User feedback detection method and device, electronic equipment and storage medium

Abstract

The invention provides a user feedback detection method, which comprises the following steps of: modeling preset language information according to a causal architecture to obtain a structural causal graph; the structural causal diagram is represented by the following formula. Wherein semantic content factors in the causal architecture, semantic style factors in the causal architecture, text information included in the language information and label information included in the language information are expressed. And performing content focusing on the language information according to a preset large language model to obtain a focusing representation. And modeling according to the focusing representation and the structural causal graph to obtain a causal decoupling model. And detecting the user feedback according to the causal decoupling model to obtain a detection result. Therefore, accurate recognition of user feedback by the language model in dialect, slang, non-standard sentence pattern and other multi-style scenes is realized.

Description

Technical Field

[0001] This application relates to the field of semantic recognition technology, and in particular to a user feedback detection method. Background Technology

[0002] With the rapid development of social media, online customer service, and e-commerce review platforms, users' textual expressions in public or interactive scenarios are becoming increasingly diverse, non-standardized, and personalized. Users of different ages, regions, or community backgrounds often exhibit specific linguistic styles when expressing opinions, emotions, or complaints, such as colloquial expressions, dialectal vocabulary, internet slang, and emoticons. The widespread existence of these stylized languages ​​poses a significant challenge to the fairness and robustness of traditional text classification models that rely on statistical correlation in detecting negative user feedback.

[0003] Current mainstream user feedback analysis methods are mostly based on downstream fine-tuning of large-scale pre-trained language models (such as BERT and RoBERTa) and employ supervised learning paradigms to distinguish between "positive / negative feedback" or "complaint categories." However, these models typically use end-to-end feature compression and representation learning strategies, lacking explicit structured attribution mechanisms. They cannot effectively distinguish between semantic content factors that truly express emotions or dissatisfaction in the text and stylistic features that merely reflect personal expression habits. When there are structural biases in the distribution of training data, such as certain dialects, regional expressions, or colloquial phrases appearing more frequently in negative samples, the model is prone to learning spurious correlations between "language style and negative labels." This can lead to systematic biases towards specific groups or expression styles during the inference stage, and even misclassify ordinary neutral complaints or suggestions as negative complaints. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a user feedback detection method, device, electronic device and storage medium.

[0005] To achieve the above objectives, this application provides a user feedback detection method, comprising: modeling preset language information based on a causal architecture to obtain a structural causal graph. The structural causal graph is represented by the following equation: .in, This represents the semantic content factor in the causal framework. This represents the semantic style factor in the causal framework. This refers to the text information included in the language information. This represents the tag information included in the language information. Based on a pre-defined large language model, the language information is focused to obtain a focused representation. A causal decoupling model is then derived by modeling the focused representation and the structural causal graph. User feedback is then detected using the causal decoupling model to obtain the detection results.

[0006] In some embodiments, content-focused language information is performed based on a pre-defined large language model to obtain a focused representation. Specifically, this includes: performing context learning on text information based on the large language model to obtain intermediate layer output; calculating the sensitivity of each token in the intermediate layer output to loss to obtain a token representation; and determining the focused representation based on the token representation and the intermediate layer output.

[0007] In some embodiments, a causal decoupling model is obtained by modeling based on the focused representation and the structural causal graph. Specifically, this includes: mapping the structural causal graph based on the focused representation to obtain latent variable information; and determining the causal decoupling model based on the latent variable information and a large language model.

[0008] In some embodiments, mapping the structural causal graph according to the focus representation to obtain latent variable information specifically includes: mapping content factors according to the focus representation to obtain content latent variables; mapping style factors according to the focus representation to obtain style latent variables; and optimizing the content latent variables and style latent variables according to a preset task prediction loss to obtain latent variable information.

[0009] In some embodiments, the latent variable information includes optimized content latent variables and optimized style latent variables. Determining a causal decoupling model based on the latent variable information and the large language model specifically includes: mapping the optimized content latent variables using a pre-defined latent variable injection module to obtain mapped content latent variables; mapping the optimized style latent variables using the same module to obtain mapped style latent variables; and injecting the mapped content latent variables and mapped style latent variables into the large language model using a residual method to obtain the causal decoupling model.

[0010] A user feedback detection device includes: a causal graph construction module, used to model preset language information based on a causal architecture to obtain a structural causal graph. The structural causal graph is represented by the following equation: .in, This represents the semantic content factor in the causal framework. This represents the semantic style factor in the causal framework. This refers to the text information included in the language information. This module represents the tag information included in the language information. The focusing module focuses the language information based on a pre-defined large language model to obtain a focused representation. The modeling module builds a causal decoupling model based on the focused representation and the structural causal graph. The detection module detects user feedback based on the causal decoupling model to obtain detection results.

[0011] In some embodiments, the focusing module specifically includes: a representation module, used to perform context learning on text information based on a large language model to obtain intermediate layer output; calculate the sensitivity of each token in the intermediate layer output to loss to obtain a token representation; and determine the focusing representation based on the token representation and the intermediate layer output.

[0012] In some embodiments, the modeling module specifically includes: a mapping module, used to map the structural causal graph according to the focused representation to obtain latent variable information; and to determine a causal decoupling model based on the latent variable information and the large language model.

[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any of the preceding descriptions.

[0014] A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform any of the methods described above.

[0015] As can be seen from the above, the user feedback detection method provided in this application models language information based on a causal architecture, sets causal relationships between semantic content and semantic style and text information and tag information respectively, focuses language information through focused representation, and finally injects latent variables and performs causal modeling on a large language model through the latent variable information obtained from the focused representation. This enables the model to distinguish between semantic content and style content in language information, thereby ensuring that the model maintains stable discrimination ability in multi-style scenarios such as dialects, slang, and non-standard sentence structures. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a user feedback detection method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a user feedback detection device provided in an embodiment of this application; Figure 3 This is a schematic diagram of a more specific electronic device hardware structure provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0021] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0022] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0023] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0024] With the rapid development of social media, online customer service, and e-commerce review platforms, users' textual expressions in public or interactive scenarios are becoming increasingly diverse, non-standardized, and personalized. Users of different ages, regions, or community backgrounds often exhibit specific linguistic styles when expressing opinions, emotions, or complaints, such as colloquial expressions, dialectal vocabulary, internet slang, and emoticons. The widespread existence of these stylized languages ​​poses a significant challenge to the fairness and robustness of traditional text classification models that rely on statistical correlation in detecting negative user feedback.

[0025] Current mainstream user feedback analysis methods are mostly based on downstream fine-tuning of large-scale pre-trained language models (such as BERT and RoBERTa) and employ supervised learning paradigms to distinguish between "positive / negative feedback" or "complaint categories." However, these models typically use end-to-end feature compression and representation learning strategies, lacking explicit structured attribution mechanisms. They cannot effectively distinguish between semantic content factors that truly express emotions or dissatisfaction in the text and stylistic features that merely reflect personal expression habits. When there are structural biases in the distribution of training data, such as certain dialects, regional expressions, or colloquial phrases appearing more frequently in negative samples, the model is prone to learning spurious correlations between "language style and negative labels." This can lead to systematic biases towards specific groups or expression styles during the inference stage, and even misclassify ordinary neutral complaints or suggestions as negative complaints.

[0026] Style bias is essentially a type of "non-causal misprediction" caused by confounding variables. This bias not only directly reduces the reliability and generalization ability of models in multilingual style scenarios, but also limits the application value of intelligent customer service, sentiment analysis, and user experience management systems in real business environments. For example, when a model incorrectly classifies user comments with regional accents as negative feedback, it may lead to incorrect customer service assignments, biased public opinion responses, and even affect the platform's fairness assessment. Furthermore, traditional attention mechanisms or saliency analysis methods cannot analyze the sources of bias within the model from a causal perspective, thus making it difficult to fundamentally decouple semantics from style.

[0027] Therefore, how to construct a user negative feedback detection method with causal structure modeling capabilities, the ability to explicitly distinguish between semantic content factors and linguistic style factors, and the ability to evaluate the model's sensitivity and robustness to style features through counterfactual intervention mechanisms has become a key problem urgently needing to be solved in the field of intelligent text understanding and user sentiment analysis. An ideal solution should effectively suppress the interference of non-causal style paths on prediction results while preserving the model's sensitivity to users' genuine dissatisfaction semantics, thereby achieving a unified improvement in fairness, interpretability, and performance.

[0028] To address the aforementioned challenges, this invention, from the perspective of causal representation learning, introduces structural causal graph modeling to model the causal path between text generation and label prediction. By explicitly decoupling and modeling semantic content latent factors and style latent factors, and combining style counterfactual sample construction, prediction consistency verification, and style invariance optimization mechanisms, a user negative feedback detection method with causal inference and bias identification capabilities is proposed. This method can significantly improve the robustness and generalization ability of the model in multi-style environments, enabling quantitative assessment and interpretable intervention of potential model biases, and providing highly reliable technical support for scenarios such as intelligent customer service, sentiment analysis, and user experience management.

[0029] like Figure 1 As shown, this application provides a user feedback detection method, including: Step S1: Model the preset language information according to the causal architecture to obtain a structural causal graph. The structural causal graph is represented by the following equation: .

[0030] in, This represents the semantic content factor in the causal framework. This represents the semantic style factor in the causal framework. This refers to the text information included in the language information. This refers to the tag information included in the language information.

[0031] In this embodiment, by constructing a semantic content factor With style factor A causal graph is used to characterize the generation mechanism of input text X and label Y. This causal modeling explicitly states that content factors directly determine the prediction results, while style factors only indirectly affect the prediction through the generated text and should not be used as a basis for classification.

[0032] Step S2: Focus the language information on the content according to the preset large language model to obtain the focused representation.

[0033] In this embodiment, the large language model can be an open-source large language model, which can effectively suppress the interference of non-causal style components by focusing on the language information based on causal attribution.

[0034] Step S3: Model the causal decoupling model based on the focused representation and the structural causal graph.

[0035] Step S4: Detect user feedback based on the causal decoupling model and obtain the detection results.

[0036] In this embodiment, language information is modeled according to a causal architecture, and causal relationships between semantic content and semantic style and text information and tag information are set respectively. Then, the language information is focused by focusing representation. Finally, the latent variable information obtained by focusing representation is used to inject latent variables and perform causal modeling on the large language model. This enables the semantic content and style content in language information to be distinguished, thereby ensuring that the model maintains stable discrimination ability in multi-style scenarios such as dialects, slang, and non-standard sentence structures.

[0037] In some embodiments, language information includes at least text information.

[0038] Step S2 specifically includes: Step S21: Context learning is performed on the text information based on the large language model to obtain the intermediate layer output.

[0039] In this embodiment, an open-source large model (such as LLaMA) is used as the semantic encoding backbone to learn the context representation of the input text X, resulting in the intermediate layer output:

[0040] in, This represents the contextual latent vector of the nth token. For hidden layer dimensions.

[0041] Step S22: Calculate the sensitivity of each token in the intermediate layer output to the loss to obtain the token representation.

[0042] To highlight task-related semantic content, this invention introduces gradient attribution and gating mechanisms in the intermediate layer to calculate the sensitivity of each token to loss. And adjust the token representation through a learnable gating function:

[0043] in For the Sigmoid function, W represents the weight of each token's contribution to the final semantics, and W,b are learnable parameters.

[0044] Step S23: Determine the focused representation based on the token representation and the intermediate layer output.

[0045] In this embodiment, the focus is represented by the following formula: .

[0046] In some embodiments, step S3 specifically includes: Step S31: Map the structural causal graph according to the focused representation to obtain latent variable information.

[0047] In this embodiment, causal separation of style and semantics is achieved by mapping text in the structural causal graph to the latent variable space.

[0048] Step S32: Determine the causal decoupling model based on the latent variable information and the large language model.

[0049] In this embodiment, latent variable information is mapped into a large language model, thereby enabling the adjustment of the parameters of the large language model and ultimately obtaining a detection model suitable for user feedback scenarios.

[0050] In some embodiments, the structural causal graph includes content factors and style factors.

[0051] Step S31 specifically includes: Step S311: Map the content factors according to the focus representation to obtain the content latent variables.

[0052] In this embodiment, for content branches, an Information Bottleneck (IB) mechanism is employed to obtain a minimum sufficient semantic representation through reparameterized sampling.

[0053] in, Output from the multilayer perceptron ensures Only retain information related to feedback prediction. A random noise vector sampled from a standard normal distribution, used to introduce a controllable perturbation. Step S312: Map the style factors according to the focus representation to obtain the style latent variables.

[0054] In this embodiment, for the style branch, a variational autoencoder structure is used to model the remaining information:

[0055] Adversarial constraints are introduced, and an auxiliary discriminator D is trained through a gradient reversal layer. To predict the label Y, thus forcing the style latent variable to contain no label information, achieving causal separation between style and semantics.

[0056] Step S313: Optimize the content latent variables and style latent variables according to the preset task prediction loss to obtain latent variable information.

[0057] In this embodiment, during training, Kullback–Leibler divergence constraints are applied to the two latent variables respectively to maintain distribution stability:

[0058] Simultaneously minimize adversarial losses This is to achieve style-tag dereferencing.

[0059] As an optional implementation, an objective function can also be set to jointly optimize the task prediction loss with the above-mentioned constraints. The overall objective function is: ; in, For cross-entropy loss, This is a weighting factor.

[0060] Stable semantic latent variables can be obtained simultaneously through end-to-end joint optimization. Style latent variables This enables causally separable representation learning.

[0061] In some embodiments, the latent variable information includes latent variables for optimizing content and latent variables for optimizing style.

[0062] Step S32 specifically includes: Step S321: Map the latent variables of the optimized content according to the preset latent variable injection module to obtain the mapped latent variables of the content.

[0063] Step S322: Map the latent variables of the optimization style according to the latent variable injection module to obtain the mapped style latent variables.

[0064] In this embodiment, the latent variables are mapped back to the intermediate layer in the large language model through the latent variable mapping module:

[0065] The MLP (Multilayer Perceptron) consists of multiple fully connected networks and nonlinear activation functions, used to process latent variables. , Each variable is mapped to a vector space with the same feature dimensions as the intermediate layer of the language model, thereby achieving dimensional alignment and controllable fusion between latent variables and internal features of the model.

[0066] Step S323: Inject the latent variables of the mapping content and the latent variables of the mapping style into the large language model according to the residual method to obtain the causal decoupling model.

[0067] In this embodiment, the residual injection method is represented by the following formula: ; in, This indicates the layer number at which latent variables are injected.

[0068] This design allows for dynamic adjustment of the model's internal representation during inference, enabling controllable generation and discrimination of semantics and style.

[0069] As an optional implementation method, when keeping the content latent variables unchanged and replacing the style latent variables... When the model generates output, it should satisfy the following:

[0070] This verifies the model's robustness to style perturbations and achieves the goal of counterfactual consistency.

[0071] As an alternative implementation, during the inference phase, the model first performs intermediate layer encoding and attribution gating on the new input text, and then infers latent variables. The system outputs the probability of negative feedback based on content latent variables. Optionally, counterfactual samples are generated by replacing style latent variables to explain prediction stability and identify potential sources of bias. The final result provides both discriminative output and an interpretable causal path description, facilitating deployment in intelligent customer service, public opinion monitoring, and user experience analysis.

[0072] As can be seen from the above embodiments of this application, the user feedback detection method provided by this application models language information according to a causal architecture, sets causal relationships between semantic content and semantic style and text information and tag information respectively, focuses language information through focused representation, and finally injects latent variables and performs causal modeling on a large language model through the latent variable information obtained by focused representation. In this way, it can distinguish between semantic content and style content in language information, thereby ensuring that the model maintains stable discrimination ability in multi-style scenarios such as dialects, slang, and non-standard sentence structures.

[0073] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0074] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0075] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a user feedback detection device.

[0076] refer to Figure 2 User feedback detection device, including: The causal graph construction module 100 is used to model the preset language information based on the causal architecture to obtain a structural causal graph. The structural causal graph is represented by the following formula: .

[0077] in, This represents the semantic content factor in the causal framework. This represents the semantic style factor in the causal framework. This refers to the text information included in the language information. This refers to the tag information included in the language information.

[0078] The focusing module 200 is used to focus the language information according to the preset large language model to obtain the focused representation.

[0079] Modeling module 300 is used to model based on the focused representation and structural causal graph to obtain a causal decoupling model.

[0080] The detection module 400 is used to detect user feedback based on the causal decoupling model and obtain the detection results.

[0081] In some embodiments, the focusing module 200 specifically includes: The representation module is used to learn the context of text information based on the large language model and obtain the intermediate layer output.

[0082] Calculate the sensitivity of each token in the intermediate layer output to the loss to obtain the token representation.

[0083] The focus representation is determined based on the token representation and the intermediate layer output.

[0084] In some embodiments, the modeling module 300 specifically includes: The mapping module is used to map the structural causal graph based on the focused representation to obtain latent variable information.

[0085] The causal decoupling model is determined based on latent variable information and the large language model.

[0086] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0087] The apparatus described above is used to implement the corresponding user feedback detection method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0088] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the user feedback detection method described in any of the above embodiments.

[0089] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0090] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0091] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0092] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0093] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0094] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0095] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0096] The electronic devices described above are used to implement the corresponding user feedback detection methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0097] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the user feedback detection method as described in any of the above embodiments.

[0098] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0099] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the user feedback detection method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0100] It should be noted that the embodiments of this application can also be further described in the following ways: A user feedback detection method, comprising: Based on the causal architecture, the preset language information is modeled to obtain a structural causal graph; the structural causal graph is represented by the following formula: ; in, This represents the semantic content factor in the causal architecture. This represents the semantic style factor in the causal architecture. This refers to the text information included in the language information. This refers to the tag information included in the language information; Based on a pre-defined large language model, the language information is content-focused to obtain a focused representation; Based on the focused representation and the structural causal graph, a causal decoupling model is obtained; The user feedback is detected based on the causal decoupling model, and the detection results are obtained.

[0101] Optionally, the language information is content-focused according to a preset large language model to obtain a focused representation, specifically including: The text information is learned from the context using the large language model to obtain the intermediate layer output. Calculate the sensitivity of each token in the intermediate layer output to the loss to obtain the token representation; The focus representation is determined based on the token representation and the intermediate layer output.

[0102] Optionally, the step of modeling based on the focused representation and the structural causal graph to obtain a causal decoupling model specifically includes: The latent variable information is obtained by mapping the structural causal graph according to the focused representation; The causal decoupling model is determined based on the latent variable information and the large language model.

[0103] Optionally, the step of mapping the structural causal graph according to the focused representation to obtain latent variable information specifically includes: The content factors are mapped according to the focus representation to obtain the content latent variables; The style factors are mapped according to the focus representation to obtain latent style variables; The content latent variables and style latent variables are optimized based on the preset task prediction loss to obtain the latent variable information.

[0104] Optionally, the latent variable information includes latent variables for optimized content and latent variables for optimized style; The causal decoupling model is determined based on the latent variable information and the large language model, specifically including: The optimized content latent variables are mapped according to the preset latent variable injection module to obtain the mapped content latent variables; The latent variable injection module maps the optimized style latent variable to obtain the mapped style latent variable. The latent variables of the mapping content and the latent variables of the mapping style are injected into the large language model according to the residual method to obtain the causal decoupling model.

[0105] A user feedback detection device, comprising: The causal graph construction module is used to model preset language information based on the causal architecture to obtain a structural causal graph; the structural causal graph is represented by the following formula: ; in, This represents the semantic content factor in the causal architecture. This represents the semantic style factor in the causal architecture. This refers to the text information included in the language information. This refers to the tag information included in the language information; The focusing module is used to focus the language information according to a preset large language model to obtain a focused representation. The modeling module is used to model based on the focused representation and the structural causal graph to obtain a causal decoupling model; The detection module is used to detect user feedback based on the causal decoupling model and obtain the detection results.

[0106] Optionally, the focusing module specifically includes: The representation module is used to perform context learning on the text information based on the large language model to obtain intermediate layer output; Calculate the sensitivity of each token in the intermediate layer output to the loss to obtain the token representation; The focus representation is determined based on the token representation and the intermediate layer output.

[0107] Optionally, the modeling module specifically includes: The mapping module is used to map the structural causal graph according to the focused representation to obtain latent variable information; The causal decoupling model is determined based on the latent variable information and the large language model.

[0108] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method described in any one of the preceding methods.

[0109] A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform any of the methods described above.

[0110] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples. Within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and many other variations exist regarding different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0111] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0112] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0113] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A user feedback detection method, comprising: Based on the causal architecture, the preset language information is modeled to obtain a structural causal graph; the structural causal graph is represented by the following formula: ； in, This represents the semantic content factor in the causal architecture. This represents the semantic style factor in the causal architecture. This refers to the text information included in the language information. This refers to the tag information included in the language information; Based on a pre-defined large language model, the language information is content-focused to obtain a focused representation; Based on the focused representation and the structural causal graph, a causal decoupling model is obtained; The user feedback is detected based on the causal decoupling model, and the detection results are obtained.

2. The user feedback detection method according to claim 1, wherein, Based on a pre-defined large language model, the language information is content-focused to obtain a focused representation, specifically including: The text information is learned from the context using the large language model to obtain the intermediate layer output. Calculate the sensitivity of each token in the intermediate layer output to the loss to obtain the token representation; The focus representation is determined based on the token representation and the intermediate layer output.

3. The user feedback detection method according to claim 1, wherein, The step of modeling based on the focused representation and the structural causal graph to obtain a causal decoupling model specifically includes: The latent variable information is obtained by mapping the structural causal graph according to the focused representation; The causal decoupling model is determined based on the latent variable information and the large language model.

4. The user feedback detection method according to claim 3, wherein, The step of mapping the structural causal graph based on the focused representation to obtain latent variable information specifically includes: The content factors are mapped according to the focus representation to obtain the content latent variables; The style factors are mapped according to the focus representation to obtain latent style variables; The content latent variables and style latent variables are optimized based on the preset task prediction loss to obtain the latent variable information.

5. The user feedback detection method according to claim 4, wherein, The latent variable information includes latent variables for optimizing content and latent variables for optimizing style; The causal decoupling model is determined based on the latent variable information and the large language model, specifically including: The optimized content latent variables are mapped according to the preset latent variable injection module to obtain the mapped content latent variables; The latent variable injection module maps the optimized style latent variable to obtain the mapped style latent variable. The latent variables of the mapping content and the latent variables of the mapping style are injected into the large language model according to the residual method to obtain the causal decoupling model.

6. A user feedback detection device, comprising: The causal graph construction module is used to model preset language information based on the causal architecture to obtain a structural causal graph; the structural causal graph is represented by the following formula: ； in, This represents the semantic content factor in the causal architecture. This represents the semantic style factor in the causal architecture. This refers to the text information included in the language information. This refers to the tag information included in the language information; The focusing module is used to focus the language information according to a preset large language model to obtain a focused representation. The modeling module is used to model based on the focused representation and the structural causal graph to obtain a causal decoupling model; The detection module is used to detect user feedback based on the causal decoupling model and obtain the detection results.

7. The user feedback detection device according to claim 6, wherein, The focusing module specifically includes: The representation module is used to perform context learning on the text information based on the large language model to obtain intermediate layer output; Calculate the sensitivity of each token in the intermediate layer output to the loss to obtain the token representation; The focus representation is determined based on the token representation and the intermediate layer output.

8. The user feedback detection device according to claim 6, wherein, The modeling module specifically includes: The mapping module is used to map the structural causal graph according to the focused representation to obtain latent variable information; The causal decoupling model is determined based on the latent variable information and the large language model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as claimed in any one of claims 1 to 5.

10. A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method of any one of claims 1 to 5.

Images