medium. A processor of an electronic device reads the computer program / instructions from the computer-readable storage medium and executes the computer program / instructions, causing the electronic device to perform the methods provided in the various optional implementations of the above embodiments.

[0260] Since the instructions stored in the storage medium can execute the steps of any model training method or information processing method provided in the embodiments of this application, the beneficial effects that any model training method or information processing method provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0261] The above provides a detailed description of a model training method, information processing method, and related equipment provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A model training method, characterized in that, include: Obtain the information processing model to be trained and the training sample set. The training sample set includes a sample question, multiple sample answers to the sample question, superiority and inferiority information labels, and argumentation information labels. The superiority and inferiority information labels include relational labels used to represent the superiority and inferiority relationship between the sample answers. The argumentation information labels are information labels that demonstrate the superiority and inferiority relationship. The sample question and the sample answer are processed by the information processing model to be trained to obtain prediction quality information and prediction argument information. The prediction quality information includes the prediction quality relationship corresponding to the relationship label. Based on the difference between the predicted superiority / inferiority information and the superiority / inferiority information label, and the difference between the predicted argumentation information and the argumentation information label, the prediction loss is determined; The information processing model to be trained is trained using the prediction loss to obtain the trained information processing model.

2. The model training method as described in claim 1, characterized in that, The process of processing the sample question and the sample answer using the information processing model to be trained to obtain prediction quality information and prediction argument information includes: The first feature processing is performed on the sample to be processed by the information processing model to be trained to obtain the first original score of the sample to be processed. The sample to be processed includes the sample question and the sample answer, and each sample to be processed corresponds to one sample answer. By comparing the first original scores, the predictive merit relationship between the sample answers in the sample to be processed is obtained; The second feature processing is performed on the sample to be processed using the information processing model to be trained, and the second original score of the sample to be processed is obtained. The predicted argument information is determined from the second original score.

3. The model training method as described in claim 2, characterized in that, The step of obtaining the prediction superiority / inferiority relationship between the sample answers in the sample to be processed by comparing the first original scores includes: For any two samples to be processed, based on the relationship label, determine the positive sample and the negative sample from the two samples to be processed; The difference between the first original score of the positive sample and the first original score of the negative sample is calculated to obtain the prediction superiority relationship between the sample answers in any two samples to be processed.

4. The model training method as described in claim 2, characterized in that, The superiority / inferiority information label also includes a superiority / inferiority category label, and the predicted superiority / inferiority information also includes the category prediction probability corresponding to the superiority / inferiority category label. The step of processing the sample question and the sample answer using the information processing model to be trained to obtain the predicted superiority / inferiority information and the predicted argument information further includes: The first original score is subjected to probability transformation to obtain the predicted probability of the category.

5. The model training method as described in claim 2, characterized in that, After performing first feature processing on the sample to be processed through the information processing model to be trained to obtain the first raw score of the sample to be processed, the process further includes: The standard deviation of the sample to be processed is calculated using the first original score. Based on the standard deviation, the target parameter value corresponding to the sample to be processed is determined. The target parameter value is used to adjust the prediction quality information corresponding to the sample to be processed, so that the adjusted prediction quality information is used to determine the prediction loss.

6. The model training method as described in claim 1, characterized in that, After processing the sample question and the sample answer using the information processing model to be trained to obtain prediction quality information and prediction argument information, the process further includes: Determine the cumulative deviation number corresponding to the sample answer, whereby the cumulative deviation number represents the cumulative number of times the predicted superiority / inferiority information corresponding to the sample answer and the corresponding superiority / inferiority information label satisfy a preset deviation. When the cumulative deviation number corresponding to the sample answer exceeds a preset number, the good and bad information label corresponding to the sample answer is reversed to obtain the reversed label, which is used to determine the prediction loss.

7. The model training method as described in claim 1, characterized in that, The prediction loss includes superiority / inferiority information loss and argumentation information loss. The determination of prediction loss based on the difference between the predicted superiority / inferiority information and the superiority / inferiority information labels, and the difference between the predicted argumentation information and the argumentation information labels, includes: The loss of the superior and inferior information is determined based on the difference between the predicted superior and inferior information and the superior and inferior information label; Based on the difference between the predicted argument information and the argument information label, the loss of the argument information is determined; The predicted loss is obtained by combining the loss of superior and inferior information and the loss of argumentation information.

8. The model training method as described in claim 1, characterized in that, The predicted argument information and the argument information label are text information. The step of determining the argument information loss based on the difference between the predicted argument information and the argument information label includes: Determine the word differences between the predicted word and the target word in the predicted argument information, wherein the target word is the word corresponding to the predicted word in the argument information tag; The loss of argumentative information is determined by the differences in the terms.

9. The model training method according to any one of claims 1 to 8, characterized in that, After training the information processing model to be trained using the prediction loss to obtain the trained information processing model, the process further includes: Obtain a test sample set, which includes test questions, multiple test answers to the test questions, and test superiority / inferiority information tags. The test superiority / inferiority information tags include test relationship tags used to represent the superiority / inferiority relationship between the test answers. The trained information processing model processes the test questions and the test answers to obtain test prediction quality information, which includes the test prediction quality relationship corresponding to the test relationship label. Based on the difference between the predicted superiority / inferiority information used for testing and the labels of the superiority / inferiority information used for testing, the evaluation index score of the post-trained information processing model is determined, and the evaluation index score is used to evaluate the performance of the post-trained information processing model.

10. An information processing method, characterized in that, include: The method involves obtaining a post-trained information processing model, a problem to be processed, and multiple unprocessed answers to the problem to be processed, wherein the post-trained information processing model is obtained by the method described in any one of claims 1 to 9. The trained information processing model processes the question and the answer to be processed to obtain target superiority information, which includes target superiority relationships used to represent the superiority relationships between the answers to be processed.

11. A model training device, characterized in that, include: The training acquisition unit is used to acquire the information processing model to be trained and the training sample set. The training sample set includes a sample question, multiple sample answers to the sample question, superiority and inferiority information labels, and argumentation information labels. The superiority and inferiority information labels include relational labels used to represent the superiority and inferiority relationship between the sample answers, and the argumentation information labels are information labels that demonstrate the superiority and inferiority relationship. The sample processing unit is used to process the sample question and the sample answer through the information processing model to be trained, and obtain prediction superiority information and prediction argumentation information, wherein the prediction superiority information includes the prediction superiority relationship corresponding to the relationship label; The loss determination unit is used to determine the prediction loss based on the difference between the predicted superiority / inferiority information and the superiority / inferiority information label, and the difference between the predicted argumentation information and the argumentation information label; The training unit is used to train the information processing model to be trained using the prediction loss, so as to obtain the trained information processing model.

12. An information processing device, characterized in that, include: A processing and acquisition unit is used to acquire a post-trained information processing model, a problem to be processed, and multiple unprocessed answers to the problem to be processed, wherein the post-trained information processing model is obtained by the method described in any one of claims 1 to 10; The information processing unit is used to process the question to be processed and the answer to be processed through the trained information processing model to obtain target superiority information, wherein the target superiority information includes a target superiority relationship used to represent the superiority relationship between the answers to be processed.

13. An electronic device, characterized in that, It includes a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to perform steps in the model training method or information processing method as described in any one of claims 1 to 10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute the steps of the model training method or information processing method according to any one of claims 1 to 10.

15. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the steps of the model training method or information processing method according to any one of claims 1 to 10.