Private dialogue data-oriented reply trust prediction method

Abstract

The embodiment of the invention discloses a reply trust prediction method for private dialogue data. A specific embodiment of the method comprises the following steps: receiving an initial model parameter; performing data enhancement processing on the local question-answer dialogue training sample set, and inputting the data to a correlation generation model to obtain a question-answer correlation value; inputting the enhanced question-answer dialogue training sample set into a privacy generation model to obtain a question-answer privacy value; performing parameter comparison learning updating on the initial model parameters to obtain updated local model parameters; sending the updated local model parameters to a server; determining the initial local question and answer privacy correlation generation model as a local question and answer privacy correlation generation model; generating a privacy correlation value set; and carrying out information flow interception control on the user side. According to the implementation mode, the security of the data and the relevance and privacy of the dialogue context can be improved, invalid interaction is reduced, loads of a user side and a server side are reduced, and the user experience feeling is improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology, and in particular to a method for predicting the credibility of responses to private conversation data. Background Art

[0002] Currently, conversation generation models are trained by learning from large amounts of conversation data to develop models that can understand language and generate coherent, relevant responses. However, ensuring the security of sample data during model training and preventing the leakage of private data is a growing concern. A common approach for assessing the privacy relevance of responses in question-and-answer conversations is to store each user's conversation sample data on the server. The server then trains an initial question-and-answer privacy generation model / initial question-and-answer relevance generation model based on the acquired conversation sample data. This trained question-and-answer privacy generation model / question-and-answer relevance generation model is then used to evaluate the privacy / relevance of the predicted conversation response set, obtaining a privacy evaluation value / relevance evaluation value set. The server then controls the interception of response messages based on the privacy evaluation value / relevance evaluation value set.

[0003] However, in practice, it has been found that when the above method is used to evaluate the privacy relevance of reply information in question-and-answer conversations, the following technical problems often occur: the conversation sample data of each user end is transmitted to the server end for unified model training, which is prone to privacy data leakage during the transmission process, requiring a large amount of transmission resources, resulting in low security of sample data, privacy data leakage, wasting a lot of transmission resources, and increasing the load on the server end; and only evaluating the reply information in the question-and-answer conversation based on a single indicator such as relevance or privacy, resulting in low accuracy of the reply information evaluation and reduced user experience.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure concept and therefore it may contain information that does not form the prior art that is already known in this country to a person of ordinary skill in the art. Summary of the Invention

[0005] The content of this disclosure is used to briefly introduce concepts that will be described in detail in the detailed description section below. The content of this disclosure is not intended to identify key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0006] Some embodiments of the present disclosure propose a reply trustworthiness prediction method for private conversation data to solve one or more of the technical problems mentioned in the above background technology section.

[0007] In a first aspect, some embodiments of the present disclosure provide a method for predicting the trustworthiness of responses to privacy-related conversation data, comprising: receiving initial model parameters sent by a server; performing the following first determination step based on the initial model parameters: performing data enhancement processing on a local question-and-answer conversation training sample set to obtain an enhanced question-and-answer conversation training sample set; inputting the enhanced question-and-answer conversation training sample set into a correlation generation model included in an initial local question-and-answer privacy correlation generation model corresponding to the initial model parameters to obtain a question-and-answer correlation value; inputting the enhanced question-and-answer conversation training sample set into a privacy generation model included in the initial local question-and-answer privacy correlation generation model to obtain a question-and-answer privacy value; and adjusting the initial model parameters of the initial local question-and-answer privacy correlation generation model according to the question-and-answer correlation value and the question-and-answer privacy value. The method comprises the steps of: performing parameter comparison learning and updating based on the received parameters to obtain updated local model parameters; sending the updated local model parameters to the above-mentioned server to perform parameter aggregation on the received updated local model parameters to obtain global model parameters, and sending the global model parameters to each user terminal; in response to not receiving the global model parameters sent by the above-mentioned server within a preset time, determining the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters as the trained local question-answer privacy relevance generation model; generating a privacy relevance value set for the obtained dialogue response information set to be predicted based on the above-mentioned trained local question-answer privacy relevance generation model; and performing information flow interception control on the user terminal corresponding to the above-mentioned dialogue response information set to be predicted based on the above-mentioned privacy relevance value set.

[0008] In a second aspect, some embodiments of the present disclosure provide a reply trust prediction device for privacy-sensitive conversation data, comprising: a receiving unit configured to receive initial model parameters sent by a server; an execution unit configured to execute the following first determination step based on the initial model parameters: performing data enhancement processing on a local question-and-answer conversation training sample set to obtain an enhanced question-and-answer conversation training sample set; inputting the enhanced question-and-answer conversation training sample set into a correlation generation model included in an initial local question-and-answer privacy correlation generation model corresponding to the initial model parameters to obtain a question-and-answer correlation value; inputting the enhanced question-and-answer conversation training sample set into a privacy generation model included in the initial local question-and-answer privacy correlation generation model to obtain a question-and-answer privacy value; and adjusting the initial model parameters of the initial local question-and-answer privacy correlation generation model according to the question-and-answer correlation value and the question-and-answer privacy value. The method comprises the steps of: performing parameter comparison learning and updating based on the received parameters to obtain updated local model parameters; sending the updated local model parameters to the above-mentioned server to perform parameter aggregation on the received updated local model parameters to obtain global model parameters, and sending the global model parameters to each user terminal; in response to not receiving the global model parameters sent by the above-mentioned server within a preset time, determining the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters as the trained local question-answer privacy relevance generation model; a generating unit is configured to generate a privacy relevance value set for the acquired dialogue response information set to be predicted based on the above-mentioned trained local question-answer privacy relevance generation model; and a controlling unit is configured to perform information flow interception control on the user terminal corresponding to the above-mentioned dialogue response information set to be predicted based on the above-mentioned privacy relevance value set.

[0009] In a third aspect, some embodiments of the present disclosure provide an electronic device comprising: one or more processors; a storage device on which one or more programs are stored, and when the one or more programs are executed by one or more processors, the one or more processors implement the method described in any implementation manner in the first aspect.

[0010] In a fourth aspect, some embodiments of the present disclosure provide a computer-readable medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method described in any implementation manner in the first aspect is implemented.

[0011] The above-mentioned embodiments of the present disclosure have the following beneficial effects: In some embodiments of the present disclosure, the reply trustworthiness prediction method for private conversation data trains a model using local user samples, thereby improving data security and protecting user privacy. Furthermore, by generating a privacy-related value set for the set of reply information to be predicted, the relevance and privacy of the conversation context can be improved, invalid interactions can be reduced, and the user experience can be improved. Furthermore, by only sending model parameters to the server, transmission resource waste and the load on the server and each user can be reduced. Specifically, the reasons for the low security of the relevant sample data, the risk of privacy data leakage, the waste of a large amount of transmission resources, the increased load on the server, and the low accuracy of reply information evaluation, which reduces the user experience, are: the transmission of conversation sample data from each user to the server for unified model training is prone to privacy data leakage during the transmission process, requiring a large amount of transmission resources, resulting in low sample data security, the risk of privacy data leakage, the waste of a large amount of transmission resources, and the increased load on the server. Furthermore, in question-and-answer conversations, only a single metric, such as relevance or privacy, is evaluated on the reply information, resulting in low accuracy of reply information evaluation and a reduced user experience. Based on this, some embodiments of the present disclosure may first receive initial model parameters sent by a server. The initial model parameters facilitate subsequent local model parameter updates on the user side. Next, based on the initial model parameters, the following first determination step is performed: First, data enhancement processing is performed on the local question-and-answer conversation training sample set to obtain an enhanced question-and-answer conversation training sample set. Processing the local question-and-answer conversation training sample set can improve the data quality of the local question-and-answer conversation training sample set. Second, the enhanced question-and-answer conversation training sample set is input into a correlation generation model included in an initial local question-and-answer privacy relevance generation model corresponding to the initial model parameters to obtain a question-and-answer relevance value. Using the correlation of the local question-and-answer conversation training sample set to train the model can improve the model training rate. Third, the enhanced question-and-answer conversation training sample set is input into a privacy generation model included in the initial local question-and-answer privacy relevance generation model to obtain a question-and-answer privacy value. Using the privacy of the local question-and-answer conversation training sample set to train the model can improve the model training rate. In the fourth step, based on the question-answer relevance and question-answer privacy values, the initial model parameters of the initial local question-answer privacy relevance generation model are updated through parameter comparison learning to obtain updated local model parameters. Here, integrating privacy and relevance training can improve the comprehensiveness and accuracy of the model. In the fifth step, the updated local model parameters are sent to the aforementioned server, which aggregates the received updated local model parameters to obtain global model parameters. The global model parameters are then sent to each user terminal.Here, model aggregation training is performed on the server side using only the local model parameters sent by each user. This can reduce transmission resource waste, increase the security of sample data from each user, minimize data privacy leaks, and reduce the load on both the user and server sides. In step 6, in response to not receiving the global model parameters sent by the server within a preset time, the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters is determined as the trained local question-answer privacy relevance generation model. Multiple rounds of training can improve the accuracy of model evaluation. Then, based on the trained local question-answer privacy relevance generation model, a set of privacy relevance values ​​is generated for the obtained set of dialogue response information to be predicted. This improves the accuracy of the generated privacy relevance values. Finally, based on the set of privacy relevance values, information flow interception control is performed on the user side corresponding to the set of dialogue response information to be predicted. This improves the accuracy of information flow interception control and reduces the load on the user side. It can be concluded that the response trust prediction method for privacy-sensitive conversation data can train the model through local samples on the user side, which can improve data security and protect privacy data. By generating a privacy-related numerical set of response information sets to be predicted, the relevance and privacy of the conversation context can be improved, invalid interactions can be reduced, and user experience can be improved. By only sending model parameters to the server side, the waste of transmission resources and the load on the server and each user side can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that components and elements are not necessarily drawn to scale.

[0013] Figure 1 is a flowchart of some embodiments of a method for predicting the credibility of a reply based on private conversation data according to the present disclosure; Figure 2 Schematic diagram of joint training of a model on each user side and a server side according to the method for predicting the trustworthiness of responses based on private conversation data disclosed in the present invention; Figure 3 1 is a schematic structural diagram of some embodiments of a device for predicting the credibility of a reply based on private conversation data according to the present disclosure; Figure 4 It is a structural diagram of an electronic device suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION

[0014] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0015] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other.

[0016] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0017] It should be noted that the modifications of "one" and "multiple" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, they should be understood as "one or more".

[0018] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0019] The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0020] Figure 1 The process 100 of some embodiments of the method for evaluating privacy relevance of conversation data according to the present disclosure is shown. The method for evaluating privacy relevance of conversation data includes the following steps: Step 101: Receive initial model parameters sent by the server.

[0021] In some embodiments, the execution entity (e.g., an electronic device) of the above-mentioned method for predicting the trustworthiness of responses for private conversation data can receive initial model parameters sent by a server via a wired or wireless connection. The server can be a server that aggregates the model parameters sent by each user terminal and sends them to each user terminal. Each user terminal can be a terminal determined by the server based on the quality of the user terminal's local sample data and the adequacy of its computing resources. The initial model parameters can be the initial model parameters of the local model sent by the server and executed by the user terminal. The initial model parameters can be weight parameters of the local question-and-answer privacy relevance generation model. The local model can be a local question-and-answer privacy relevance generation model. The local question-and-answer privacy relevance generation model can be a neural network model that assesses the relevance of responses in a question-and-answer conversation to the preceding sentence in the question-and-answer conversation and whether the responses contain private data.

[0022] Step 102: Based on the initial model parameters, perform the following first determination step: Step 1021: perform data enhancement processing on the local question-answering dialogue training sample set to obtain an enhanced question-answering dialogue training sample set.

[0023] In some embodiments, the execution entity may perform data enhancement processing on the local question-answering dialogue training sample set to obtain an enhanced question-answering dialogue training sample set. The local question-answering dialogue training samples in the local question-answering dialogue training sample set may be samples locally owned by the user terminal, including question-answering dialogue questions and question reply information. It should be noted that, in a static data scenario, the local question-answering dialogue training sample set of each cyclic training round of the initial local question-answering privacy relevance generation model is the same. In a dynamic data scenario, the local question-answering dialogue training sample set of each cyclic training round of the initial local question-answering privacy relevance generation model is different, and the training samples of this cyclic training round are samples obtained by merging the data corresponding to the previous cyclic training round with the newly generated data increment.

[0024] In some optional implementations of some embodiments, the local question-and-answer dialogue training samples include: local question-and-answer dialogue context samples and local question-and-answer dialogue response samples. The local question-and-answer dialogue context samples may be questions in a question-and-answer dialogue. For example, the local question-and-answer dialogue context samples may be "What's your phone number?" The local question-and-answer dialogue response samples may be responses to the local question-and-answer dialogue context samples. For example, the local question-and-answer dialogue response samples may be "You can ask my secretary for my phone number."

[0025] Optionally, performing data enhancement processing on the local question-answering dialogue training sample set to obtain an enhanced question-answering dialogue training sample set may include the following steps: In the first step, for each local question-answering dialogue training sample in the above local question-answering dialogue training sample set, the following second determination step is performed: Sub-step 1: Input the local question-and-answer conversation context samples included in the local question-and-answer conversation training samples into a trained local conversation generation model to obtain a generated context response information set. The local conversation generation model may be a neural network model that generates response information for the input local question-and-answer conversation context samples. For example, the local conversation generation model may be a Chat Generative Pre-trained Transformer (ChatGPT). The generated context response information in the generated context response information set may be response information for the local question-and-answer conversation context samples generated by the local conversation generation model.

[0026] Sub-step 2, respectively determining the response privacy value of each generated context response information in the generated context response information set and the local question-and-answer dialogue response sample, as a generated response privacy value set and a sample response privacy value. The generated response privacy value in the generated response privacy value set can represent the degree to which the generated context response information contains private data. The sample response privacy value can represent the degree to which the local question-and-answer dialogue response sample contains private data. In practice, the execution entity can use a regular expression to perform regular matching processing on each generated context response information in the generated context response information set and the local question-and-answer dialogue response sample with a preset privacy database, to obtain a generated matching degree value set and a sample matching degree value, as the generated response privacy value set and the sample response privacy value. The preset privacy database can be a pre-built database that stores private data.

[0027] Sub-step 3: Input the local Q&A dialogue training sample into the trained dialogue response relevance generation model to obtain a sample context-response relevance value. The response relevance generation model may be a neural network model that evaluates the semantic relevance between the input local Q&A dialogue response sample or generated context-response information and the input local Q&A dialogue context sample. For example, the response relevance generation model may be a BERTScore model. The sample context-response relevance value may represent the semantic relevance between the local Q&A dialogue training sample and the local Q&A dialogue training sample. For example, the local Q&A dialogue context sample may be "What is your contact number?" The first local Q&A dialogue response sample may be "You can ask my secretary for my contact number." The second local Q&A dialogue response sample may be "The weather is great today." Therefore, the relevance between the first local Q&A dialogue response sample and the local Q&A dialogue context sample is greater than the relevance between the second local Q&A dialogue response sample and the local Q&A dialogue context sample.

[0028] In sub-step 4, the local Q&A conversation context sample and each generated context response in the generated context response information set are input into the trained conversation response relevance generation model to obtain a generated context response relevance value set. The generated context response relevance values ​​in the generated context response relevance value set can represent the degree of semantic association between the local Q&A conversation context sample and the generated context response information.

[0029] Sub-step 5, based on the above-mentioned generative response privacy value set, the above-mentioned sample response privacy value, the above-mentioned sample previous response relevance value and the above-mentioned generative previous response relevance value set, the above-mentioned generative previous response information set is filtered to obtain the filtered generative response information set.

[0030] As an example, the execution subject may first filter out at least one generative above reply information whose corresponding generative above reply relevance value is greater than the sample above reply relevance value from the generative above reply information set. Then, multiple generative above reply relevance values ​​whose corresponding generative above reply privacy value in the at least one generative above reply information is greater than or equal to the sample above reply privacy value are determined as positive sample above reply information samples. Afterwards, multiple generative above reply relevance values ​​whose corresponding generative above reply privacy value in the at least one generative above reply information is less than the sample above reply privacy value are determined as negative sample above reply information samples. Finally, the positive sample above reply information samples and the negative sample above reply information samples are determined as the filtered generative reply information set.

[0031] Sub-step 6: Determine the above-mentioned local question-answering dialogue training samples and the above-mentioned filtered generative response information set as enhanced question-answering dialogue training samples.

[0032] In some optional implementations of some embodiments, respectively determining the response privacy value of each generated context response information in the generated context response information set and the response privacy value of the local question-and-answer dialogue response sample as the generated response privacy value set and the sample response privacy value may include the following steps: In the first step, for each generated context reply information in the generated context reply information set, the following third determination step is performed: Sub-step 1: Perform triple extraction processing on the above-mentioned generated context reply information to obtain a generated context reply triple set. The generated context reply triples in the above-mentioned generated context reply triple set may be triples containing entity segmentations and entity segmentation relationships in the generated context reply information.

[0033] As an example, the execution entity may utilize a Text2Graph model to perform triple extraction processing on the generative contextual response information to obtain a set of generative response triples. The Text2Graph model may be a model that performs entity recognition, relationship extraction, and triple normalization on the input generative contextual response information to obtain a structured knowledge graph triple output. The Text2Graph model may be a model based on an end-to-end generative model framework. The end-to-end generative model may be a T5 model (Text to Text Transfer Transformer) or a GPT model (Generative Pre-trained Transformer).

[0034] Sub-step 2: Perform private data identification on each generated response triple in the generated response triple set to obtain a private data identification result set. The private data identification results in the private data identification result set may indicate whether the generated response triple contains private data. In practice, the execution entity may match the generated response triple set with the pre-set privacy database to obtain the private data identification result set.

[0035] Sub-step 3: determining at least one private data identification result in the private data identification result set that does not contain private data.

[0036] Sub-step 4: Determine the ratio of the number of generated response triplets to the target number of response triplets as the generated response privacy value. The number of generated response triplets is the number of private data identification results included in the at least one private data identification result, and the target number of response triplets is the number of private data identification results included in the private data identification result set. The higher the generated response privacy value, the more private the generated response information is.

[0037] In the second step, the above-mentioned local question-and-answer dialogue response sample is determined as the generative context response information, and the above-mentioned third determination step is performed again to obtain the generative response privacy value as the sample response privacy value.

[0038] In step 1022, the enhanced question-answer dialogue training sample set is input into the correlation generation model included in the initial local question-answer privacy correlation generation model corresponding to the initial model parameters to obtain the question-answer correlation value.

[0039] In some embodiments, the execution entity may input the enhanced question-answering dialogue training sample set into the correlation generation model included in the initial local question-answering privacy correlation generation model corresponding to the initial model parameters to obtain a question-answering correlation value. The correlation generation model may be a neural network model that performs correlation evaluation on the enhanced local question-answering dialogue context sample and the enhanced local question-answering dialogue reply sample included in the input enhanced question-answering dialogue training sample. The correlation generation model may be a network model composed of a BERTScore model and a multi-layer perceptron. The question-answering correlation value may represent the degree of semantic association between the enhanced local question-answering dialogue context sample and the enhanced local question-answering dialogue reply sample. The initial local question-answering privacy correlation generation model corresponding to the initial model parameters may be a model parameter of the initial local question-answering privacy correlation generation model as the initial parameter.

[0040] As an example, the execution entity may first input the enhanced question-answer dialogue training sample set into the BERTScore model included in the relevance generation model to obtain a set of question-answer context-response similarity feature vectors. Then, the set of question-answer context-response similarity feature vectors may be input into a multilayer perceptron for nonlinear processing to obtain a question-answer relevance value.

[0041] Step 1023: Input the enhanced question-answer dialogue training sample set into the privacy generation model included in the initial local question-answer privacy relevance generation model to obtain the question-answer privacy value.

[0042] In some embodiments, the execution entity may input the enhanced question-and-answer dialogue training sample set into a privacy generation model included in the initial local question-and-answer privacy relevance generation model to obtain a question-and-answer privacy value. The question-and-answer privacy value may represent the degree to which the enhanced local question-and-answer dialogue response samples contain private data. The privacy generation model may be a model that assesses the degree to which the enhanced local question-and-answer dialogue response samples included in the input enhanced question-and-answer dialogue training sample set contain private data. The privacy generation model may first input the enhanced question-and-answer dialogue training sample set into a Text2Graph model to obtain multiple response triple sets; then, using a preset privacy database, determine the privacy value of each response triple set in the multiple response triple sets to obtain a privacy value set; and finally, determine the average of the privacy value sets as the question-and-answer privacy value model.

[0043] In step 1024, based on the question-answer relevance value and the question-answer privacy value, the initial model parameters of the initial local question-answer privacy relevance generation model are updated through parameter comparison learning to obtain updated local model parameters.

[0044] In some embodiments, the execution entity may perform parameter comparison learning and updating on the initial model parameters of the initial local question-answer privacy relevance generation model based on the question-answer relevance value and the question-answer privacy value, thereby obtaining updated local model parameters. The updated local model parameters may be parameters obtained by updating the weight parameter values ​​corresponding to the model weight parameters.

[0045] As an example, the execution entity may first input the question-answer relevance value and the question-answer privacy value into a generative model contrastive loss function to obtain a generative model loss function value. The generative model loss function may be a contrastive learning loss function. Then, the generative model loss function value is input into the partial derivative of the contrastive learning loss function with respect to a weight parameter to obtain a model contrastive gradient. Finally, the difference between the initial model parameters and the model contrastive gradient is determined as the updated local model parameters.

[0046] In some optional implementations of some embodiments, performing parameter comparison learning and updating on the model parameters of the initial local question-answer privacy relevance generation model based on the question-answer relevance value and the question-answer privacy value to obtain updated local model parameters may include the following steps: The first step is to determine initial correlation weights and initial privacy weights for the question-answer correlation value and the question-answer privacy value. Both the initial correlation weights and the initial privacy weights can be preset values. For example, both the initial correlation weights and the initial privacy weights can be 0.5.

[0047] The second step is to dynamically adjust the initial relevance weight and the initial privacy weight to obtain adjusted relevance weight and adjusted privacy weight. The dynamic weight adjustment can be based on the data characteristics of the domain to which the local question-and-answer dialogue training sample set belongs. For example, if the local question-and-answer dialogue training sample set is from the medical field, the adjusted relevance weight can be 0.4, and the adjusted privacy weight can be 0.6.

[0048] The third step is to determine the product of the adjusted relevance weight value and the question-answer relevance value as the relevance loss function value.

[0049] The fourth step is to determine the product of the adjusted privacy weight value and the question-answer privacy value as the privacy loss function value.

[0050] Step 5: Determine the sum of the above correlation loss function value and the above privacy loss function value as the model loss function value.

[0051] Step 6: Based on the model loss function value, determine the local model gradient value of the initial local question-answer privacy relevance generation model. The local model gradient value can be the update direction of the initial model parameters, that is, how to update the weight parameters to minimize the loss function.

[0052] As an example, the execution entity may input the model loss function value into the partial derivative function of the model loss function corresponding to the model loss function value with respect to the weight parameter to obtain a local model gradient value.

[0053] Step 7: Determine the product of a preset model learning rate for the initial local question-answer privacy-related relevance generation model and the local model gradient value as the target model parameter value. The preset model learning rate may be a pre-set value that controls the update step size of the weight parameters of the initial local question-answer privacy-related relevance generation model.

[0054] In the eighth step, the difference between the parameter value corresponding to the initial model parameter and the target model parameter value is determined to obtain the updated local model parameter value.

[0055] In the ninth step, the initial model parameters and the updated local model parameter values ​​are determined as updated local model parameters.

[0056] In the process of adopting technical solutions to solve the above-mentioned technical problem 1, the following technical problem often arises: when training the initial local question-answer privacy relevance generation model, how to determine the impact of privacy and relevance on the accuracy of the initial local question-answer privacy relevance generation model, so as to improve the evaluation accuracy of the initial local question-answer privacy relevance generation model, thereby improving the accuracy of the model parameters, reducing the number of training rounds for the initial local question-answer privacy relevance generation model and the waste of computing resources on the user side, and improving training efficiency. Regarding the above-mentioned technical problem 2, the conventional solution generally adopts fixed weights to determine the loss function of the initial local question-answer privacy relevance generation model with respect to privacy and relevance. However, the above-mentioned conventional solution still has the following problems: due to the use of fixed weights, the applicability to question-answer dialogue data in specific application scenarios is low, which reduces the accuracy of the local question-answer privacy relevance generation model. In order to improve the accuracy of the model, more rounds of model training are required, resulting in the waste of computing resources and transmission resources on the user side, reducing model training efficiency and user experience. Considering the shortcomings of the above-mentioned conventional solutions and combining the advantages and technical status of the dynamic weight adjustment technology possessed by the inventor's company, we decided to adopt the following solution: Optionally, the dynamic weight adjustment process of the initial correlation weight value and the initial privacy weight value to obtain the adjusted correlation weight value and the adjusted privacy weight value may include the following steps: The first step is to obtain a user rating value and dialogue response requirement information for the dialogue response information set to be predicted. The user rating value may represent the user's satisfaction with the dialogue response information set to be predicted. The user rating value may be a value between 1 and 5. The dialogue response requirement information may include information about the user's requirements for the dialogue response information set to be predicted. For example, the dialogue response requirement information may include requirements for a sensitive information filtering task.

[0057] The second step is to determine the difference between the above user rating value and the preset value as the user rating difference.

[0058] The third step is to determine the product of the above user rating difference and the above preset model learning rate as the user weight control value.

[0059] The fourth step is to respectively determine the difference between the above-mentioned initial correlation weight value and the above-mentioned initial privacy weight value and the above-mentioned user weight control value as the user correlation weight value and the user privacy weight value.

[0060] Step 5: Determine the conversation domain type information for the predicted conversation reply information set and the corresponding conversation context information set. The conversation domain type information may include information about the domains to which the predicted conversation reply information set and the corresponding conversation context information set belong. For example, the conversation domain type information may include, but is not limited to, at least one of the following: healthcare, finance, and e-commerce.

[0061] Step 6: Determine a domain relevance weight and a domain privacy weight based on the conversation domain type information. The domain relevance weight can represent the level of attention paid to the semantic relevance of the conversation domain type information. The domain privacy weight can represent the level of attention paid to the privacy of the conversation domain type information.

[0062] As an example, the execution entity may use the conversation domain type information to search a preset domain weight setting table to obtain domain relevance weight values ​​and domain privacy weight values. The preset domain weight setting table may be a pre-set weight value for different conversation domains in terms of contextual relevance and privacy.

[0063] In the seventh step, based on the above-mentioned dialogue response requirement information, a requirement relevance weight value and a requirement privacy weight value are determined. The requirement relevance weight value can represent the degree to which the dialogue response requirement information focuses on relevance. The requirement privacy weight value can represent the degree to which the dialogue response requirement information focuses on privacy.

[0064] As an example, the execution entity may utilize a grid search algorithm to reply to the demand information in the conversation and determine a demand relevance weight value and a demand privacy weight value.

[0065] In the eighth step, the user relevance weight value, the field relevance weight value and the demand relevance weight value are weighted and summed to obtain the target relevance weight value.

[0066] In the ninth step, the user privacy weight value, the domain privacy weight value, and the demand privacy weight value are weighted and summed to obtain a target privacy weight value.

[0067] In the tenth step, the target relevance weight value and the target privacy weight value are normalized to obtain the normalized relevance weight value and the normalized privacy weight value as the adjusted relevance weight value and the adjusted privacy weight value.

[0068] In the eleventh step, the initial local question-answer privacy relevance generation model is trained based on the adjusted relevance weight value and the adjusted privacy weight value, and the trained model parameters obtained after the model training are sent to the server.

[0069] The above technical solution and its related contents, as an inventive point of an embodiment of the present disclosure, solve the second technical problem mentioned in the background technology: "Due to the use of fixed weights, the applicability of question-answering conversation data for specific application scenarios is low, which reduces the accuracy of the local question-answering privacy relevance generation model. In order to improve the accuracy of the model, more rounds of model training and model parameter transmission are required, resulting in waste of computing resources and transmission resources on the user side, and reducing model training efficiency and user experience." The factors that lead to the waste of computing resources and transmission resources on the user side and reduce model training efficiency and user experience are often as follows: Due to the use of fixed weights, the applicability of question-answering conversation data for specific application scenarios is low, which reduces the accuracy of the local question-answering privacy relevance generation model. In order to improve the accuracy of the model, more rounds of model training and model parameter transmission are required. If the above factors are solved, it is possible to reduce the waste of computing resources and transmission resources on the user side, improve model training efficiency and user experience, and make the evaluation results more in line with actual needs. To achieve this effect, the present disclosure first determines initial relevance weights and initial privacy weights, facilitating subsequent dynamic adjustment of these initial relevance weights and initial privacy weights to better suit the specific application scenario of the predicted conversation response information. Then, the initial relevance weights and initial privacy weights are dynamically adjusted based on three factors: user rating values, conversation response requirement information, and conversation response requirement information. Comprehensively considering these three different perspectives, the initial relevance weights and initial privacy weights are designed to better suit the application scenario and user needs of the predicted conversation response information, thereby improving the accuracy of the adjusted relevance weights and adjusted privacy weights. Finally, the initial local question-answer privacy relevance generation model is trained based on the adjusted relevance weights and adjusted privacy weights, and the trained model parameters obtained after model training are sent to the server. This reduces the number of training rounds required for the initial local question-answer privacy relevance generation model to reach a predetermined accuracy, improving model training efficiency, reducing training time and user-side computing resources, and reducing the waste of transmission resources for communicating with the server. This improves user experience and reduces the training load on both the user and server sides.

[0070] Step 1025 : Send the updated local model parameters to the server, perform parameter aggregation on the received updated local model parameters to obtain global model parameters, and send the global model parameters to each user terminal.

[0071] In some embodiments, the execution entity may send the updated local model parameters to the server to aggregate the received updated local model parameters to obtain global model parameters, and send the global model parameters to each user terminal. The global model parameters may be the average of the updated local model parameters. The user terminal may be a terminal that receives the model parameters sent by the server and uses locally stored samples to perform local model training. The execution entity may be any of the user terminals.

[0072] Step 1026: In response to not receiving the global model parameters sent by the server within the preset time, the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters is determined as the trained local question-answer privacy relevance generation model.

[0073] In some embodiments, in response to not receiving the global model parameters from the server within a preset time, the execution entity may determine the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters as the trained local question-answer privacy relevance generation model. The preset time may be a pre-set time at which the user receives information from the server. Failure to receive the global model parameters from the server within the preset time indicates that model training on both the server and the user is complete. Figure 2 A schematic diagram shows the joint training of the local question-answering privacy relevance generation model between various user ends and the server end using the local question-answering dialogue training sample set.

[0074] Step 103: Generate a privacy relevance value set for the acquired dialogue response information set to be predicted based on the trained local question-answer privacy relevance generation model.

[0075] In some embodiments, the execution entity may generate a set of privacy relevance values ​​for the obtained set of dialogue response information to be predicted based on the trained local question-and-answer privacy relevance generation model. The dialogue response information to be predicted in the set of dialogue response information to be predicted may be a response information from a previous conversation awaiting privacy relevance assessment. The privacy relevance values ​​in the set of privacy relevance values ​​may be a weighted sum of question-and-answer privacy and question-and-answer relevance.

[0076] As an example, the above-mentioned dialogue response information set to be predicted is input into the trained local question-answering privacy relevance generation model to perform privacy relevance evaluation and obtain a privacy relevance value set.

[0077] Optionally, before generating a privacy relevance value set for the acquired dialogue response information set to be predicted based on the trained local question-answer privacy relevance generation model, the following steps may be performed: In response to receiving the global model parameters sent by the server within a preset time, the global model parameters are determined as the initial model parameters, and the local question-answering dialogue training sample set is incrementally processed to obtain a local incremental question-answering dialogue training sample set as the local question-answering dialogue training sample set, so as to perform the first determination step again. In the case of dynamic samples, this can be a sample set that is partially fused with the historical local question-answering dialogue training sample set from the previous round of training and the local question-answering dialogue training sample set newly generated by the user end. In the case of static samples, the local incremental question-answering dialogue training sample set can be the sample set for initial model training.

[0078] Step 104 : Based on the privacy relevance value set, information flow interception control is performed on the user terminal corresponding to the predicted dialogue reply information set.

[0079] In some embodiments, based on the above-mentioned privacy relevance value set, information flow interception control is performed on the user terminal corresponding to the above-mentioned conversation reply information set to be predicted.

[0080] As an example, the execution entity may first filter out at least one privacy relevance value from the set of privacy relevance values ​​that is less than or equal to a preset privacy relevance threshold. Then, the execution entity may control the interception of the set of predicted conversation response information corresponding to the at least one privacy relevance value, and transmit at least one predicted conversation response information that is greater than the preset privacy relevance threshold to the user terminal for display.

[0081] The above-mentioned embodiments of the present disclosure have the following beneficial effects: In some embodiments of the present disclosure, the reply trustworthiness prediction method for private conversation data trains a model using local user samples, thereby improving data security and protecting user privacy. Furthermore, by generating a privacy-related value set for the set of reply information to be predicted, the relevance and privacy of the conversation context can be improved, invalid interactions can be reduced, and the user experience can be improved. Furthermore, by only sending model parameters to the server, transmission resource waste and the load on the server and each user can be reduced. Specifically, the reasons for the low security of the relevant sample data, the risk of privacy data leakage, the waste of a large amount of transmission resources, the increased load on the server, and the low accuracy of reply information evaluation, which reduces the user experience, are: the transmission of conversation sample data from each user to the server for unified model training is prone to privacy data leakage during the transmission process, requiring a large amount of transmission resources, resulting in low sample data security, the risk of privacy data leakage, the waste of a large amount of transmission resources, and the increased load on the server. Furthermore, in question-and-answer conversations, only a single metric, such as relevance or privacy, is evaluated on the reply information, resulting in low accuracy of reply information evaluation and a reduced user experience. Based on this, some embodiments of the present disclosure may first receive initial model parameters sent by a server. The initial model parameters facilitate subsequent local model parameter updates on the user side. Next, based on the initial model parameters, the following first determination step is performed: First, data enhancement processing is performed on the local question-and-answer conversation training sample set to obtain an enhanced question-and-answer conversation training sample set. Processing the local question-and-answer conversation training sample set can improve the data quality of the local question-and-answer conversation training sample set. Second, the enhanced question-and-answer conversation training sample set is input into a correlation generation model included in an initial local question-and-answer privacy relevance generation model corresponding to the initial model parameters to obtain a question-and-answer relevance value. Using the correlation of the local question-and-answer conversation training sample set to train the model can improve the model training rate. Third, the enhanced question-and-answer conversation training sample set is input into a privacy generation model included in the initial local question-and-answer privacy relevance generation model to obtain a question-and-answer privacy value. Using the privacy of the local question-and-answer conversation training sample set to train the model can improve the model training rate. In the fourth step, based on the question-answer relevance and question-answer privacy values, the initial model parameters of the initial local question-answer privacy relevance generation model are updated through parameter comparison learning to obtain updated local model parameters. Here, integrating privacy and relevance training can improve the comprehensiveness and accuracy of the model. In the fifth step, the updated local model parameters are sent to the aforementioned server, which aggregates the received updated local model parameters to obtain global model parameters. The global model parameters are then sent to each user terminal.Here, model aggregation training is performed on the server side using only the local model parameters sent by each user. This can reduce transmission resource waste, increase the security of sample data from each user, minimize data privacy leaks, and reduce the load on both the user and server sides. In step 6, in response to not receiving the global model parameters sent by the server within a preset time, the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters is determined as the trained local question-answer privacy relevance generation model. Multiple rounds of training can improve the accuracy of model evaluation. Then, based on the trained local question-answer privacy relevance generation model, a set of privacy relevance values ​​is generated for the obtained set of dialogue response information to be predicted. This improves the accuracy of the generated privacy relevance values. Finally, based on the set of privacy relevance values, information flow interception control is performed on the user side corresponding to the set of dialogue response information to be predicted. This improves the accuracy of information flow interception control and reduces the load on the user side. It can be concluded that the response trust prediction method for privacy-sensitive conversation data can train the model through local samples on the user side, which can improve data security and protect privacy data. By generating a privacy-related numerical set of response information sets to be predicted, the relevance and privacy of the conversation context can be improved, invalid interactions can be reduced, and user experience can be improved. By only sending model parameters to the server side, the waste of transmission resources and the load on the server and each user side can be reduced.

[0082] Further references Figure 3 As an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of a device for predicting the credibility of a reply for private conversation data. These device embodiments are similar to Figure 1 Corresponding to the method embodiments shown, the reply credibility prediction device for private conversation data can be specifically applied to various electronic devices.

[0083] like Figure 3As shown, a reply credibility prediction device 300 for private conversation data includes: a receiving unit 301, an executing unit 302, a generating unit 303 and a controlling unit 304. The receiving unit 301 is configured to: receive initial model parameters sent by the server. The execution unit 302 is configured to: based on the initial model parameters, perform the following first determination step: perform data enhancement processing on the local question-and-answer dialogue training sample set to obtain an enhanced question-and-answer dialogue training sample set; input the enhanced question-and-answer dialogue training sample set into the correlation generation model included in the initial local question-and-answer privacy correlation generation model corresponding to the initial model parameters to obtain a question-and-answer correlation value; input the enhanced question-and-answer dialogue training sample set into the privacy generation model included in the initial local question-and-answer privacy correlation generation model to obtain a question-and-answer privacy value; perform parameter comparison learning and update on the initial model parameters of the initial local question-and-answer privacy correlation generation model based on the question-and-answer correlation value and the question-and-answer privacy value to obtain updated local model parameters; send the updated local model parameters to the above-mentioned server to perform parameter aggregation on the received updated local model parameters to obtain global model parameters, and send the global model parameters to each user terminal; in response to not receiving the global model parameters sent by the above-mentioned server within a preset time, determine the initial local question-and-answer privacy correlation generation model corresponding to the updated local model parameters as the trained local question-and-answer privacy correlation generation model. The generation unit 303 is configured to generate a privacy relevance value set for the obtained set of dialogue response information to be predicted based on the trained local question-answer privacy relevance generation model. The control unit 304 is configured to perform information flow interception control on the user terminal corresponding to the set of dialogue response information to be predicted based on the privacy relevance value set.

[0084] It is understandable that the various units recorded in the reply credibility prediction device 300 for private conversation data and the reference Figure 1 Therefore, the operations, features and beneficial effects described above for the method are also applicable to the reply credibility prediction device 300 for private conversation data and the units contained therein, and will not be repeated here.

[0085] Reference below Figure 4 , which shows a structural schematic diagram of an electronic device (eg, an electronic device) 400 suitable for implementing some embodiments of the present disclosure. Figure 4 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0086] like Figure 4As shown, electronic device 400 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes based on programs stored in a read-only memory (ROM) 402 or programs loaded from a storage device 408 into a random access memory (RAM) 403. RAM 403 also stores various programs and data required for the operation of electronic device 400. Processing device 301, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to bus 404.

[0087] Typically, the following devices may be connected to the I / O interface 405: an input device 406 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 407 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 408 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 409. The communication device 409 may allow the electronic device 400 to communicate with other devices wirelessly or by wire to exchange data. Figure 4 The electronic device 400 is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead. Figure 3 Each block shown in the figure may represent one device, or may represent multiple devices as needed.

[0088] In particular, according to some embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In some such embodiments, the computer program can be downloaded and installed from a network via the communication device 409, or installed from the storage device 408, or installed from the ROM 402. When the computer program is executed by the processing device 401, the above-mentioned functions defined in the method of some embodiments of the present disclosure are performed.

[0089] It should be noted that in some embodiments of the present disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. Computer-readable storage media may include, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or components, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In some embodiments of the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device, or component. Furthermore, in some embodiments of the present disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. This propagated data signal may take a variety of forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wire, optical cable, RF (radio frequency), or any suitable combination thereof.

[0090] In some embodiments, the client and server can communicate using any currently known or later developed network protocol, such as HTTP (Hypertext Transfer Protocol), and can be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or later developed network.

[0091] The above-mentioned computer-readable medium may be included in the above-mentioned electronic device; or it may exist independently without being assembled into the electronic device. The above-mentioned computer-readable medium carries one or more programs. When the above-mentioned one or more programs are executed by the electronic device, the electronic device: receives the initial model parameters sent by the server; based on the initial model parameters, performs the following first determination step: performs data enhancement processing on the local question-and-answer dialogue training sample set to obtain an enhanced question-and-answer dialogue training sample set; inputs the enhanced question-and-answer dialogue training sample set into the correlation generation model included in the initial local question-and-answer privacy correlation generation model corresponding to the initial model parameters to obtain a question-and-answer correlation value; inputs the enhanced question-and-answer dialogue training sample set into the privacy generation model included in the initial local question-and-answer privacy correlation generation model to obtain a question-and-answer privacy value; and according to the question-and-answer correlation value and the question-and-answer privacy value, performs a data enhancement process on the local question-and-answer privacy correlation generation model. The initial model parameters are updated through parameter comparison learning to obtain updated local model parameters; the updated local model parameters are sent to the above-mentioned server to perform parameter aggregation on the received updated local model parameters to obtain global model parameters, and the global model parameters are sent to each user terminal; in response to not receiving the global model parameters sent by the above-mentioned server within a preset time, the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters is determined to be the trained local question-answer privacy relevance generation model; based on the above-mentioned trained local question-answer privacy relevance generation model, a privacy relevance value set is generated for the obtained dialogue response information set to be predicted; based on the above-mentioned privacy relevance value set, information flow interception control is performed on the user terminal corresponding to the above-mentioned dialogue response information set to be predicted.

[0092] Computer program code for performing the operations of some embodiments of the present disclosure may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0093] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the module, program segment, or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of the boxes in the block diagram and / or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0094] The units described in some embodiments of the present disclosure may be implemented in software or in hardware. The units described may also be provided in a processor. For example, they may be described as follows: a processor includes a receiving unit, an execution unit, a generation unit, and a control unit. The names of these units do not, in some cases, constitute limitations on the units themselves. For example, the receiving unit may also be described as a "unit that receives initial model parameters sent by the server."

[0095] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0096] The above descriptions are merely some preferred embodiments of the present disclosure and illustrate the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present disclosure is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A method for predicting the trustworthiness of responses based on private conversation data, comprising: Receive the initial model parameters sent by the server; Based on the initial model parameters, the following first determination step is performed: Perform data enhancement processing on the local question-answering dialogue training sample set to obtain an enhanced question-answering dialogue training sample set; Input the enhanced question-answering dialogue training sample set into the correlation generation model included in the initial local question-answering privacy correlation generation model corresponding to the initial model parameters to obtain the question-answer correlation value; Input the enhanced question-answering dialogue training sample set into the privacy generation model included in the initial local question-answering privacy relevance generation model to obtain the question-answering privacy value; Based on the question-answer relevance value and the question-answer privacy value, the initial model parameters of the initial local question-answer privacy relevance generation model are updated through parameter comparison learning to obtain the updated local model parameters; Sending the updated local model parameters to the server, performing parameter aggregation on the received updated local model parameters to obtain global model parameters, and sending the global model parameters to each user terminal; In response to not receiving the global model parameters sent by the server within a preset time, determining the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters as the trained local question-answer privacy relevance generation model; Generating a privacy relevance value set for the acquired dialogue response information set to be predicted based on the trained local question-answering privacy relevance generation model; According to the privacy relevance value set, information flow interception control is performed on the user terminal corresponding to the to-be-predicted conversation reply information set.

2. The method according to claim 1, wherein Before generating a privacy relevance value set for the acquired conversation response information set to be predicted based on the trained local question-answer privacy relevance generation model, the method further includes: In response to receiving the global model parameters sent by the server within a preset time, the global model parameters are determined as the initial model parameters, and the local question-answering dialogue training sample set is incrementally processed to obtain a local incremental question-answering dialogue training sample set as the local question-answering dialogue training sample set, so as to execute the first determination step again.

3. The method according to claim 1, wherein Local question-answering dialogue training samples include: local question-answering dialogue context samples and local question-answering dialogue response samples; and The data enhancement processing is performed on the local question-answering dialogue training sample set to obtain the enhanced question-answering dialogue training sample set, including: For each local question-answering dialogue training sample in the local question-answering dialogue training sample set, the following second determination step is performed: Inputting the local question-answering dialogue context samples included in the local question-answering dialogue training samples into the trained local dialogue generation model to obtain a generated context response information set; Determine the response privacy value of each generated context response information in the generated context response information set and the response sample of the local question-and-answer dialogue as the generated response privacy value set and the sample response privacy value; Inputting the local question-answering dialogue training sample into the trained dialogue response relevance generation model to obtain the sample response relevance value; Inputting the local question-and-answer conversation context sample and each generated context reply information in the generated context reply information set into the trained conversation reply relevance generation model to obtain a generated context reply relevance value set; Filtering the generated context response information set according to the generated context response privacy value set, the sample context response privacy value, the sample context response relevance value, and the generated context response relevance value set to obtain a filtered generated context response information set; The local question-answering dialogue training sample and the screened generative response information set are determined as the enhanced question-answering dialogue training sample.

4. The method according to claim 3, wherein: The step of respectively determining the response privacy values ​​of each generated context response information in the generated context response information set and the response privacy value of the local question-and-answer dialogue response sample as the generated response privacy value set and the sample response privacy value includes: For each generated context reply information in the generated context reply information set, perform the following third determination step: Performing triple extraction processing on the generated preceding reply information to obtain a generated reply triple set; Performing private data identification on each generated response triple in the generated response triple set to obtain a private data identification result set; Determining that the private data identification result set represents at least one private data identification result that does not contain private data; Determining a ratio of the number of generated response triplets to the target number of response triplets as a generated response privacy value, wherein the number of generated response triplets is the number of private data identification results included in the at least one private data identification result, and the target number of response triplets is the number of private data identification results included in the private data identification result set; The local question-and-answer dialogue response sample is determined as the generated context response information, so as to perform the third determination step again to obtain the generated response privacy value as the sample response privacy value.

5. The method according to claim 1, wherein The method performs parameter comparison learning and updating on the initial model parameters of the initial local question-answer privacy relevance generation model based on the question-answer relevance value and the question-answer privacy value to obtain updated local model parameters, including: Determining an initial correlation weight value and an initial privacy weight value of the question-answer correlation value and the question-answer privacy value; Performing dynamic weight adjustment processing on the initial correlation weight value and the initial privacy weight value to obtain an adjusted correlation weight value and an adjusted privacy weight value; Determine the product of the adjusted relevance weight value and the question-answer relevance value as the relevance loss function value; Determine the product of the adjusted privacy weight value and the question-answer privacy value as the privacy loss function value; Determine the sum of the relevance loss function value and the privacy loss function value as the model loss function value; Determining a local model gradient value of the initial local question-answer privacy relevance generation model according to the model loss function value; Determine the product of a preset model learning rate of the initial local question-answer privacy relevance generation model and the local model gradient value as the target model parameter value; Determine the difference between the parameter value corresponding to the initial model parameter and the target model parameter value to obtain an updated local model parameter value; The initial model parameters and the updated local model parameter values ​​are determined as updated local model parameters.

6. A device for predicting the credibility of responses based on private conversation data, comprising: A receiving unit, configured to receive initial model parameters sent by the server; The execution unit is configured to perform the following first determination step based on the initial model parameters: performing data enhancement processing on the local question-answering dialogue training sample set to obtain an enhanced question-answering dialogue training sample set; Inputting the enhanced question-answer dialogue training sample set into the correlation generation model included in the initial local question-answer privacy correlation generation model corresponding to the initial model parameters to obtain a question-answer correlation value; inputting the enhanced question-answer dialogue training sample set into the privacy generation model included in the initial local question-answer privacy correlation generation model to obtain a question-answer privacy value; performing parameter comparison learning and updating on the initial model parameters of the initial local question-answer privacy correlation generation model based on the question-answer correlation value and the question-answer privacy value to obtain updated local model parameters; sending the updated local model parameters to the server to perform parameter aggregation on the received updated local model parameters to obtain global model parameters, and sending the global model parameters to each user terminal; In response to not receiving the global model parameters sent by the server within a preset time, determining the initial local question-answer privacy relevance generation model corresponding to the updated local model parameters as the trained local question-answer privacy relevance generation model; a generating unit configured to generate a privacy relevance value set for the acquired dialogue response information set to be predicted based on the trained local question-answering privacy relevance generation model; The control unit is configured to perform information flow interception control on the user terminal corresponding to the to-be-predicted dialogue reply information set according to the privacy relevance value set.

7. An electronic device comprising: one or more processors; a storage device having one or more programs stored thereon, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 5.

8. A computer-readable medium having a computer program stored thereon, wherein: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images