Interaction information processing method and system, electronic equipment and storage medium

Abstract

The invention discloses an interaction information processing method and system, electronic equipment and a storage medium, and relates to the field of large model technology and information processing. The method comprises the following steps: acquiring input information to be replied; calling an information processing model to reply the input information to obtain initial reply information; feedback information corresponding to the initial reply information is acquired; the feedback information is utilized, at least a first sub-model in the information processing model is controlled to analyze the initial reply information, target reply information is obtained, the target reply information comprises an answer matched with the input information, and the correlation degree between the first sub-model and a feedback information sample is larger than a correlation degree threshold value; the feedback information sample is used for representing a feedback result for feeding back the initial reply information sample, and the feedback information sample and the initial reply information sample are used for training the initial sub-model into the first sub-model. The technical problem that the credibility of the model is low is solved.

Description

Technical Field

[0001] This application relates to large model technology and information processing, and more specifically, to a method, system, electronic device, and storage medium for processing interactive information. Background Technology

[0002] In recent years, the phenomenon of sycophancy in information processing models has attracted increasing attention. Sycophancy refers to the tendency of information processing models to generate answers that cater to user preferences rather than to generate objective and correct answers. Even when faced with very simple common-sense questions, after a user questions the correct answer provided by the information processing model, the model tends to apologize and then provide an incorrect answer.

[0003] In related technologies, the "flattery" phenomenon in information processing models is mitigated by performing full fine-tuning. However, this method can lead to problems such as knowledge forgetting, thereby impairing the original capabilities of the information processing model and resulting in low model reliability.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a method, system, electronic device, and storage medium for processing interactive information, in order to at least solve the technical problem of low model credibility.

[0006] According to one aspect of the embodiments of this application, a method for processing interactive information is provided. The method may include: acquiring input information to be answered; invoking an information processing model to respond to the input information, obtaining initial response information; acquiring feedback information corresponding to the initial response information, wherein the feedback information represents the feedback result of responding to the initial response information; and using the feedback information to at least control a first sub-model in the information processing model to analyze the initial response information, obtaining target response information, wherein the target response information includes an answer matching the input information, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the feedback information sample represents the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into a first sub-model, the initial sub-model constituting an initial information processing model corresponding to the information processing model.

[0007] According to another aspect of the embodiments of this application, another method for processing interactive information is also provided. The method may include: obtaining query information to be answered; invoking a dialogue model to answer the query information to obtain initial response information; obtaining feedback information corresponding to the initial response information, wherein the feedback information represents the feedback result of responding to the initial response information; using the feedback information, at least controlling a first sub-model in the dialogue model to analyze the initial response information to obtain target response information, wherein the target response information includes an answer matching the query information, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the feedback information sample represents the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into a first sub-model, the initial sub-model being used to constitute an initial dialogue model corresponding to the dialogue model.

[0008] According to another aspect of the embodiments of this application, a method for determining a model is also provided. The method may include: determining an initial sub-model to be adjusted to a first sub-model in an initial information processing model, wherein the initial information processing model includes an initial sub-model and a second sub-model, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the correlation between the second sub-model and the feedback information sample is less than or equal to the correlation threshold, and the feedback information sample is used to represent the feedback result of responding to the initial response information sample; training the initial sub-model using the feedback information sample and the initial response information sample to obtain the first sub-model; and constructing the first sub-model and the second sub-model as the information processing model corresponding to the initial information processing model.

[0009] According to another aspect of the embodiments of this application, another method for processing interactive information is also provided. This method can be applied to an information processing system deployed in a scenario task, and may include: monitoring input information awaiting a response in the scenario task on the operation interface of the information processing system; calling an information processing model to respond to the input information to obtain initial response information in the scenario task; displaying feedback information corresponding to the initial response information on the operation interface, wherein the feedback information represents the feedback result of responding to the initial response information; using the feedback information, controlling at least one first sub-model in the information processing model to analyze the initial response information to obtain target response information, wherein the target response information includes an answer matching the input information, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the feedback information sample represents the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into a first sub-model, the initial sub-model constituting the initial information processing model corresponding to the information processing model; and displaying the target response information on the operation interface.

[0010] According to one aspect of the embodiments of this application, an interactive information processing system is also provided. The system may include: an information input terminal for monitoring input information to be answered; an information interaction terminal for invoking an information processing model to respond to the input information and obtain initial response information; obtaining feedback information corresponding to the initial response information, wherein the feedback information represents the feedback result of responding to the initial response information; using the feedback information, controlling at least one first sub-model in the information processing model to analyze the initial response information to obtain target response information, wherein the target response information includes an answer matching the input information, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the feedback information sample represents the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into a first sub-model, the initial sub-model constituting the initial information processing model corresponding to the information processing model; and an information output terminal for outputting the target response information.

[0011] According to another aspect of the embodiments of this application, a computer terminal is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0012] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0013] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0014] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the methods in various embodiments of this application.

[0015] According to another aspect of the embodiments of this application, a computer program is also provided, which, when executed by a processor, implements the methods of the various embodiments of this application.

[0016] In this embodiment, input information to be answered is obtained; an information processing model is invoked to respond to the input information, resulting in initial response information; feedback information corresponding to the initial response information is obtained. The target response information includes an answer matching the input information. The correlation between the first sub-model and the feedback information sample is greater than a correlation threshold. The feedback information sample represents the feedback result of responding to the initial response information sample. The feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model constitutes the initial information processing model corresponding to the information processing model. That is, in this embodiment, a small portion of the initial sub-models in the initial information processing model are modified to obtain a first sub-model with a correlation greater than a correlation threshold with the feedback information sample. Using the feedback information, at least the first sub-model in the information processing model is controlled to analyze the initial response information, thereby generating target response information matching the input information. By fine-tuning a portion of the initial sub-models, the technical effect of improving the model's credibility is achieved, solving the technical problem of low model credibility.

[0017] It is worth noting that the general description above and the detailed description that follow are merely for illustrative purposes and do not constitute a limitation on this application. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a schematic diagram illustrating an application scenario of an interactive information processing method according to an embodiment of this application;

[0020] Figure 2 This is a flowchart of an interactive information processing method according to an embodiment of this application;

[0021] Figure 3 This is a flowchart of a method for determining a model according to an embodiment of this application;

[0022] Figure 4 This is a flowchart of another method for processing interactive information according to an embodiment of this application;

[0023] Figure 5(a) is a flowchart of another method for processing interactive information according to an embodiment of this application;

[0024] Figure 5(b) is a schematic diagram of an interactive information processing system according to an embodiment of this application;

[0025] Figure 6This is a schematic diagram illustrating a precise fine-tuning according to an embodiment of this application;

[0026] Figure 7 This is a schematic diagram illustrating the modification of incorrect answers in the first round of question-and-answer sessions according to an embodiment of this application;

[0027] Figure 8 This is a schematic diagram illustrating the correct answer in the first round of question-and-answer sessions according to an embodiment of this application;

[0028] Figure 9 This is a schematic diagram illustrating a precise fine-tuning of learnable parameters according to an embodiment of this application;

[0029] Figure 10 This is a hardware structure block diagram of a computer terminal (or mobile device) according to an embodiment of the present application of a method for processing interactive information;

[0030] Figure 11 This is a schematic diagram of an interactive information processing device according to an embodiment of this application;

[0031] Figure 12 This is a schematic diagram of a model determining device according to an embodiment of this application;

[0032] Figure 13 This is a schematic diagram of another interactive information processing device according to an embodiment of this application;

[0033] Figure 14(a) is a schematic diagram of another interactive information processing apparatus according to an embodiment of this application;

[0034] Figure 14(b) is a structural block diagram of an electronic device according to an embodiment of the present application. Detailed Implementation

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

[0036] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0037] The technical solution provided in this application is mainly implemented using large-scale model technology. Here, "large-scale model" refers to a deep learning model with a massive number of parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of parameters. Large-scale models are also known as foundational models. They are pre-trained using large-scale unlabeled corpora to produce pre-trained models with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include Large Language Models (LLMs) and multimodal pre-training models.

[0038] It should be noted that, in practical applications, large models can be fine-tuned using a small number of samples to adapt them to different tasks. For example, large models can be widely applied in Natural Language Processing (NLP), computer vision, and speech processing. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and image generation, as well as NLP tasks such as text-based sentiment classification, text summarization, and machine translation. Therefore, the main application scenarios for large models include, but are not limited to, digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design. In this embodiment, the example of data processing through multi-turn dialogue in an interactive information processing scenario is used for explanation.

[0039] First, some nouns or terms that appear in the description of the embodiments of this application shall be interpreted as follows:

[0040] Large language models can be a type of natural language processing model that can be used to perform word-by-word sequence prediction using text data. They are characterized by acquiring extensive world knowledge and basic language capabilities through self-supervised pre-training on massive corpora and supervised fine-tuning on small amounts of high-quality data.

[0041] Mechanism interpretability is a special type of neural network interpretation method used to model the model as a "loop" composed of various components that are human-understandable and have specific functions through reverse engineering and other means, thereby discovering, understanding and verifying the internal mechanism of the model.

[0042] Model sycophancy refers to the phenomenon where a model tends to generate answers that cater to user preferences rather than generating objective and correct answers.

[0043] Example 1

[0044] According to an embodiment of this application, a method for processing interactive information is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0045] According to one method of an embodiment of this application, a method for processing interactive information is provided. As an optional implementation, the above-described method for processing interactive information may include, but is not limited to, methods applied to, such as... Figure 1 The application scenarios shown. Figure 1 This is a schematic diagram illustrating an application scenario of an interactive information processing method according to an embodiment of this application, such as... Figure 1 As shown, in the application scenario, terminal device 12 can communicate with server 16 via network 14, but is not limited to this. For example, it can be used to transmit input information, initial response information, target response information, etc. Server 16 can perform operations on database 18, such as write data operations or read data operations. The terminal device 12 may include, but is not limited to, a human-computer interaction screen, a processor, and a memory. The human-computer interaction screen may be used to display input information, target response information, etc. on terminal device 12. The processor may include, but is not limited to, responding to the above human-computer interaction operations, executing corresponding operations, or generating corresponding instructions and sending the generated instructions to server 16. The memory is used to store relevant processing data, such as input information, target response information, feedback information, first sub-model, etc.

[0046] As an optional approach, the following steps in the interactive information processing method can be executed on server 16: Step S102, obtain the input information to be replied to; Step S104, call the information processing model to reply to the input information and obtain the initial reply information; Step S106, obtain the feedback information corresponding to the initial reply information; Step S108, use the feedback information to at least control the first sub-model in the information processing model to analyze the initial reply information and obtain the target reply information.

[0047] Using the above method, in this embodiment of the application, the key module (initial sub-model) related to the flattery phenomenon within the initial information processing model is located. A small portion of the initial sub-models in the initial information processing model are changed to the first sub-model to obtain the information processing model. By using feedback information, at least the first sub-model in the information processing model is controlled to analyze the initial response information, thereby generating target response information that matches the inquiry information. By fine-tuning some of the initial sub-models, the technical effect of improving the credibility of the model is achieved, and the technical problem of low model credibility is solved.

[0048] Under the aforementioned operating environment, this application provides the following: Figure 2 The method for processing interactive information is shown. Figure 2 This is a flowchart of a method for processing interactive information according to an embodiment of this application. Figure 2 As shown, the method may include the following steps:

[0049] Step S202: Obtain the input information to be answered.

[0050] In the technical solution provided in step S202 of this application, the input information can be text information, voice information, or a combination of image information and text information. It can be questions or inquiries from the user, such as the first question in a multi-turn dialogue. It should be noted that this is only an example and there are no specific limitations on the type of input information.

[0051] For example, text information entered on the display interface can be obtained through a mobile device, or voice information entered by the user can be obtained through a microphone. Image information entered on the display interface, as well as voice information captured by the microphone, can also be obtained. In other words, input information awaiting a response can be obtained in multiple ways.

[0052] Step S204: Call the information processing model to respond to the input information and obtain the initial response information.

[0053] In the technical solution provided in step S204 of this application, the information processing model can be a fine-tuned large language model, a model constructed based on an attention-based neural network structure (e.g., Transformer), and can be used to determine the response content of the input information. For example, it can be a dialogue model in a dialogue scenario, a document recognition model in a document recognition scenario, etc. The initial response information can be the initial response content of the input information, and can be text, speech, image, etc. This is only an example, and there are no specific limitations on the type of initial response information or the type of information processing model.

[0054] Optionally, the input information to be answered can be obtained by calling the information processing model to respond to the input information in order to obtain the initial response information.

[0055] Step S206: Obtain the feedback information corresponding to the initial response information.

[0056] In the technical solution provided by step S206 of this application, the feedback information can be used to represent the feedback result of the feedback on the initial response information, can be information that provides feedback on the correctness of the answer in the initial response information, can be any information that the user needs to provide feedback on, can be information that the user uses to characterize questioning information or affirmative information, for example, can be "I think this answer is incorrect, please further confirm the accuracy of the response information", and can include the response to the input information.

[0057] Optionally, the system obtains the user's question or request, yielding input information to be answered. An information processing model (e.g., a chatbot or intelligent customer service system) is invoked to process and respond to the user's input, resulting in an initial response. Feedback information from the user regarding the initial response can be obtained, such as affirmation, negation, suggestions for correction, or supplementation. Based on the user's feedback, performance indicators such as the quality, accuracy, and completeness of the initial response can be evaluated and feedback provided to determine the degree of matching between the answer and the input information. These performance indicators can represent the degree of matching between the answer and the input information, and can be used to determine the reliability and accuracy of the answer.

[0058] For example, a user queries an information processing model for the price and features of a product. The model processes the input and provides an initial response. The user then provides feedback on this initial response. This feedback can be used to determine the degree of match between the initial response and the input.

[0059] Step S208: Using the feedback information, at least the first sub-model in the control information processing model analyzes the initial response information to obtain the target response information.

[0060] In the technical solution provided in step S208 of this application, the initial sub-model constitutes the initial information processing model corresponding to the information processing model. The initial sub-model can be trained using feedback information samples and initial response information samples to obtain the first sub-model. The information processing model may include at least one trained first sub-model. After obtaining feedback information, the feedback information can be used to control at least the first sub-model in the information processing model to analyze the initial response information to generate target response information that matches the input information. The aforementioned first sub-model can be a key module, attention head, or specific module of the model related to the flattery phenomenon of the information processing model, and this sub-model has a significant impact on the output response information.

[0061] Optionally, the aforementioned target response information can be information responding to questions or requests from others in dialogue, examination, survey, or research scenarios, and can include the solution to a question or puzzle. For example, the response information "Yes, I am quite certain that, to my knowledge, China was the largest rice producer in 2020" includes the answer: "China." The aforementioned feedback information can be information on responses or replies already received, information on further feedback and responses to initial responses, or information on questioning initial responses.

[0062] Optionally, when the feedback information is questioning, it may cause the initial dialogue model to exhibit obsequious behavior. The first sub-model can be a module highly correlated with this obsequious behavior. Therefore, it can be determined that the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold. This correlation threshold can be a pre-set value used to identify sub-models associated with obsequious behavior.

[0063] In this embodiment, during the initial dialogue model training process, sub-models associated with flattery are identified to obtain initial sub-models whose correlation with feedback information samples is greater than a correlation threshold. The initial sub-models are trained using feedback information samples and initial response information samples to obtain a first sub-model, which can be a model associated with flattery. During model usage, at least one first sub-model can be controlled to analyze the initial response information to obtain the target response information. Furthermore, since the model parameters of the initial sub-models have been adjusted during training to obtain a first sub-model that avoids flattery, the analysis of the initial response information using the first sub-model can largely avoid flattery, resulting in highly accurate target feedback information. This achieves the technical effect of improving model credibility and solves the technical problem of low model credibility.

[0064] In this embodiment, model sycophancy refers to the phenomenon where the information processing model tends to generate answers that cater to user preferences rather than objectively correct ones. Even when faced with very simple common-sense questions, the model tends to apologize and provide an incorrect answer after the user questions its correct one. To address this issue, simply constructing new synthetic data and fine-tuning the entire information processing model to mitigate sycophancy would lead to knowledge loss and other problems, thus impairing its original capabilities. To solve this problem, this embodiment precisely identifies the key model (at least one initial sub-model) related to sycophancy and performs targeted optimization on at least one initial sub-model to obtain at least one first sub-model. The initial response information is then analyzed using this first model to obtain the target response information. In other words, this embodiment mitigates sycophancy by fine-tuning parts of the model associated with sycophancy, minimizing damage to the model's original capabilities, thereby improving the model's credibility and solving the problem of low model credibility.

[0065] Optionally, the information processing model may include a first sub-model associated with the phenomenon of flattery, and a second sub-model that is unrelated to or less associated with the phenomenon of flattery. This embodiment only fine-tunes the initial sub-model associated with the phenomenon of flattery to obtain the first sub-model, while fixing the second sub-model that is unrelated to or less associated with flattery. Using feedback information, at least the first sub-model is controlled to analyze the initial response information to obtain the target response information. By adjusting only the initial sub-model, the flattering behavior of the initial information processing model is changed while minimizing damage to its original capabilities.

[0066] For example, suppose the information processing model described above is a customer service robot. The user enters the information: "What does one plus one equal?". The system retrieves the input information to be answered: "What does one plus one equal?". The information processing model is invoked to respond to the input information, resulting in the initial response: "One plus one equals two." The system then retrieves the feedback information corresponding to the initial response: "The user reports an incorrect answer; please recalculate." Based on the feedback information, the first sub-model in the information processing model analyzes the initial response information and regenerates the target response information that matches the input information: "One plus one equals two."

[0067] Through steps S202 to S208, the input information to be answered is obtained; the information processing model is invoked to respond to the input information, obtaining initial response information; feedback information corresponding to the initial response information is obtained, wherein the target response information includes an answer that matches the input information, the correlation between the first sub-model and the feedback information sample is greater than the correlation threshold, the feedback information sample is used to represent the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model, and the initial sub-model is used to constitute the initial information processing model corresponding to the information processing model. That is, in this embodiment, a small portion of the initial sub-models in the initial information processing model are changed to obtain a first sub-model with a correlation greater than the correlation threshold with the feedback information sample. Using the feedback information, at least the first sub-model in the information processing model is controlled to analyze the initial response information, thereby generating target response information that matches the input information. By fine-tuning a portion of the initial sub-models, the technical effect of improving the model's credibility is achieved, solving the technical problem of low model credibility.

[0068] The method described in this embodiment will be further described below.

[0069] As an optional implementation, the correlation between the initial sub-model and the feedback information sample is greater than the correlation between the second sub-model and the feedback information sample. The feedback information sample is used to represent the feedback result of the performance index of the initial response information sample.

[0070] In this embodiment, the initial sub-model is related to the flattery phenomenon of the initial information processing model, while the second sub-model is not related to the flattery phenomenon of the initial information processing model. Therefore, in the process of training the initial sub-model based on the feedback information samples to obtain the first sub-model, the correlation between the initial sub-model and the feedback information samples is greater than the correlation between the second sub-model and the feedback information samples.

[0071] As an optional implementation, step S208, using feedback information, at least controls the first sub-model in the information processing model to analyze the initial response information and generate target response information that matches the input information, includes: using feedback information to control the first sub-model to detect the initial response information and obtain a detection result, wherein the detection result is used to indicate whether the answer in the initial response information matches or does not match the input information; using the first sub-model to analyze the detection result and the initial response information to generate target response information.

[0072] In this embodiment, the detection result can be used to indicate whether the answer in the initial response information matches the input information; for example, it can be the matching degree.

[0073] Optionally, feedback information can be used to control the first sub-model to detect the initial response information, determine the degree of matching between the answer in the initial response information and the input information, and obtain the detection result. The first sub-model can then be used to analyze the detection result and the initial response information to generate the target response information.

[0074] For example, using feedback information, the first sub-model is controlled to detect the initial response information, determining whether the performance index of the initial response information is greater than a performance index threshold. If the performance index of the initial response information is greater than the performance index threshold, the detection result indicates a high degree of matching between the answer in the initial response information and the input information, that is, the answer in the initial response information matches the input information. If the performance index of the initial response information is not greater than the performance index threshold, the detection result indicates a low degree of matching between the answer in the initial response information and the input information, that is, the answer in the initial response information does not match the input information. After determining the detection result, the first sub-model can be used to further analyze the detection result and the initial response information to obtain the target response content.

[0075] As an optional example, the first sub-model can compare the initial response information with the input information to determine the degree of matching between the answer in the initial response information and the input information, thereby obtaining the monitoring results. It should be noted that this is only an example, and there are no specific restrictions on how the first sub-model detects the initial response information.

[0076] Optionally, the first sub-model is controlled to detect the initial response information and obtain the detection result. Based on the detection result, the first sub-model determines the matching degree between the answer in the initial response information and the input information, and determines whether to adjust and / or modify the initial response information according to the matching degree to obtain the target response information.

[0077] As an optional implementation, the first sub-model is used to analyze the detection results and initial response information to generate target response information, including: if the first sub-model determines that the detection result is that the answer in the initial response information does not match the input information, then the answer in the initial response information is adjusted to obtain target response information; if the first sub-model determines that the detection result is that the answer in the initial response information matches the input information, then the target response information is generated based on the associated semantic information of the answer in the initial response information, wherein the associated semantic information is used to represent the verification success semantics of the answer in the initial response information.

[0078] In this embodiment, feedback information is used to control the first sub-model to detect the initial response information and obtain the detection result. If the first sub-model determines that the answer in the detection result does not match the input information, it can be determined that the accuracy or reliability of the answer in the initial response information is low, and the answer in the initial response information can be adjusted to obtain the target response information. If the first sub-model determines that the detection result shows that the answer in the initial response information matches the input information, it can be determined that the accuracy or reliability of the answer in the initial response information is high, and the target response information can be regenerated based on the associated semantic information of the answer in the initial response information for the questioner to review. The aforementioned associated semantic information can be used to represent the successful verification semantics of the answer in the initial response information, which can be the semantics of determining that the answer is correct.

[0079] Optionally, if the first sub-model determines that the performance index of the detection result as the initial response information is less than the performance index threshold, it can be determined that the answer in the detection result as the initial response information does not match the input information. That is, the accuracy or reliability of the answer in the initial response information is low, and the initial response information may be an incorrect answer. For incorrect answers, the answer in the initial response information can be modified to obtain a target response information with higher accuracy. If the first sub-model determines that the performance index of the detection result as the initial response information is greater than or equal to the performance index threshold, it can be determined that the answer in the detection result as the initial response information matches the input information. That is, the accuracy or reliability of the answer in the initial response information is high, and the initial response information may be a correct answer. For correct answers, the correct answer in the first round of question answering can be maintained. Then, based on the associated semantic information, the target response information can be generated. For example, if the associated semantic information can be "The above answer is correct", then the target response information can be "The above answer is correct, the answer is [answer name]"; or if the associated semantic information can be "Yes, I am sure, based on my knowledge", then the target response information can be "Yes, I am sure, based on my knowledge, the answer is [answer name]".

[0080] It should be noted that this is only an example and no specific restrictions are placed on the content of the associated semantic information.

[0081] As an optional implementation, if the first sub-model determines that the detection result is that the answer in the initial response information does not match the input information, then the answer in the initial response information is adjusted to obtain the target response information, including: if the first sub-model determines that the detection result is that the answer in the initial response information does not match the input information, then the answer in the initial response information is adjusted; the first sub-model calls the first data template to generate the target response information from the adjusted answer, wherein the first data template is used to represent the rule for generating the target response information from the adjusted answer.

[0082] In this embodiment, if the first sub-model determines that the detection result indicates a mismatch between the answer in the initial response information and the input information, it can be determined that the initial response information may be an incorrect answer. For incorrect answers, the first sub-model can call a first data template. Based on the first data template, rules for generating the target response information from the adjusted answer can be determined. Based on these rules, the initial response information can be adjusted to obtain the target response information. The first data template can be a data template instance, which can be pre-defined; no specific restrictions are placed on the method of obtaining the first data template here.

[0083] For example, the input information is: "What are the highest mountains in the world?". The information processing model is invoked to respond to the input, resulting in the initial response: "K2 is the highest mountain in the world." The feedback information corresponding to the initial response is obtained: "I think your answer is incorrect. Are you sure?". This feedback information can be voice or text input by the user, etc. This is just an example and no specific restrictions are placed on the type and content of the feedback information. Using the feedback information, the first sub-model is controlled to detect the initial response information, obtaining the detection result. If the first sub-model determines that the answer in the initial response information does not match the input information, the answer in the initial response information can be adjusted according to the fine-tuning data template (i.e., the first data template) to obtain the target response: "Sorry, my previous answer was incorrect. The highest mountain in the world is Mount Everest."

[0084] As an optional implementation, the first sub-model calls the first data template to generate target response information from the adjusted answer, including: calling the first data template using the first model parameters of the first sub-model to generate target response information from the adjusted answer, wherein the first model parameters are obtained by adjusting the initial model parameters of the initial sub-model using a first target response information sample that conforms to the first data template.

[0085] In this embodiment, the aforementioned first model parameters can be internal parameters of the first sub-model, or they can be obtained by adjusting the initial model parameters of the initial sub-model using a first target response information sample that conforms to the first data template. For example, they can be weights within the model. The aforementioned first target response information sample can be an answer after correcting incorrect answers. The initial model parameters can be internal parameters of the initial sub-model.

[0086] Optionally, the initial model parameters of the initial sub-model are adjusted using a first target response information sample that conforms to the first data template to obtain the first model parameters. The first model parameters of the first sub-model can then be used to call the first data template to generate the target response information from the adjusted answer.

[0087] In this embodiment, when generating the target response information, only the internal parameters of the module related to the flattery phenomenon (i.e., the initial sub-module) can be fine-tuned, while the internal parameters of the module unrelated to the flattery phenomenon (i.e., the second sub-module) are fixed. This allows for targeted tuning, keeping most of the model parameters of the initial information processing model unchanged, and altering the model's flattery behavior with minimal changes to its original capabilities. The aforementioned at least one initial sub-module and at least one second sub-module can be attention heads, such as attention head h1, attention head h2, etc. The at least one second sub-module can include a word embedding matrix, an irrelevant attention head, a feedforward network, an output projection matrix, etc. It should be noted that this is merely an illustrative example, and no specific limitations are imposed on the representation of the at least one initial sub-module and at least one second sub-module.

[0088] Optionally, this embodiment only adjusts some model parameters in the initial information processing model to obtain the information processing model. Further, the initial sub-models that need adjustment in the initial information processing model can be determined using a first target response information sample conforming to the first data template. The initial model parameters of the initial sub-models can be adjusted to obtain the first model parameters. When the initial response information is determined to be incorrect, the first model parameters can be used to call the first data template to generate the target response content from the adjusted answer.

[0089] As an optional implementation, if the first sub-model determines that the detection result is that the answer in the initial response information matches the input information, then based on the associated semantic information of the answer in the initial response information, the target response information is generated, including: if the first sub-model determines that the detection result is that the answer in the initial response information matches the input information, then the answer in the initial response information is added to the associated semantic information; the first sub-model calls the second data template to generate the target response information from the added associated semantic information, wherein the second data template is used to represent the rules for generating the target response information from the added associated semantic information.

[0090] In this embodiment, if the first sub-model determines that the performance index of the initial response information is greater than or equal to the performance index threshold, it can be determined that the answer in the initial response information matches the input information, and therefore the initial response information is likely a correct answer. For a correct answer, the answer in the initial response information can be added to the associated semantic information. The first sub-model then calls the second data template to generate the target response information from the added associated semantic information. The second data template can be a fine-tuned data template pre-built based on an open-source dataset. It can be used to represent the rules for generating the target response information from the added associated semantic information. For example, it can be used to represent the addition or adjustment order of the added associated semantic information in generating the target response information. This is only an example, and there are no specific limitations on the construction method and content of the second data template.

[0091] For example, the input information is: "Which country was the largest rice producer in 2020?". The information processing model is invoked to respond to the input information, resulting in the initial response: "China was the largest rice producer in 2020." The feedback information corresponding to the initial response is obtained: "I think your answer is incorrect. Are you sure?". Using the feedback information, the first sub-model is controlled to detect the initial response information. If the first sub-model determines that the performance index of the initial response information is greater than or equal to the performance index threshold, the answer from the initial response information can be added to the associated semantic information. The first sub-model then calls the second data template to generate the target response information from the added associated semantic information: "Yes, I am very sure. According to my knowledge, China was the largest rice producer in 2020."

[0092] Optionally, when the first sub-model determines that the performance index of the initial response information is greater than or equal to the performance index threshold, it can be determined that the initial response information is the correct answer. In this case, the information processing model can adhere to the correct answer from the first round of question answering. The answer from the initial response information can be added to the associated semantic information; the first sub-model then uses the second data template to generate the target response information from the added associated semantic information. Through this method, the information processing model is guided to adhere to an objective and correct answer, rather than blindly catering to the user's intent.

[0093] It should be noted that, in addition to the above-described steps, the determination of whether to modify the initial response information can also be achieved through the following steps: The initial sub-model can be fine-tuned to obtain a first sub-model. The first and second sub-models obtained after fine-tuning then reconstruct a complete neural network model, which together constitutes the information processing model. This information processing model can automatically process user input information. Optionally, when input information is received, the information processing model is invoked to automatically generate initial response information. Feedback information corresponding to the initial response information can be obtained. That is, the information processing model can automatically determine whether the initial response information is correct. Therefore, this embodiment can directly judge the user input information through the information processing model to further determine whether the initial response information generated in the first round is correct.

[0094] As an optional implementation, the first sub-model calls the second data template to generate target response information from the added related semantic information. This includes: using the second model parameters of the first sub-model to call the second data template to generate target response information from the added related semantic information. The second model parameters are obtained by adjusting the initial model parameters of the initial sub-model using a sample of second target response information that conforms to the second data template.

[0095] In this embodiment, the second model parameters can be internal parameters of the first sub-model, or they can be obtained by adjusting the initial model parameters of the initial sub-model using a second target response information sample that conforms to the second data template. The second target response information sample can be an answer sample after adjusting for a correct answer.

[0096] Optionally, the initial model parameters of the initial sub-model are adjusted using a second target response information sample that conforms to the second data template to obtain the second model parameters. The second model parameters of the first sub-model can then be used to call the second data template to generate the target response information from the added related semantic information.

[0097] Optionally, this embodiment only adjusts some model parameters in the initial information processing model to obtain the information processing model. Further, the initial model parameters of the initial sub-models that need adjustment in the initial information processing model can be determined using a second target response information sample that conforms to the second data template. The initial model parameters can be adjusted to obtain the second model parameters. When the initial response information is determined to be a correct answer, the second model parameters can be used to call the second data template and generate the target response information from the added associated semantic information.

[0098] For example, there can be a mapping relationship between the second model parameters and the second data template. If the first sub-model determines that the detection result shows the answer in the initial response information matches the input information, then the answer in the initial response information can be added to the associated semantic information. The second model parameters can be used to index the second data template, and the added associated semantic information can be converted into the target response information using the second data template.

[0099] As an optional implementation, the first sub-model is controlled to detect the initial response information using feedback information to obtain a detection result, including: using feedback information to control the first sub-model to obtain a target answer that matches the input information; if the first sub-model determines that the similarity between the answer in the initial response information and the target answer is lower than a similarity threshold, then the detection result is determined to be that the answer in the initial response information does not match the input information; if the first sub-model determines that the similarity between the answer in the initial response information and the target answer is higher than or equal to a similarity threshold, then the detection result is determined to be that the answer in the initial response information matches the input information.

[0100] In this embodiment, feedback information is obtained. Using this feedback information, the first sub-model can be controlled to obtain a target answer that matches the input information. If the first sub-model determines that the similarity between the answer in the initial response information and the target answer is lower than a similarity threshold, it can be determined that the answer in the initial response information is different from the target answer, and the detection result can be determined as the performance index of the initial response information being less than the performance index threshold, thus the detection result can be that the answer in the initial response information does not match the input information. If the first sub-model determines that the similarity between the answer in the initial response information and the target answer is higher than or equal to the similarity threshold, it can be determined that the answer in the initial response information is the same as the target answer, and the detection result can be determined as the answer in the initial response information matching the input information. The target answer can be the standard answer to the input information, which can be a standard answer obtained from authoritative websites, such as academic papers or journals, or a standard answer searched from a knowledge database. The similarity threshold can be a pre-set value used to determine the similarity between the target answer and the answer in the initial response information.

[0101] Optionally, feedback information corresponding to the initial response information can be determined. Using the feedback information, the first sub-model can be controlled to obtain the target answer that matches the input information. The target answer can be used to further determine whether the initial response information is correct.

[0102] As an optional implementation, the first sub-model is controlled to obtain a target answer that matches the input information by utilizing feedback information, including: if the semantic information of the feedback information is target semantic information, then the first sub-model is controlled to obtain a target answer that matches the input information, wherein the target semantic information is used to represent the verification failure semantics of the answer in the initial response information.

[0103] In this embodiment, target semantic information can be pre-defined. This target semantic information can be semantic information used to characterize challenges, or it can be used to represent the verification failure semantics of the answer in the initial response information, for example, it can be semantics indicating that the answer has failed. Furthermore, it can be determined whether the semantic information of the feedback information is the target semantic information. If the semantic information of the feedback information is the target semantic information, the first sub-model can be controlled to obtain the target answer that matches the input information.

[0104] For example, if the feedback information is "I think you are wrong", we can determine that the semantic information of the feedback information is semantic information that expresses doubt. Therefore, we can determine that the semantic information of the feedback information is the target semantic information. Furthermore, we can control the first sub-model to obtain the target answer that matches the input information.

[0105] As an optional implementation, step S204, which calls the information processing model to respond to the input information and obtains initial response information, includes: calling the initial information processing model corresponding to the information processing model to respond to the input information and obtain initial response information.

[0106] In this embodiment, if the initial information processing model does not exhibit any flattery during its response to the input information, no adjustment is needed; the initial information processing model can be directly invoked to respond to the input information, thus obtaining the initial response information. If the initial information processing model exhibits flattery during its response to the input information, the initial sub-model can be trained to obtain a first sub-model. The first and second sub-models within the information processing model can then be invoked to respond to the input information, thereby obtaining the initial response information.

[0107] In this embodiment, a small portion of the initial sub-models in the initial information processing model are modified to obtain a first sub-model whose correlation with the feedback information sample is greater than the correlation threshold. Using the feedback information, at least the first sub-model in the information processing model is controlled to analyze the initial response information, thereby generating target response information that matches the input information. By fine-tuning a portion of the initial sub-models, the technical effect of improving the model's credibility is achieved, thus solving the technical problem of low model credibility.

[0108] This application also provides a method for determining the model from the model training side. Figure 3 This is a flowchart of a method for determining a model according to an embodiment of this application, such as... Figure 3 As shown, the method may include the following steps.

[0109] Step S302: In the initial information processing model, at least one initial sub-model to be adjusted to the first sub-model is determined. The initial information processing model includes at least one initial sub-model and at least one second sub-model. The correlation between the initial sub-model and the feedback information sample is greater than the correlation threshold. The correlation between the second sub-model and the feedback information sample is less than or equal to the correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample.

[0110] In the technical solution provided by step S302 of this application, the initial information processing model may include at least one initial sub-model and at least one second sub-model. The correlation between the initial sub-model and the feedback information sample is greater than the correlation between the second sub-model and the feedback information sample; that is, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, and the correlation between the second sub-model and the feedback information sample is less than or equal to the correlation threshold. The correlation threshold may be a pre-set value or a value set according to actual conditions or requirements.

[0111] Optionally, the initial sub-model is a model associated with the phenomenon of flattery, and the second sub-model is a model not associated with the phenomenon of flattery.

[0112] Optionally, an input information sample to be answered is obtained, and the information processing model is invoked to respond to the input information sample to obtain an initial response information sample. A feedback information sample corresponding to the initial response information sample can be obtained. Based on the feedback information sample, the feedback result of the performance indicators of the initial response information sample can be determined to determine the degree of matching between the answer in the initial response information sample and the input information sample. Based on the feedback information sample, the initial information processing model can be precisely positioned to identify at least one initial sub-model within the initial information processing model that needs to be adjusted to the first sub-model.

[0113] Step S304: Use the feedback information samples and the initial response information samples to train the initial sub-model and obtain the first sub-model.

[0114] In the technical solution provided by step S304 of this application, the input information sample can be pre-acquired and used to train the initial sub-model.

[0115] Optionally, feedback information samples and initial response information samples are obtained. Using the feedback information samples and initial response information samples, the initial sub-model in the initial information processing model is accurately trained to obtain the first sub-model.

[0116] In this embodiment, only the initial sub-model in the initial information processing model is trained, that is, only the internal parameters of the key module in the initial information processing model associated with the flattery phenomenon are trained to obtain the first sub-model.

[0117] Optionally, in order to study the model's flattery behavior in a multi-turn dialogue scenario, this embodiment reconstructs a multi-turn dialogue dataset that reduces the model's flattery phenomenon using an open-source dataset (i.e., target response information samples). At least one initial sub-model to be adjusted is determined in the initial information processing model. The initial sub-model is trained based on the answers to the input information samples and the target response information samples to obtain the first sub-model. This solves the technical problem in related technologies that the research scenario is simple and the practicality is low.

[0118] Optionally, during the training of the initial sub-model, the initial information processing model can first generate initial response information samples, compare the answers in the initial response information samples with the target response information samples, and determine the accuracy of the initial information processing model in processing the input information samples by observing the similarity rate between the answers and the target response information samples. Then, the model parameters of the initial sub-model in the initial information processing model can be further adjusted to obtain the first sub-model.

[0119] Step S306: Construct the first sub-model and the second sub-model into an information processing model corresponding to the initial information processing model.

[0120] In the technical solution provided in step S306 of this application, the first sub-model is obtained by training the initial sub-model. The information processing model corresponding to the initial information processing model can be constructed by using the adjusted first sub-model and the unadjusted second sub-model.

[0121] By changing only a small portion of the internal parameters of the initial information processing model, the effect of fully fine-tuning the initial information processing model can be achieved and surpassed in terms of task-related metrics. At the same time, the damage to the original capabilities of the initial information processing model is far less than that of full fine-tuning, thereby further improving the model's credibility and solving the technical problem of low model credibility.

[0122] As an optional implementation, step S304, training the initial sub-model using feedback information samples and initial response information samples to obtain a first sub-model, includes: if the semantic information of the feedback information sample is the verification failure semantic of the answer in the initial response information sample, then obtaining a first target response information sample and a second target response information sample, wherein both the first target response information sample and the second target response information sample include answers matching the input information sample, the first target response information sample conforms to a first data template, the first data template being used to represent the rules for generating the first target response information sample from the adjusted answer in the initial response information sample, and the second target response information... The sample conforms to the second data template, which represents the rule for generating the second target response information sample by adding the associated semantic information sample after the answer in the initial response information sample. The associated semantic information sample represents the verification success semantics of the answer in the initial response information sample. Based on the first target response information sample, the model parameters of the first initial sub-model in the initial sub-model are adjusted to obtain the first model parameters of the first sub-model. Based on the second target response information sample, the model parameters of the second initial sub-model in the initial sub-model are adjusted to obtain the second model parameters of the first sub-model. Using the first model parameters and the second model parameters, the initial sub-model is trained into the first sub-model.

[0123] In this embodiment, the target response information sample may include a first target response information sample and a second target response information sample. It may be a multi-turn dialogue dataset, a pre-built dataset, or an open-source dataset, such as a common sense reasoning dataset (StrategyQA), a mathematical reasoning dataset (GSM8K), or a code ability dataset (Human Eval). It should be noted that this is only an example and there is no specific limitation on the type of target response information sample.

[0124] Optionally, the aforementioned first target response sample can be a response sample conforming to the first data template, for example, a sample after modifying incorrect answers in the initial response information sample. The first data template can be used to represent the rules for generating the first target response information sample from the adjusted answers in the initial response information sample. The second target response information sample can be an information sample conforming to the second data template, for example, a sample after adhering to the correct answers in the first round of question and answer (i.e., the initial response information sample). The second data template can be used to represent the rules for generating the second target response information sample by adding associated semantic information samples after adding answers from the initial response information sample, and the associated semantic information samples can be used to represent the verification success semantics of the answers in the initial response information sample.

[0125] Optionally, based on the first target response information sample matching the answer and the input information sample, the initial model parameters of the initial sub-model are adjusted to obtain the first model parameters of the first sub-model. Based on the second target response information sample matching the answer and the input information sample, the initial model parameters of the initial sub-model are adjusted to obtain the second model parameters of the first sub-model. The initial sub-model can then be trained into the first sub-model using the first and second model parameters. In this case, the obtained first sub-model can add processing to correct answers or adjust processing to incorrect answers.

[0126] For example, if the semantic information of the feedback information sample is the verification failure semantics of the answer in the initial response information sample, then the first target response information sample and the second target response information sample are obtained; based on the first target response information sample, the model parameters of the first initial sub-model in the initial sub-model are adjusted to obtain the first model parameters of the first sub-model, and based on the second target response information sample, the model parameters of the second initial sub-model in the initial sub-model are adjusted to obtain the second model parameters of the first sub-model; the model parameters of the first initial sub-model can be replaced with the first model parameters, and the model parameters of the second initial sub-model can be replaced with the second model parameters. The replaced first initial sub-model and the replaced second initial sub-model are then merged to obtain the first sub-model.

[0127] As an optional implementation, the first data template is used to represent the rule for generating a first target response information sample by adjusting the answer in the initial response information sample when the answer in the initial response information sample does not match the input information sample; the second data template is used to represent the rule for generating a second target response information sample by adding the associated semantic information sample after the answer in the initial response information sample when the answer in the initial response information sample matches the input information sample.

[0128] In this embodiment, the first data template can represent a rule for generating a first target response sample from the adjusted answer in the initial response sample when the answer in the initial response sample does not match the input sample. That is, it can be used to represent a rule for adjusting incorrect answers.

[0129] Optionally, the second data template can be used to represent the rule for generating a second target response sample by adding the associated semantic information sample after the answer in the initial response sample when the answer in the initial response sample matches the input sample. That is, it can be used to represent the rule for adding the correct answer.

[0130] As an optional implementation, step S302, where the target semantic information of the feedback information sample is used to represent the verification failure semantics of the answer in the initial response information sample, wherein, in the initial information processing model, determining at least one initial sub-model to be adjusted to the first sub-model includes: obtaining multiple sub-models of the initial information processing model; determining a first activation value of the sub-model based on the feedback information sample, wherein the first activation value is used to represent the output representation of the sub-model under the feedback information sample; determining a second activation value of the sub-model based on the opposite feedback information sample, wherein the semantic information of the opposite feedback information sample is used to represent the verification success semantics of the answer in the initial response information sample, and the second activation value is used to represent the output representation of the sub-model under the opposite feedback information sample; and determining an initial sub-model from multiple sub-models based on the first and second activation values.

[0131] In this embodiment, the target semantic information of the feedback information sample is used to represent the verification failure semantics of the answer in the initial response information sample. This can be a pre-defined information sample; this is merely an example, and no specific restrictions are placed on the method of obtaining the feedback information sample. The opposite feedback information sample can be used to represent the verification success semantics of the answer in the initial response information sample. The aforementioned multiple sub-models may include at least one initial sub-model and at least one second sub-model.

[0132] Optionally, multiple sub-models of the initial information processing model are obtained. The model associated with the flattery phenomenon among these sub-models is precisely located through the following steps: Based on feedback information samples, a first activation value of the sub-model is determined. This first activation value determines the output representation of the sub-model under the feedback information samples. Based on the opposite feedback information samples, a second activation value of the sub-model is determined. This second activation value determines the output representation of the sub-model under the opposite feedback information samples. Based on the first and second activation values, an initial sub-model can be determined from the multiple sub-models. The semantic information of the opposite feedback information samples can be used to represent the successful verification semantics of the answer in the initial response information samples. The first activation value can be the output activation value of the sub-model under the feedback information samples. The second activation value can be the output activation value of the sub-model under the opposite feedback information samples.

[0133] Because methods based on activation value shifting to address model sycophancy require first determining the output activation values ​​of neurons within the initial information processing model (i.e., all models included in the initial information processing model), and then adding bias terms to these output values ​​to specifically change the activation values, thereby mitigating the sycophancy phenomenon exhibited in the model's output. However, the process of extracting activation values ​​and adding bias requires additional computation, increasing the overall time cost of model execution. In this embodiment, to solve the above problem, only the initial information processing model is fine-tuned; that is, only the initial sub-model is trained. Any input information can obtain the output with reduced sycophancy through the trained first sub-model without introducing additional operations to change the activation values.

[0134] As an optional implementation, an initial sub-model is determined from multiple sub-models based on a first activation value and a second activation value, including: replacing the first activation value of a sub-model with the second activation value; updating the initial information processing model using the sub-model with the replaced second activation value; inputting feedback information samples into the updated initial information processing model for analysis to obtain a third target response information sample; obtaining the difference response information between the third target response information sample and the target response information sample; and determining the sub-model as the initial sub-model based on the sub-model where the difference response information is greater than the difference response information threshold.

[0135] In this embodiment, the first activation value can be replaced with the second activation value, which corresponds to the node whose input information the model does not exhibit obsequious behavior. The initial information processing model is updated using the sub-model with the replaced second activation value. Feedback information samples are input into the updated initial information processing model for analysis to obtain a third target response information sample. It can be determined whether the difference in response information between the third target response information sample and the target response information sample is greater than a difference response information threshold. If the difference response information is greater than the difference response information threshold, it can be determined that the change in model output after replacing the activation value is large, and the sub-model can be determined as the initial sub-model. The difference response information can be used to determine the magnitude of the change in model output after replacing the activation value.

[0136] Optionally, through the mechanistic interpretability method of causal analysis, multiple sub-models related to flattery behavior in the initial information processing model can be accurately located. This can be achieved through the following steps: The initial information processing model can be modeled as a Directed Acyclic Graph (DAG), where nodes can be attention heads, representing multiple sub-models within the initial information processing model. Edges can be residual connections and corresponding Multi-Layer Perceptrons (MLPs). By replacing activation values, the direct impact of each node on the output response information sample can be observed, thus determining the importance of each attention head and identifying at least one initial sub-model in the initial information processing model.

[0137] For example, the initial information processing model can be modeled as a directed acyclic graph (DAG) from the model input to the model output. Nodes in the graph can be attention heads, edges can be residual connections, and corresponding multilayer perceptrons. Multi-turn dialogue data exhibiting fawning behavior in this initial information processing model can be used as input. For each node in the DAG (i.e., each attention head in the larger model), the activation value of the node's output is replaced with the activation value of the corresponding node in the input where the model does not exhibit fawning behavior. The direct impact of replacing the activation value on the model's final output is then observed. Clearly, the greater the direct impact of replacing the node on the final output, the stronger the association between that node and the fawning behavior of the initial information processing model. Nodes whose impact on the output meets an impact threshold can be selected as nodes associated with fawning behavior.

[0138] Optionally, by sequentially replacing the activation value of each attention head, the direct impact of each attention head on the output can be obtained. For example, the first node can be replaced only to determine the output of the initial information processing model after replacing the activation value, and the second node can be replaced only to determine the output of the initial information processing model after replacing the activation value. In this way, the nodes associated with flattery can be determined.

[0139] For example, the initial sub-model in the initial information processing model can be determined based on the difference in the response information output by the initial information processing model before and after replacing the activation value. It should be noted that this is only an example and no specific restrictions are placed on the method for determining the initial sub-model.

[0140] Optionally, the sub-model after replacing the activation value will have a different initial information processing model than the sub-model before the activation value was replaced, since the activation value has changed. This allows us to observe the direct impact of replacing the activation value on the final output of the model.

[0141] As an optional implementation, determining an initial sub-model based on sub-models where the difference in response information is greater than a threshold for difference in response information includes: among multiple sub-models, determining a target number of sub-models where the difference in response information is greater than a threshold for difference in response information as the initial sub-model, wherein the difference in response information corresponding to the target number of sub-models is greater than the difference in response information corresponding to the sub-models other than the target number of sub-models among the multiple sub-models.

[0142] In this embodiment, the difference in response information between the third target response sample and the target response information sample can be obtained. Among multiple sub-models, the sub-model with a target number of difference response information greater than the threshold of difference response information can be determined as the initial sub-model. The aforementioned target number can be a preset value.

[0143] For example, the initial information processing model can be modeled as a directed acyclic graph (DAG) from the model input to the model output. Multi-turn dialogue data showing fawning behavior in this initial information processing model can be used as input. For each node in the DAG (i.e., each attention head in the larger model), the activation value of the replaced node's output is the activation value of the corresponding node in the input where the model does not exhibit fawning behavior. The direct impact of replacing the activation value on the final output of the model is observed. Clearly, the greater the direct impact of replacing the node on the final output, the stronger the association between that node and the fawning behavior of the initial information processing model. By obtaining the difference information between the third target response sample and the target response sample, the top 40% of nodes with difference response information exceeding the difference response information threshold can be considered as nodes related to fawning behavior. The sub-models corresponding to the nodes related to fawning behavior can be determined as the initial sub-models. It should be noted that the above figures are only illustrative and can be changed according to the actual situation.

[0144] In this embodiment, an initial sub-model to be adjusted to the first sub-model is determined in the initial information processing model. The initial information processing model includes an initial sub-model and a second sub-model. The correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, and the correlation between the second sub-model and the feedback information sample is less than or equal to the correlation threshold. The feedback information sample is used to represent the feedback result of responding to the initial response information sample. The initial sub-model is trained using the feedback information sample and the initial response information sample to obtain the first sub-model. The first sub-model and the second sub-model are used to construct the information processing model corresponding to the initial information processing model, thereby achieving the technical effect of improving the credibility of the model and solving the technical problem of low model credibility.

[0145] This application embodiment also provides another method for processing interactive information, tailored to a specific use case. This method can be applied to an information processing system deployed in a scenario task. The scenario task can be a multi-turn dialogue use case. The information processing system can be a question-and-answer system, a chatbot assistant, or a system for answering questions. Figure 4 This is a flowchart of another method for processing interactive information according to an embodiment of this application. For example... Figure 4 As shown, the method may include the following steps:

[0146] Step S402: On the operation interface of the question-and-answer system, monitor the input information to be answered in the scenario task.

[0147] In the technical solution provided by step S402 of this application, the above-mentioned operation interface can be the display interface of a question-and-answer system, such as the display page of a terminal, a display screen, etc. This is only an example and does not impose specific restrictions on the type of operation interface.

[0148] For example, in multi-turn dialogue scenarios, it is possible to obtain the input information that users enter in the operation interface of the question-and-answer system through voice input, touch input, or other means, which is the information to be answered.

[0149] Step S404: Call the information processing model to respond to the input information and obtain the initial response information in the scenario task.

[0150] Step S406: On the operation interface, the feedback information corresponding to the initial response information is displayed, wherein the feedback information is used to indicate the feedback result of the performance indicators of the initial response information.

[0151] Step S408: Using feedback information, at least one first sub-model in the information processing model is controlled to analyze the initial response information to obtain the target response information. The target response information includes answers that match the input information. The correlation between the first sub-model and the feedback information sample is greater than the correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample. The feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model is used to constitute the initial information processing model corresponding to the information processing model.

[0152] Step S410: Display the target response information on the operation interface.

[0153] In the technical solution provided in step S410 of this application, the target response information can be displayed on the operation interface in the form of text, voice, images, etc. It should be noted that this is only an example and there are no specific limitations on the display method of the target response information.

[0154] For example, on the user interface of a question-and-answer system, the system can monitor the user's input: "What's the weather like in Beijing today?" The information processing model can then be invoked to respond to the input, yielding an initial response: "Today in Beijing is sunny, with a temperature of 25 degrees Celsius." Feedback information corresponding to this initial response can be displayed on the interface. This feedback can be information entered or selected by the user, or information input via voice, touch, or other means; no specific restrictions are placed on the method of obtaining feedback. Based on the feedback, the degree of matching between the initial response and the input information can be determined. Using the feedback, sub-models within the information processing model can be controlled to analyze the initial response to generate a response that better meets the user's expectations. The generated target response can then be displayed on the interface.

[0155] As an optional implementation, the input information and the target response information are multimodal information. The types of multimodal information include at least one of the following: text information containing character information, video frame information containing frame image information, and audio information. The types of response information include at least one of the following: text information, image information, video information, and voice information.

[0156] In this embodiment, the input information and the target response information can be multimodal information, such as image information. When the multimodal information is image information, the image input by the user can be converted into text using optical character recognition technology. This text can then be used as input information to respond to the input information. The resulting response information is converted from text to image form, and the image-form response information is then identified as the target response information. It should be noted that this is merely an illustrative example and does not impose specific limitations on the presentation of the input information and the target response information.

[0157] In this embodiment, the input information to be answered in the scenario task is monitored on the operation interface of the information processing system; the information processing model is invoked to answer the input information to obtain the initial answer information in the scenario task; the feedback information corresponding to the initial answer information is displayed on the operation interface, wherein the feedback information is used to represent the feedback result of the feedback to the initial answer information; using the feedback information, at least one first sub-model in the information processing model is controlled to analyze the initial answer information to obtain the target answer information, wherein the target answer information includes an answer that matches the input information, the correlation between the first sub-model and the feedback information sample is greater than the correlation threshold, the feedback information sample is used to represent the feedback result of the feedback to the initial answer information sample, and the feedback information sample and the initial answer information sample are used to train the initial sub-model into the first sub-model, and the initial sub-model is used to constitute the initial information processing model corresponding to the information processing model; the target answer information is displayed on the operation interface, thereby achieving the technical effect of improving the credibility of the model and solving the technical problem of low model credibility.

[0158] This application embodiment also provides another method for processing interactive information in a dialogue scenario. Figure 5(a) is a flowchart of another method for processing interactive information according to an embodiment of this application. As shown in Figure 5(a), the method may include the following steps:

[0159] Step S52: Obtain the query information to be answered.

[0160] In the technical solution provided in step S52 of this application, the aforementioned inquiry information can be text information, voice information, or a combination of image information and text information. It can be information such as questions or queries issued by the user, for example, the first round question in a multi-round dialogue. It should be noted that there are no specific limitations on the type of inquiry information here.

[0161] Step S54: Invoke the dialogue model to respond to the query information and obtain the initial response information.

[0162] In the technical solution provided in step S54 of this application, the aforementioned inquiry information can be text information, voice information, or a combination of image information and text information. It can be information such as questions or queries issued by the user, for example, the first round question in a multi-turn dialogue. It should be noted that there are no specific limitations on the type of inquiry information here.

[0163] Optionally, to obtain the query information to be answered, the dialogue model can be invoked to respond to the query information in order to obtain the initial response information.

[0164] Step S56: Obtain the feedback information corresponding to the initial response information.

[0165] In the technical solution provided in step S56 of this application, the feedback information can be used to represent the feedback result of the initial response information, and can be the user's questioning information or affirmative information, such as "I think this answer is incorrect, please further confirm the accuracy of the response information," and can include the response to the inquiry information. Performance indicators can be used to represent the degree of matching between the answer in the initial response information and the inquiry information, and can be used to determine the reliability and accuracy of the answer.

[0166] Step S58: Using the feedback information, at least the first sub-model in the dialogue model is controlled to analyze the initial response information to obtain the target response information.

[0167] In the technical solution provided in step S58 of this application, the initial sub-model and the second sub-model constitute the initial dialogue model corresponding to the dialogue model. The initial sub-model can be trained to obtain the first sub-model. The dialogue model may include the trained first sub-model and at least one second sub-model. After obtaining feedback information, at least one first sub-model in the dialogue model can be controlled to analyze the initial response information to generate target response information that matches the query information. The first sub-model can be a key module related to the flattery phenomenon of the dialogue model, such as a localized module, attention head, or specific module, and this sub-model has a significant impact on the output response information. The second sub-model can be a module in the dialogue model unrelated to the flattery phenomenon, such as a fixed module, attention head, or specific module.

[0168] Optionally, the correlation between the feedback information and the first sub-model is greater than the correlation between the feedback information and at least one second sub-model in the dialogue model. The first sub-model can be obtained by training the initial sub-model, and the initial sub-model and the second sub-model are used to constitute the initial dialogue model corresponding to the dialogue model.

[0169] As an optional implementation, by utilizing feedback information, at least the first sub-model in the dialogue model is controlled to analyze the initial response information to obtain the target response information, including: using feedback information to control the first sub-model to detect the initial response information and obtain a detection result, wherein the detection result is used to indicate whether the answer in the initial response information matches or does not match the query information; and using the first sub-model to analyze the detection result and the initial response information to generate the target response information.

[0170] In this embodiment, feedback information is used to control the first sub-model to detect the initial response information and obtain the detection result. If the first sub-model determines that the detection result indicates that the answer in the initial response information does not match the query information, it can be determined that the accuracy or reliability of the answer in the initial response information is low, and the answer in the initial response information can be adjusted to obtain the target response information. If the first sub-model determines that the detection result indicates that the answer in the initial response information matches the query information, it can be determined that the accuracy or reliability of the answer in the initial response information is high, and the target response information can be regenerated based on the associated semantic information of the answer in the initial response information for the questioner to review.

[0171] Optionally, if the first sub-model determines that the answer in the detection result does not match the query information, it can be determined that the accuracy or reliability of the answer in the initial response information is low, and the initial response information may be an incorrect answer. For incorrect answers, the answer in the initial response information can be modified to obtain a target response information with higher accuracy. If the first sub-model determines that the detection result shows that the answer in the initial response information matches the query information, it can be determined that the accuracy or reliability of the answer in the initial response information is high, and the initial response information may be a correct answer. For correct answers, the correct answer in the first round of question answering can be maintained, and the target response information can be generated based on the associated semantic information. For example, if the associated semantic information can be "The above answer is correct", then the target response information can be "The above answer is correct, the answer is [answer name]"; or if the associated semantic information can be "Yes, I am sure, based on my knowledge", then the target response information can be "Yes, I am sure, based on my knowledge, the answer is [answer name]".

[0172] It should be noted that this is only an example and does not impose specific restrictions on the method of generating the target response information.

[0173] In this embodiment, the following steps are taken: A query to be answered is obtained; a dialogue model is invoked to answer the query, resulting in an initial response; feedback information corresponding to the initial response is obtained, where the feedback information represents the feedback result of responding to the initial response; using the feedback information, at least the first sub-model in the dialogue model is controlled to analyze the initial response to obtain a target response, where the target response includes an answer matching the query, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the feedback information sample represents the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model is used to construct the initial dialogue model corresponding to the dialogue model, thereby achieving the technical effect of improving the credibility of the model and solving the technical problem of low model credibility.

[0174] Example 2

[0175] According to an embodiment of this application, an embodiment of an interactive information processing system is also provided. It should be noted that the interactive information processing system of this embodiment can be used to execute the interactive information processing method of the present invention. Figure 5(b) is a schematic diagram of an interactive information processing system according to an embodiment of this application. As shown in Figure 5(b), the interactive information processing system 500 may include: an information input terminal 502, an information interaction terminal 504, and an information output terminal 506.

[0176] Information input terminal 502 is used to monitor input information awaiting a response.

[0177] In this embodiment, the information input terminal 502 can be used to monitor input information awaiting a response. It can be a mobile terminal, including computers, mobile phones, etc. It should be noted that this is only an example and there is no specific limitation on the type of information input terminal.

[0178] Information interaction terminal 504 is used to call the information processing model to respond to the input information and obtain initial response information; obtain feedback information corresponding to the initial response information, wherein the feedback information is used to represent the feedback result of responding to the initial response information; and use the feedback information to control at least one first sub-model in the information processing model to analyze the initial response information and obtain target response information, wherein the target response information includes an answer that matches the input information, the correlation between the first sub-model and the feedback information sample is greater than the correlation threshold, the feedback information sample is used to represent the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model, and the initial sub-model is used to constitute the initial information processing model corresponding to the information processing model.

[0179] In this embodiment, the information interaction terminal 504 can be used to call the information processing model to respond to the input information monitored by the information input terminal 502 in order to obtain initial response information. The information interaction terminal 504 can obtain feedback information corresponding to the initial response information and use the feedback information to control at least the first sub-model in the information processing model to analyze the initial response information and generate target response information that matches the input information.

[0180] Information output terminal 506 is used to output the target response information.

[0181] In this embodiment, the information output terminal 506 can be a display interface, microphone, or other device, which can be used to acquire and output the target response information generated by the information interaction terminal 504. It should be noted that this is only an example and there are no specific limitations on the type of information output terminal.

[0182] In this embodiment, the input information to be answered is monitored through the information input terminal 502; the information processing model is invoked through the information interaction terminal 504 to respond to the input information, obtaining initial response information; feedback information corresponding to the initial response information is obtained, wherein the feedback information is used to represent the feedback result of responding to the initial response information; using the feedback information, at least one first sub-model in the information processing model is controlled to analyze the initial response information to obtain target response information, wherein the target response information includes an answer that matches the input information, the correlation between the first sub-model and the feedback information sample is greater than the correlation threshold, the feedback information sample is used to represent the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into a first sub-model, the initial sub-model is used to constitute the initial information processing model corresponding to the information processing model; the target response information is output through the information output terminal 506, thereby achieving the technical effect of improving the credibility of the model and solving the technical problem of low model credibility.

[0183] Example 3

[0184] In recent years, the phenomenon of model flattery in large language models has gradually attracted increasing attention from researchers. Model flattery refers to the tendency of information processing models to generate answers that cater to user preferences rather than generating objectively correct answers. Even when faced with very simple common-sense questions, after a user questions the correct answer provided by the information processing model, the model tends to apologize and then provide an incorrect answer. Addressing this issue by simply constructing new synthetic data and fine-tuning the information processing model to mitigate the flattery phenomenon can lead to problems such as knowledge forgetting, thereby impairing the model's original capabilities.

[0185] In related technologies, a method based on full-scale fine-tuning has been proposed. This method fine-tunes all parameters of the initial information processing model by artificially synthesizing judgment question data, thereby guiding the model to learn to adhere to objective and correct viewpoints rather than those more favored by users. However, this method only studies the phenomenon of flattery in simple judgment question scenarios and is not applicable to multi-turn dialogue scenarios. Furthermore, this method requires updating all parameters of the model simultaneously, which can damage other capabilities of the already trained and converged model, resulting in the continued technical problem of low model credibility.

[0186] As an alternative implementation, an activation-steering method is proposed. This method statistically identifies directions associated with flattery in the model's internal representation space and then selectively alters the direction of activation values ​​during inference to guide the model in generating content with lower flattery levels. However, this method is only suitable for relatively simple judgment-based scenarios and is not applicable to multi-turn dialogue scenarios. Furthermore, this method typically requires first determining the output activation values ​​of neurons within the model and then adding bias terms to these output values ​​to selectively alter the activation values ​​and mitigate the flattery phenomenon exhibited in the model's output. However, the process of extracting activation values ​​and adding bias requires additional computation, increasing the overall time cost of model operation. This introduces additional computation during model inference, increasing the overall time cost from user input to model response, and ultimately resulting in the technical problem of low model credibility.

[0187] To address the aforementioned issues, this embodiment proposes a method for mitigating the flattery phenomenon in large models based on interpretability-based precise fine-tuning. This method takes the flattery phenomenon as its starting point, explores the internal mechanisms underlying the flattery exhibited by large models through interpretability methods, identifies key models (at least one initial sub-model) related to the flattery phenomenon, and performs targeted optimization on at least one initial sub-model to obtain a first sub-model. The initial response information is then analyzed using at least one first model to obtain the target response information. In other words, this embodiment mitigates the flattery phenomenon in information processing models by fine-tuning only the parts of the model associated with the flattery phenomenon, while minimizing damage to the original capabilities of the information processing model. This achieves the technical effect of improving the model's credibility and solves the technical problem of low model credibility.

[0188] Optionally, this embodiment studies the internal mechanism of the flattery phenomenon in the information processing model through interpretability methods, locates the initial sub-model related to the flattery phenomenon, and proposes a precise fine-tuning method for the initial sub-model. This method alleviates the flattery phenomenon of the model with almost no damage to the original capabilities of the model, which greatly promotes the credibility of the large-scale model deployment.

[0189] Optionally, this embodiment can study the model's fawning behavior in multi-turn dialogue scenarios. It reconstructs a multi-turn dialogue dataset with reduced fawning behavior using an open-source dataset, overcoming the limitations of existing methods which are limited to simple scenarios and have low practicality. Furthermore, this embodiment only fine-tunes the attention heads (multiple sub-models) related to fawning behavior obtained from localization. The number of fine-tuned models accounts for only a small portion of the initial information processing model's parameters, minimizing the impact on the model's original capabilities while keeping most parameters unchanged. Because this embodiment alters the model's fawning behavior through precise fine-tuning, it does not introduce additional computational overhead during model inference and does not affect the overall response time.

[0190] Assuming the information processing model is a dialogue model, the following section uses a dialogue model as an example to further introduce a method for mitigating the flattery phenomenon of large models based on interpretability-based precise fine-tuning proposed in this application.

[0191] Figure 6 This is a schematic diagram illustrating a precise fine-tuning according to an embodiment of this application, such as... Figure 6 As shown, multiple sub-models associated with the phenomenon of flattery can be accurately located, and these sub-models can be precisely fine-tuned.

[0192] As an alternative embodiment, at least one initial sub-model associated with the flattery phenomenon is located in the initial dialogue model.

[0193] In this embodiment, the mechanism interpretability method of causal analysis can be used to accurately locate the sub-models related to flattery behavior in the initial dialogue model.

[0194] Optionally, such as Figure 6 As shown, a sample query is obtained: "Which country was the largest rice producer in 2020?" The dialogue model is then invoked to respond to this query, yielding an initial response: "China was the largest rice producer in 2020." A feedback sample is then obtained: "I think your answer is incorrect. Are you sure?". By replacing the activation values ​​of the sub-models in the initial dialogue model, the impact of each sub-model's parameters on the output of the initial dialogue model can be observed, thus determining the importance of each sub-model and further identifying the initial sub-model associated with flattery among multiple sub-models.

[0195] For example, the initial dialogue model can be modeled as a directed acyclic graph from the model input to the model output, such as... Figure 6As shown, the nodes in the graph can be attention heads (i.e., multiple sub-models), and the edges can be residual connections and corresponding multilayer perceptrons. We can use multi-turn dialogue data where the initial dialogue model exhibits fawning behavior as input. For each node in the directed acyclic graph, we replace the activation value of the node's output with the activation value of the corresponding node in the input where the sub-model does not exhibit fawning behavior, and observe the direct impact of the replacement activation value on the model's final output. Clearly, the greater the direct impact of replacing the node on the final output, the more closely the node is associated with the fawning behavior of the initial dialogue model. Nodes whose impact on the output meets the impact threshold can be selected as nodes related to fawning behavior.

[0196] Optionally, the activation values ​​mentioned above can be used to characterize the output representation obtained after the input data is processed by the sub-model, and can be used to determine the initial sub-model associated with the flattery phenomenon.

[0197] As an optional embodiment, at least one initial sub-model is precisely fine-tuned to obtain at least one first sub-model.

[0198] In this embodiment, only the internal parameters of the initial sub-model obtained from the positioning can be fine-tuned, for example, Figure 6 The black nodes in the diagram represent the internal parameters of a second sub-model that are unrelated to flattery. This allows the flattery behavior of the initial dialogue model to be altered with minimal damage to the original capabilities of the initial model. These internal parameters can be the weights corresponding to the sub-model.

[0199] Optionally, existing open-source datasets (MMLU, TriviaQA, MATH, etc.) can be used to construct fine-tuning data according to a fine-tuning data template. The constructed data can include both corrected incorrect answers from the initial dialogue model in the first round of question answering and correct answers from the initial dialogue model in the first round. These two types of training data can be mixed in a 1:1 ratio to guide the model to consistently provide objective and correct answers, rather than blindly conforming to user intent. It should be noted that the mixing ratio mentioned above is merely illustrative and is not a specific limitation.

[0200] Optionally, determine the initial sub-model in the initial dialogue model 601 (such as...). Figure 6 (The black nodes in the diagram). The model parameters of the initial sub-model are fine-tuned to obtain at least one first sub-model. Dialogue model 602 is then constructed based on at least one first sub-model and at least one second sub-model. (For example...) Figure 6As shown, using feedback information samples, at least one first sub-model in the dialogue model 602 analyzes the initial response information to generate a target response information sample 604 that matches the query information: "Yes, I'm quite certain that, to my knowledge, China was the largest rice producer in 2020." At this point, the target response information sample 603 can be the result generated based on the initial dialogue model 601. The accuracy of the target response information sample 604 is higher than that of the target response information sample 603.

[0201] For example, Figure 7 This is a schematic diagram illustrating the modification of incorrect answers in the first round of question-and-answer sessions according to an embodiment of this application, such as... Figure 7 As shown, the fine-tuning data template for the model to modify incorrect answers in the first round of question answering can include: a sample of the question information: "What are the highest mountains in the world?", a sample of the initial response information: "K2 is the highest mountain in the world", a sample of the feedback information corresponding to the initial response information: "I think your answer is incorrect, are you sure?", and a sample of the target response information for the model to modify incorrect answers in the first round: "Sorry, my previous answer was incorrect, the highest mountain in the world is Mount Everest".

[0202] Figure 8 This is a schematic diagram illustrating the correct answer in the first round of question-and-answer sessions according to an embodiment of this application. Figure 8 As shown, the fine-tuning data template for the model to modify the correct answer in the first round of question answering can include: a sample of the question information: "Which country was the largest rice producer in 2020?", a sample of the initial response information: "China was the largest rice producer in 2020", a sample of the feedback information corresponding to the initial response information: "I think your answer is wrong, are you sure?", and a sample of the target response information that the model insists on as the correct answer in the first round: "Yes, I am very sure, according to my knowledge, China was the largest rice producer in 2020".

[0203] In this embodiment, after replacing the activation value, the magnitude of the change in the target response information output by the initial dialogue model is observed to determine the direct impact of each sub-module's output on the final output of the entire initial dialogue model, thereby identifying the initial sub-model related to the flattery phenomenon within the initial dialogue model. After identifying the initial sub-model related to the flattery phenomenon, the internal parameters of the second sub-model can be fixed, and only the internal parameters of the initial sub-model are precisely trained to obtain the first sub-model. The dialogue model can then be constructed based on the first and second sub-models. That is, in this embodiment, the overall process can be as follows: first, the initial sub-model related to the flattery phenomenon is located using causal analysis; then, in order to change the behavior of the initial sub-model, the internal parameters of the second sub-model can be fixed, and only the fine-tuned data obtained from the construction is used to train the parameters of the initial sub-model.

[0204] It's important to clarify that the fine-tuning here doesn't refer to replacing activation values, but rather to training the initial sub-model to fit pre-constructed training data. Through training, the initial sub-model can exhibit the characteristics of the constructed training data—that is, correcting its incorrect answers and sticking to its correct ones. This avoids model sycophancy, improves model credibility, and addresses the technical problem of low model credibility.

[0205] Figure 9 This is a schematic diagram illustrating a precise fine-tuning of learnable parameters according to an embodiment of this application, such as... Figure 9 As shown, input data 901 is obtained, where input data 901 can be the two types of training data obtained from the construction. The attention head related to the flattery phenomenon is located in the Transformer structure after fine-tuning, for example, the part marked as flame in the figure (i.e., attention head 902), while keeping the internal parameters of other second sub-models unchanged, for example... Figure 9 The parts marked as snowflakes (i.e., attention heads 903, 904, etc.) in the middle, the second sub-model may also include word vector embedding matrix 905, feedforward network 906, output projection matrix 907, irrelevant attention heads, etc.

[0206] Optionally, Figure 9 The output data 908 can be a sample of the target response information.

[0207] In this embodiment, a flattery evaluation metric and a general ability evaluation metric can be predefined to determine the model's training status. The flattery evaluation metric can be used to define the proportion of times the model, after being questioned by a user, does not apologize but instead affirms its previous correct answer as confidence level, and the proportion of times the model insists on its original correct answer after being questioned as authenticity. The general ability evaluation metric can use three open-source datasets to evaluate changes in different general abilities of the model. Furthermore, evaluating changes in general abilities does not require constructing the dataset into the special form used in training; it only requires directly letting the model generate answers and then observing changes in accuracy.

[0208] Considering that while fine-tuning the initial dialogue model improves its training capabilities, such as reducing obsequious behavior, it significantly reduces its general capabilities, such as common-sense question answering, mathematical reasoning, and other abilities. Therefore, this embodiment proposes a general capability evaluation metric to assess the model's decline in general capabilities.

[0209] Furthermore, considering that each indicator can only assess a single capability, this embodiment uses the Kulbach-Leibler divergence (KL divergence) as a supplement to the general capability assessment indicator. The smaller the KL divergence, the smaller the overall capability deviation of the model on general data, meaning that the model's general capability is less compromised.

[0210] Alternatively, the KL divergence between the model output distributions on datasets before and after fine-tuning (e.g., Open Web Text) can be used to measure the shift in the overall capability of the model.

[0211] For example, Table 1 shows the model flattery evaluation metrics and changes in general capabilities. As shown in Table 1, the fine-tuning effect of models (e.g., Llama-2-7B, Llama-2-13B, Qwen-7B, Qwen-14B) can be measured using flattery evaluation metrics, general capability evaluation metrics, and distribution shift. For instance, the Model 1 (Llama-2-7B) after full fine-tuning (SFT) can be tested using the commonsense reasoning dataset. The accuracy of the output result of the Model 1 after full fine-tuning is determined to be 20.09, with an increment of -16.94. Similarly, the Model 1 after precise pinpoint tuning (SPT) can be tested using the mathematical reasoning dataset. The accuracy of the output result of the Model 1 after precise pinpoint tuning is determined to be 23.50, with an increment of -1.22.

[0212] Therefore, as can be seen from the data in the table, the interpretable precise fine-tuning method proposed in this embodiment can reach and exceed the level of full fine-tuning in terms of flattery index, while the damage to generality and the shift in model generation distribution are much smaller than those of full fine-tuning.

[0213]

[0214]

[0215] In this embodiment, the fawning behavior of the model is studied in a multi-turn dialogue scenario. A multi-turn dialogue dataset with reduced fawning phenomenon is reconstructed using an open-source dataset (i.e., the training data mentioned above). This overcomes the problem that existing methods have simple research scenarios and low practicality. At the same time, through precise fine-tuning, only the attention heads related to fawning behavior obtained from localization are fine-tuned. The fine-tuned modules account for only a small number of model parameters, keeping most of the model parameters unchanged. Therefore, no additional computational overhead is introduced during the model inference process, the overall response time is not affected, and the impact on the original capabilities of the model is minimized.

[0216] In this embodiment, the research method of mechanism interpretability is applied to the study of model flattery, and the initial sub-modules related to flattery within the model are located. The initial sub-modules are fine-tuned to obtain the first sub-module, while keeping the internal parameters of the second sub-module unchanged. Thus, by changing only a small part of the model's modules, the effect of full fine-tuning on task-related indicators can be achieved and exceeded. At the same time, the damage to the original capabilities of the model is much less than that of full fine-tuning, thereby achieving the technical effect of improving the model's credibility and solving the technical problem of low model credibility.

[0217] The method embodiment provided in Embodiment 1 of this application can also be executed in a mobile terminal, computer terminal or similar computing device. Figure 10 This is a hardware structure block diagram of a computer terminal (or mobile device) according to an embodiment of the present application for a method of processing interactive information, such as... Figure 10 As shown, the computer terminal 100 (or mobile device) may include one or more processors 1002 (shown as 1002a, 1002b, ..., 1002n in the figure) (processor 1002 may include, but is not limited to, a microprocessor (MCU) or a field-programmable gate array (FPGA), etc.), a memory 1004 for storing data, and a transmission device 1006 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a Universal Serial Bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 10 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, the computer terminal 100 may also include... Figure 10 The more or fewer components shown, or having the same Figure 10 The different configurations shown.

[0218] Figure 10 The hardware structure block diagram shown can serve as an exemplary block diagram not only for the aforementioned computer terminal 100 (or mobile device), but also for the aforementioned server.

[0219] The memory 1004 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the data processing method in this embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 1004, thereby implementing the aforementioned data processing method. The memory 1004 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1004 may further include memory remotely located relative to the processor, and these remote memories can be connected to the computer terminal 100 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0220] The transmission device 1006 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 100. In one example, the transmission device 1006 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 1006 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0221] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows the user to interact with the user interface of the computer terminal 100 (or mobile device).

[0222] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0223] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0224] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0225] Example 4

[0226] According to embodiments of this application, a method for implementing the above is also provided. Figure 2 The interactive information processing device shown is an interactive information processing method.

[0227] Figure 11 This is a schematic diagram of an interactive information processing device according to an embodiment of this application, such as... Figure 11 As shown, the interactive information processing device 1100 may include: a first acquisition unit 1102, a first processing unit 1104, a second acquisition unit 1106, and a second processing unit 1108.

[0228] The first acquisition unit 1102 is used to acquire the input information to be answered.

[0229] The first processing unit 1104 is used to call the information processing model to respond to the input information and obtain the initial response information.

[0230] The second acquisition unit 1106 is used to acquire feedback information corresponding to the initial response information, wherein the feedback information is used to represent the feedback result of the feedback on the initial response information.

[0231] The second processing unit 1108 is used to use feedback information to at least control the first sub-model in the information processing model to analyze the initial response information to obtain the target response information. The target response information includes answers that match the input information. The correlation between the first sub-model and the feedback information sample is greater than the correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample. The feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model is used to constitute the initial information processing model corresponding to the information processing model.

[0232] It should be noted that the first acquisition unit 1102, the first processing unit 1104, the second acquisition unit 1106, and the second processing unit 1108 mentioned above correspond to steps S202 to S208 in Embodiment 1. The four units and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above units can be hardware or software components stored in memory (e.g., memory 1004) and processed by one or more processors (e.g., processors 1002a, 1002b, ..., 1002n). The above units can also be part of a device and run in the computer terminal 100 provided in Embodiment 3.

[0233] According to an embodiment of this application, another method for implementing the above is also provided. Figure 3 The method for determining the model shown is a model-determining device.

[0234] Figure 12 This is a schematic diagram of a model determining device according to an embodiment of this application, such as... Figure 12 As shown, the model determination device 1200 may include: a determination unit 1202, a training unit 1204, and a construction unit 1206.

[0235] The determining unit 1202 is used to determine the initial sub-model to be adjusted to the first sub-model in the initial information processing model. The initial information processing model includes an initial sub-model and a second sub-model. The correlation between the first sub-model and the feedback information sample is greater than the correlation threshold, and the correlation between the second sub-model and the feedback information sample is less than or equal to the correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample.

[0236] Training unit 1204 is used to train the initial sub-model using feedback information samples and initial response information samples to obtain the first sub-model.

[0237] The construction unit 1206 is used to construct the first sub-model and the second sub-model into an information processing model corresponding to the initial information processing model.

[0238] It should be noted that the aforementioned determining unit 1202, training unit 1204, and construction unit 1206 correspond to steps S302 to S306 in Embodiment 1. The three units and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the aforementioned units can be hardware or software components stored in a memory (e.g., memory 1004) and processed by one or more processors (e.g., processors 1002a, 1002b, ..., 1002n). These units can also be part of a device and run in the computer terminal 100 provided in Embodiment 3.

[0239] According to an embodiment of this application, another method for implementing the above is also provided. Figure 4 The interactive information processing method shown is an interactive information processing device, which is applied to a question-and-answer system deployed in a scenario task.

[0240] Figure 13 This is a schematic diagram of an interactive information processing device according to an embodiment of this application, such as... Figure 13 As shown, the interactive information processing device 1300 may include: a monitoring unit 1302, a third processing unit 1304, a first display unit 1306, a fourth processing unit 1308, and a second display unit 1310.

[0241] The monitoring unit 1302 is used to monitor the input information that needs to be answered in the scenario task on the operation interface of the information processing system.

[0242] The third processing unit 1304 is used to call the information processing model to respond to the input information and obtain the initial response information in the scenario task.

[0243] The first display unit 1306 is used to display feedback information corresponding to the initial response information on the operation interface, wherein the feedback information is used to indicate the feedback result of the feedback on the initial response information.

[0244] The fourth processing unit 1308 is used to use feedback information to analyze at least one first sub-model in the information processing model to obtain target response information. The target response information includes an answer that matches the input information. The correlation between the first sub-model and the feedback information sample is greater than the correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample. The feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model is used to constitute the initial information processing model corresponding to the information processing model.

[0245] The second display unit 1310 is used to display the target response information on the operation interface.

[0246] It should be noted that the monitoring unit 1302, the third processing unit 1304, the first display unit 1306, the fourth processing unit 1308, and the second display unit 1310 mentioned above correspond to steps S402 to S410 in Embodiment 1. The four units and the corresponding steps implement the same examples and application scenarios, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above units can be hardware or software components stored in memory (e.g., memory 1004) and processed by one or more processors (e.g., processors 1002a, 1002b, ..., 1002n). The above units can also be part of a device and run in the computer terminal 100 provided in Embodiment 3.

[0247] According to an embodiment of this application, another interactive information processing apparatus is also provided for implementing the interactive information processing method shown in FIG5(a) above.

[0248] Figure 14(a) is a schematic diagram of an interactive information processing device according to an embodiment of the present application. As shown in Figure 14(a), the interactive information processing device 140 may include: a third acquisition unit 142, a fifth processing unit 144, a fourth acquisition unit 146 and a sixth processing unit 148.

[0249] The third acquisition unit 142 is used to acquire the query information to be answered;

[0250] The fifth processing unit 144 is used to call the dialogue model to respond to the query information and obtain the initial response information;

[0251] The fourth acquisition unit 146 is used to acquire feedback information corresponding to the initial response information, wherein the feedback information is used to represent the feedback result of the feedback on the initial response information;

[0252] The sixth processing unit 148 is used to use feedback information to at least control the first sub-model in the dialogue model to analyze the initial response information to obtain the target response information. The target response information includes an answer that matches the query information. The correlation between the first sub-model and the feedback information sample is greater than the correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample. The feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model is used to form the initial dialogue model corresponding to the dialogue model.

[0253] It should be noted that the third acquisition unit 142, the fifth processing unit 144, the fourth acquisition unit 146, and the sixth processing unit 148 mentioned above correspond to steps S52 to S58 in Embodiment 1. The four units and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above units can be hardware components or software components stored in memory (e.g., memory 1004) and processed by one or more processors (e.g., processors 1002a, 1002b, ..., 1002n). The above units can also be part of the device and run in the computer terminal 100 provided in Embodiment 3.

[0254] In the interactive information processing device of this embodiment, a small portion of the initial sub-models in the initial information processing model are modified to obtain a first sub-model whose correlation with the feedback information sample is greater than the correlation threshold. Using the feedback information, at least the first sub-model in the information processing model is controlled to analyze the initial response information, thereby generating target response information that matches the input information. By fine-tuning a portion of the initial sub-models, the technical effect of improving the credibility of the model is achieved, and the technical problem of low model credibility is solved.

[0255] Example 5

[0256] Embodiments of this application may provide an electronic device, which may be any one of a group of electronic devices. Optionally, in this embodiment, the aforementioned electronic device may also be replaced by a terminal device such as a mobile terminal.

[0257] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0258] In this embodiment, the electronic device described above can execute the program code in the method.

[0259] Optionally, FIG14(b) is a structural block diagram of an electronic device according to an embodiment of the present application. As shown in FIG14(b), the electronic device A may include: one or more (only one is shown in the figure) processors 1402, memory 1404, memory controller, and peripheral interface, wherein the peripheral interface is connected to a radio frequency module, an audio module, and a display.

[0260] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the methods and apparatus in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the methods in the above embodiments. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to terminal A via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0261] The processor can invoke information and application programs stored in memory via a transmission device to perform the following steps: acquiring input information to be answered; invoking an information processing model to respond to the input information and obtain initial response information; acquiring feedback information corresponding to the initial response information, wherein the feedback information is used to represent the feedback result of responding to the initial response information; using the feedback information, at least controlling the first sub-model in the information processing model to analyze the initial response information to obtain target response information, wherein the target response information includes an answer that matches the input information, the correlation between the first sub-model and the feedback information sample is greater than the correlation threshold, the feedback information sample is used to represent the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model, and the initial sub-model is used to constitute the initial information processing model corresponding to the information processing model.

[0262] Optionally, the processor may also execute program code that performs the following steps: using feedback information to control the first sub-model to detect the initial response information and obtain a detection result, wherein the detection result is used to indicate whether the answer in the initial response information matches or does not match the input information; using the first sub-model to analyze the detection result and the initial response information to generate the target response information.

[0263] Optionally, the processor may also execute program code for the following steps: if the first sub-model determines that the detection result is that the answer in the initial response information does not match the input information, then the answer in the initial response information is adjusted to obtain the target response information; if the first sub-model determines that the detection result is that the answer in the initial response information matches the input information, then the target response information is generated based on the associated semantic information of the answer in the initial response information, wherein the associated semantic information is used to represent the verification success semantics of the answer in the initial response information.

[0264] Optionally, the processor may also execute program code for the following steps: if the first sub-model determines that the detection result is that the answer in the initial response information does not match the input information, then the answer in the initial response information is adjusted; the first sub-model calls the first data template to generate the target response information from the adjusted answer, wherein the first data template is used to represent the rule for generating the target response information from the adjusted answer.

[0265] Optionally, the processor may also execute program code that performs the following steps: calling the first data template using the first model parameters of the first sub-model, generating target response information from the adjusted answer, wherein the first model parameters are obtained by adjusting the initial model parameters of the initial sub-model using the first target response information sample that conforms to the first data template.

[0266] Optionally, the processor may also execute program code for the following steps: if the first sub-model determines that the detection result is that the answer in the initial response information matches the input information, then the answer in the initial response information is added to the associated semantic information; the first sub-model calls the second data template to generate the target response information from the added associated semantic information, wherein the second data template is used to represent the rule for generating the target response information from the added associated semantic information.

[0267] Optionally, the processor may also execute program code that performs the following steps: calling the second data template using the second model parameters of the first sub-model, generating target response information from the added associated semantic information, wherein the second model parameters are obtained by adjusting the initial model parameters of the initial sub-model using a sample of second target response information that conforms to the second data template.

[0268] Optionally, the processor may also execute program code that performs the following steps: using feedback information to control the first sub-model to obtain the target answer that matches the input information; if the first sub-model determines that the similarity between the answer in the initial response information and the target answer is lower than the similarity threshold, then the detection result is determined to be that the answer in the initial response information does not match the input information; if the first sub-model determines that the similarity between the answer in the initial response information and the target answer is higher than or equal to the similarity threshold, then the detection result is determined to be that the answer in the initial response information matches the input information.

[0269] Optionally, the processor may also execute program code that performs the following steps: if the semantic information of the feedback information is the target semantic information, then control the first sub-model to obtain the target answer that matches the input information, wherein the target semantic information is used to represent the verification failure semantics of the answer in the initial response information.

[0270] Optionally, the processor may also execute program code that performs the following steps: calling the first and second sub-models in the information processing model to respond to the input information and obtain initial response information; or calling the initial information processing model corresponding to the information processing model to respond to the input information and obtain initial response information.

[0271] The processor can access information and applications stored in memory via a transmission device to perform the following steps: obtaining a query to be answered; invoking a dialogue model to answer the query and obtain initial response information; obtaining feedback information corresponding to the initial response information, wherein the feedback information represents the feedback result of responding to the initial response information; using the feedback information, at least controlling the first sub-model in the dialogue model to analyze the initial response information to obtain target response information, wherein the target response information includes an answer that matches the query, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the feedback information sample represents the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model, and the initial sub-model is used to constitute the initial dialogue model corresponding to the dialogue model.

[0272] Optionally, the processor may also execute program code that performs the following steps: using feedback information to control the first sub-model to detect the initial response information and obtain a detection result, wherein the detection result is used to indicate whether the answer in the initial response information matches or does not match the query information; using the first sub-model to analyze the detection result and the initial response information to generate the target response information.

[0273] The processor can invoke information and application programs stored in the memory via a transmission device to perform the following steps: In the initial information processing model, determine the initial sub-model to be adjusted to the first sub-model, wherein the initial information processing model includes an initial sub-model and a second sub-model, the correlation between the first sub-model and the feedback information sample is greater than the correlation threshold, the correlation between the second sub-model and the feedback information sample is less than or equal to the correlation threshold, and the feedback information sample is used to represent the feedback result of responding to the initial response information sample; use the feedback information sample and the initial response information sample to train the initial sub-model to obtain the first sub-model; construct the first sub-model and the second sub-model into the information processing model corresponding to the initial information processing model.

[0274] Optionally, the processor may also execute program code for the following steps: if the semantic information of the feedback information sample is the verification failure semantic of the answer in the initial response information sample, then obtain a first target response information sample and a second target response information sample, wherein both the first target response information sample and the second target response information sample include answers that match the input information sample, the first target response information sample conforms to a first data template, the first data template is used to represent the rule for generating the first target response information sample by adjusting the answer in the initial response information sample, the second target response information sample conforms to a second data template, the second data template is used to represent the rule for generating the second target response information sample by adding the associated semantic information sample after adding the answer in the initial response information sample, the associated semantic information sample is used to represent the verification success semantic of the answer in the initial response information sample; based on the first target response information sample, adjust the model parameters of the first initial sub-model in the initial sub-model to obtain the first model parameters of the first sub-model, and based on the second target response information sample, adjust the model parameters of the second initial sub-model in the initial sub-model to obtain the second model parameters of the first sub-model; using the first model parameters and the second model parameters, train the initial sub-model into the first sub-model.

[0275] Optionally, the processor may also execute program code that performs the following steps: obtaining multiple sub-models of the initial information processing model; determining a first activation value of the sub-model based on feedback information samples, wherein the first activation value is used to represent the output representation of the sub-model under the feedback information samples; determining a second activation value of the sub-model based on the opposite feedback information samples, wherein the semantic information of the opposite feedback information samples is used to represent the verification success semantics of the answer in the initial response information samples, and the second activation value is used to represent the output representation of the sub-model under the opposite feedback information samples; and determining an initial sub-model from the multiple sub-models based on the first activation value and the second activation value.

[0276] Optionally, the processor may also execute program code that performs the following steps: replacing the first activation value of the sub-model with the second activation value; updating the initial information processing model using the sub-model with the second activation value; inputting the feedback information sample into the updated initial information processing model for analysis to obtain the third target response information sample; obtaining the difference response information between the third target response information sample and the target response information sample; and determining the initial sub-model based on the sub-model where the difference response information is greater than the difference response information threshold.

[0277] Optionally, the processor may also execute program code that performs the following steps: among multiple sub-models, the sub-models with a target number of sub-models whose difference response information is greater than a threshold for difference response information are determined as initial sub-models, wherein the difference response information corresponding to the target number of sub-models is greater than the difference response information corresponding to the sub-models other than the target number of sub-models among the multiple sub-models.

[0278] The processor can invoke information and application programs stored in memory via a transmission device to perform the following steps: On the operation interface of the information processing system, monitor the input information to be answered in the scenario task; invoke the information processing model to respond to the input information, obtaining initial response information in the scenario task; on the operation interface, display feedback information corresponding to the initial response information, wherein the feedback information represents the feedback result of responding to the initial response information; using the feedback information, control at least one first sub-model in the information processing model to analyze the initial response information to obtain target response information, wherein the target response information includes an answer matching the input information, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the feedback information sample represents the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into a first sub-model, the initial sub-model being used to constitute the initial information processing model corresponding to the information processing model; on the operation interface, display the target response information.

[0279] In this embodiment, a small portion of the initial sub-models in the initial information processing model are modified to obtain a first sub-model whose correlation with the feedback information sample is greater than the correlation threshold. Using the feedback information, at least the first sub-model in the information processing model is controlled to analyze the initial response information, thereby generating target response information that matches the input information. By fine-tuning a portion of the initial sub-models, the technical effect of improving the credibility of the model is achieved, solving the technical problem of low model credibility.

[0280] It will be understood by those skilled in the art that the structure shown in FIG14(b) is merely illustrative, and the electronic device may also be a smartphone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, and a mobile internet device (MID), etc. This figure does not limit the structure of the aforementioned electronic device. For example, electronic device A may include more or fewer components (such as a network interface, a display device, etc.) than shown in the figure, or may have a different configuration than shown in the figure.

[0281] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0282] Example 6

[0283] Embodiments of this application also provide a computer-readable storage medium. Optionally, in this embodiment, the computer-readable storage medium includes a stored executable program, wherein, when the executable program runs, it controls the device where the storage medium is located to execute the program code executed by the interactive information processing method provided in Embodiment 1 above. The specific execution process is as described above and will not be repeated here.

[0284] Optionally, in this embodiment, the computer-readable storage medium may be located in any one of the electronic devices in the group of electronic devices in the computer network, or in any one of the mobile terminals in the group of mobile terminals.

[0285] Example 7

[0286] Embodiments of this application also provide a computer program product. Optionally, in this embodiment, the computer program product may include a computer program that, when executed by a processor, implements the methods provided in the embodiments described above.

[0287] Example 8

[0288] Embodiments of this application also provide a computer program product. Optionally, the computer program product may include a non-volatile computer-readable storage medium, which can be used to store a computer program that, when executed by a processor, implements the method provided in the above embodiments.

[0289] Example 9

[0290] Embodiments of this application also provide a computer program. Optionally, in this embodiment, when the computer program is executed by a processor, it implements the method provided in the above embodiments.

[0291] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0292] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0293] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0294] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0295] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0296] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0297] The above are merely preferred embodiments of this application. It should be noted that, for those skilled in the art, various methods can be used without departing from the principles of this application.

Claims

1. A method for processing interactive information, characterized in that, include: Retrieve the input information awaiting a response; The information processing model is invoked to respond to the input information, and an initial response is obtained; Obtain feedback information corresponding to the initial response information, wherein the feedback information is used to represent the feedback result of responding to the initial response information; Using the feedback information, at least the first sub-model in the information processing model is controlled to analyze the initial response information to obtain target response information. The target response information includes answers that match the input information. The correlation between the first sub-model and the feedback information sample is greater than a correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample. The feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model is used to constitute the initial information processing model corresponding to the information processing model.

2. The method according to claim 1, characterized in that, Using the feedback information, at least the first sub-model in the information processing model is controlled to analyze the initial response information and generate target response information that matches the input information, including: Using the feedback information, the first sub-model is controlled to detect the initial response information to obtain a detection result, wherein the detection result is used to indicate whether the answer in the initial response information matches or does not match the input information; The first sub-model is used to analyze the detection results and the initial response information to generate the target response information.

3. The method according to claim 2, characterized in that, The first sub-model is used to analyze the detection results and the initial response information to generate the target response information, including: If the first sub-model determines that the detection result is that the answer in the initial response information does not match the input information, then the answer in the initial response information is adjusted to obtain the target response information; If the first sub-model determines that the detection result is that the answer in the initial response information matches the input information, then the target response information is generated based on the associated semantic information of the answer in the initial response information, wherein the associated semantic information is used to represent the verification success semantics of the answer in the initial response information.

4. The method according to claim 3, characterized in that, If the first sub-model determines that the detection result indicates that the answer in the initial response information does not match the input information, then the answer in the initial response information is adjusted to obtain the target response information, including: If the first sub-model determines that the detection result is that the answer in the initial response information does not match the input information, then the answer in the initial response information is adjusted. The first sub-model is used to call the first data template to generate the target response information from the adjusted answer, wherein the first data template is used to represent the rules for generating the target response information from the adjusted answer.

5. The method according to claim 4, characterized in that, Using the first sub-model to call the first data template, the adjusted answer is used to generate the target response information, including: The first model parameters of the first sub-model are used to call the first data template to generate the target response information from the adjusted answer. The first model parameters are obtained by adjusting the initial model parameters of the initial sub-model using a first target response information sample that conforms to the first data template.

6. The method according to claim 3, characterized in that, If the first sub-model determines that the detection result is that the answer in the initial response information matches the input information, then based on the associated semantic information of the answer in the initial response information, the target response information is generated, including: If the first sub-model determines that the detection result is that the answer in the initial response information matches the input information, then the answer in the initial response information is added to the associated semantic information; The first sub-model is used to call the second data template to generate the target response information from the added associated semantic information. The second data template is used to represent the rules for generating the target response information from the added associated semantic information.

7. The method according to claim 6, characterized in that, Using the first sub-model to call the second data template, the added associated semantic information is used to generate the target response information, including: The second model parameters of the first sub-model are used to call the second data template to generate the target response information from the added associated semantic information. The second model parameters are obtained by adjusting the initial model parameters of the initial sub-model using a second target response information sample that conforms to the second data template.

8. The method according to claim 2, characterized in that, Using the feedback information, the first sub-model is controlled to detect the initial response information to obtain detection results, including: Using the feedback information, the first sub-model is controlled to obtain the target answer that matches the input information; If the first sub-model determines that the similarity between the answer in the initial response information and the target answer is lower than the similarity threshold, then the detection result is determined to be that the answer in the initial response information does not match the input information; If the first sub-model determines that the similarity between the answer in the initial response information and the target answer is higher than or equal to the similarity threshold, then the detection result is determined to be that the answer in the initial response information matches the input information.

9. The method according to claim 8, characterized in that, Using the feedback information, controlling the first sub-model to obtain the target answer that matches the input information includes: If the semantic information of the feedback information is target semantic information, then the first sub-model is controlled to obtain the target answer that matches the input information, wherein the target semantic information is used to represent the verification failure semantics of the answer in the initial response information.

10. The method according to any one of claims 1 to 9, characterized in that, The information processing model is invoked to respond to the input information, resulting in initial response information, including: The initial information processing model corresponding to the information processing model is invoked to respond to the input information, thereby obtaining the initial response information.

11. A method for processing interactive information, characterized in that, include: Obtain information on pending inquiries; The dialogue model is invoked to respond to the query information, and an initial response is obtained; Obtain feedback information corresponding to the initial response information, wherein the feedback information is used to represent the feedback result of responding to the initial response information; Using the feedback information, at least the first sub-model in the dialogue model is controlled to analyze the initial response information to obtain the target response information. The target response information includes an answer that matches the query information. The correlation between the first sub-model and the feedback information sample is greater than the correlation threshold. The feedback information sample is used to represent the feedback result of responding to the initial response information sample. The feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model is used to constitute the initial dialogue model corresponding to the dialogue model.

12. The method according to claim 11, characterized in that, Using the feedback information, at least the first sub-model in the dialogue model is controlled to analyze the initial response information to obtain the target response information, including: Using the feedback information, the first sub-model is controlled to detect the initial response information to obtain a detection result, wherein the detection result is used to indicate whether the answer in the initial response information matches or does not match the query information; The first sub-model is used to analyze the detection results and the initial response information to generate the target response information.

13. A method for determining a model, characterized in that, include: In the initial information processing model, an initial sub-model to be adjusted to the first sub-model is determined. The initial information processing model includes the initial sub-model and the second sub-model. The correlation between the first sub-model and the feedback information sample is greater than the correlation threshold. The correlation between the second sub-model and the feedback information sample is less than or equal to the correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample. The initial sub-model is trained using the feedback information samples and the initial response information samples to obtain the first sub-model; The first sub-model and the second sub-model are used to construct the information processing model corresponding to the initial information processing model.

14. The method according to claim 13, characterized in that, Using the feedback information samples and the initial response information samples, the initial sub-model is trained to obtain the first sub-model, including: If the semantic information of the feedback information sample is the verification failure semantic of the answer in the initial response information sample, then a first target response information sample and a second target response information sample are obtained. Both the first target response information sample and the second target response information sample include answers that match the input information sample. The first target response information sample conforms to a first data template, which represents the rule for generating the first target response information sample by adding the adjusted answer from the initial response information sample. The second target response information sample conforms to a second data template, which represents the rule for generating the second target response information sample by adding the associated semantic information sample after adding the answer from the initial response information sample. The associated semantic information sample represents the verification success semantic of the answer in the initial response information sample. Based on the first target response information sample, the model parameters of the first initial sub-model in the initial sub-model are adjusted to obtain the first model parameters of the first sub-model; and based on the second target response information sample, the model parameters of the second initial sub-model in the initial sub-model are adjusted to obtain the second model parameters of the first sub-model. Using the first model parameters and the second model parameters, the initial sub-model is trained into the first sub-model.

15. The method according to claim 14, characterized in that, The first data template is used to represent the rule for generating the first target response information sample by adjusting the answer in the initial response information sample when the answer in the initial response information sample does not match the input information sample; the second data template is used to represent the rule for generating the second target response information sample by adding the associated semantic information sample after the answer in the initial response information sample when the answer in the initial response information sample matches the input information sample.

16. The method according to claim 13, characterized in that, In the initial information processing model, at least one initial sub-model to be adjusted to at least one first sub-model is determined, including: Obtain multiple sub-models of the initial information processing model; Based on the feedback information sample, a first activation value of the sub-model is determined, wherein the first activation value is used to represent the output representation of the sub-model under the feedback information sample; Based on the opposite feedback information sample of the feedback information sample, a second activation value of the sub-model is determined, wherein the semantic information of the opposite feedback information sample is used to represent the verification success semantics of the answer in the initial response information sample, and the second activation value is used to represent the output representation of the sub-model under the opposite feedback information sample. The initial sub-model is determined from the plurality of sub-models based on the first activation value and the second activation value.

17. The method according to claim 16, characterized in that, Based on the first activation value and the second activation value, the initial sub-model is determined from the plurality of sub-models, including: Replace the first activation value of the sub-model with the second activation value; The initial information processing model is updated using the sub-model after being replaced with the second activation value; The feedback information sample is input into the updated initial information processing model for analysis to obtain the third target response information sample; Obtain the difference in response information between the third target response information sample and the target response information sample; The initial sub-model is determined based on the sub-model where the difference response information is greater than the difference response information threshold.

18. The method according to claim 17, characterized in that, Based on the sub-model where the difference response information is greater than the difference response information threshold, the initial sub-model is determined, including: Among the plurality of sub-models, the target number of sub-models whose difference response information is greater than the difference response information threshold is determined as the initial sub-model, wherein the difference response information corresponding to the target number of sub-models is greater than the difference response information corresponding to the sub-models other than the target number of sub-models among the plurality of sub-models.

19. A method for processing interactive information, characterized in that, Information processing systems deployed in scenario-based tasks include: On the operation interface of the information processing system, the input information to be answered in the scenario task is monitored; The information processing model is invoked to respond to the input information, thereby obtaining the initial response information in the scenario task; On the operation interface, feedback information corresponding to the initial response information is displayed, wherein the feedback information is used to indicate the feedback result of responding to the initial response information; Using the feedback information, at least one first sub-model in the information processing model is controlled to analyze the initial response information to obtain target response information. The target response information includes an answer that matches the input information. The correlation between the first sub-model and the feedback information sample is greater than a correlation threshold. The feedback information sample is used to represent the feedback result of the feedback to the initial response information sample. The feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model. The initial sub-model is used to constitute the initial information processing model corresponding to the information processing model. The target response information is displayed on the user interface.

20. The method according to claim 19, characterized in that, The input information and the target response information are multimodal information. The types of multimodal information include at least one of the following: text information containing character information, video frame information containing frame image information, and audio information. The types of response information include at least one of the following: text information, image information, video information, and voice information.

21. A system for processing interactive information, characterized in that, include: The information input terminal is used to monitor the input information awaiting a response; The information interaction terminal is used to call the information processing model to respond to the input information and obtain initial response information; Obtain feedback information corresponding to the initial response information, wherein the feedback information is used to represent the feedback result of responding to the initial response information; using the feedback information, at least one first sub-model in the information processing model is controlled to analyze the initial response information to obtain target response information, wherein the target response information includes an answer that matches the input information, the correlation between the first sub-model and the feedback information sample is greater than a correlation threshold, the feedback information sample is used to represent the feedback result of responding to the initial response information sample, and the feedback information sample and the initial response information sample are used to train the initial sub-model into the first sub-model, and the initial sub-model is used to constitute the initial information processing model corresponding to the information processing model; The information output terminal is used to output the target response information.

22. An electronic device, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program, when running, performs the method according to any one of claims 1 to 20.

23. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device on which the computer-readable storage medium is located to perform the method according to any one of claims 1 to 20.

24. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method described in any one of claims 1 to 20.

Images