Method and system for determining voice reply strategy based on dynamic risk assessment

Abstract

The invention relates to the technical field of voice interaction, in particular to a method and system for determining a voice reply strategy based on dynamic risk assessment, and the method comprises the steps: obtaining user input and calling a pre-trained large language model to output a risk score in response to a voice interaction request initiated by a user; grading the risk scores based on a preset risk threshold, wherein the risk scores comprise high risks, medium risks and low risks; if the risk is high, immediately entering a strategic refusal response process; if the risk is medium risk, entering a multi-round clarification dialogue process, and continuously recalculating a risk score based on the constructed clarification question and the answer of the user until the risk score is not in a medium risk interval; and if the risk is low, directly entering a normal interaction process. According to the invention, a strategy which can perform fine-grained evaluation and dynamic management and control on a multi-level risk scene and can adaptively switch rejection and clarification in the same closed-loop process can be constructed in a voice dialogue system, so as to give consideration to the interactive experience of safety compliance and natural smoothness.

Description

Technical Field

[0001] The present application relates to the field of voice interaction technology, and in particular to a method and system for determining a voice response strategy based on dynamic risk assessment. Background Art

[0002] Currently, with the widespread application of large-scale pre-trained language models in voice dialogue systems, human-computer interaction is gradually shifting from rule-driven to intelligent, deep learning-based systems. Neural network-based dialogue systems can learn natural language expressions and contextual associations from massive amounts of data, enabling a more fluid interactive experience. However, open-domain dialogue also poses security and compliance risks. When user input contains sensitive, illegal, or injection-like attack intent, the system must effectively identify and prevent the leakage of risky information or misleading responses while ensuring question and answer quality and user experience. This has become a fundamental issue that needs to be addressed in current voice interaction technology.

[0003] Existing voice conversation security management often employs a dual-track mechanism: On the one hand, a risk detection module based on fixed thresholds or rules directly maps high-risk or out-of-bounds requests to pre-defined rejection template responses; on the other hand, when confidence in user intent is insufficient, it triggers single or multiple rounds of clarification and follow-up questions to obtain additional information. Typical examples include attack defense devices that use fixed template rejection and truncate responses when sensitive information leakage risks are detected; dual-model judgment based on functional intent, which directly executes templated rejection when the intent is identified as outside the business scope; and multi-round question-and-answer systems that initiate multiple rounds of clarification when intent is unclear, but do not perform real-time risk reassessment. While these solutions improve security or intent identification accuracy in their respective areas, they still suffer from fragmented decision-making, blunt rejection responses, and a disconnect between clarification and risk assessment.

[0004] Therefore, while existing technologies can rapidly intercept high-risk requests and gradually clarify ambiguous intent, they lack a unified policy switching mechanism based on a risk gradient. This prevents a smooth transition between strategic rejection and gradual clarification within a single framework, nor does it provide users with contextual modification guidance or dynamically adjust the defense depth based on supplemental information. The core technical challenge arising from this is how to build a system within a voice dialogue system that can both perform fine-grained assessment and dynamic control of multi-level risk scenarios, while also adaptively switching between rejection and clarification strategies within the same closed-loop process, ensuring a balanced experience of security compliance and a natural and smooth interaction experience. Summary of the Invention

[0005] This application provides a method and system for determining voice response strategies based on dynamic risk assessment. This system builds a system in a voice dialogue system that can not only perform fine-grained assessment and dynamic control of multi-level risk scenarios, but also adaptively switch between rejection and clarification strategies within the same closed-loop process, balancing safety and compliance with a natural and smooth interactive experience. This application provides the following technical solutions: In a first aspect, the present application provides a method for determining a voice response strategy based on dynamic risk assessment, the method comprising: In response to a voice interaction request initiated by a user, the system obtains user input and calls a pre-trained large language model to output a risk score. Classifying the risk scores based on a preset risk threshold, with the risk levels including high risk, medium risk, and low risk; If the risk level is high, immediately enter the strategic rejection response process; If the risk level is medium, a multi-round clarification dialogue process is initiated, where the risk score is continuously recalculated based on the constructed clarification questions and the user's answers until the risk score is no longer in the medium risk range. If the risk level is low, directly enter the normal interaction process.

[0006] In a specific implementation scheme, obtaining user input and calling a pre-trained large language model to output a risk score includes: The user's voice request is transcribed into text through the speech recognition module, and the pre-trained generative large language model is called to perform semantic analysis on the text input. Semantic features, contextual information and potential risk factors are extracted to comprehensively evaluate the semantic content of the user's request, and the risk score S is output in real time through the built-in risk calculation mechanism.

[0007] In a specific embodiment, the risk scores are divided into levels based on a preset risk threshold, and the risk levels include high risk, medium risk and low risk. The risk score S is compared with the pre-set double threshold and Compare in sequence: When When When .

[0008] In a specific implementation scheme, if the risk level is high risk, immediately entering the strategic rejection response process includes: Based on the text content entered by the user and its contextual dialogue information, the pre-trained generative large language model is called to perform in-depth semantic analysis of the request. During the analysis process, the specific feature information that causes high-risk judgments in the request is identified and extracted. Based on the understanding of the user's intention, the analysis clearly defines the specific reasons why the request is judged as high-risk. The large language model generates a rejection response text based on the current context and simultaneously outputs the specific reasons for the rejection.

[0009] In a specific implementation scheme, if the risk level is medium risk, a multi-round clarification dialogue process is initiated, and the risk score is continuously recalculated based on the constructed clarification questions and the user's answers until the risk score is no longer in the medium risk range. When the current risk score S is detected to be in the medium risk range, the large language model analyzes the semantic features of the current input to construct a vague and guiding clarification question. After the user provides supplementary information based on the clarification question, the pre-trained generative large language model is re-called to perform semantic understanding and feature extraction on the new input, and the risk score S is updated and recalculated. If the new risk score S rises to the high-risk range, the clarification process will be terminated immediately and the strategic rejection process will be initiated, generating a rejection response that includes risk descriptions and recommendations. If the new risk score S is reduced to the low-risk range and the user's intention can be clarified based on the current semantic features, the clarification process ends and the normal response process begins. If the new risk score is still in the medium risk range, the next round of clarification questions will be generated, the user will provide additional information, the score will be evaluated again, and the score recalculation process will be repeated until the score jumps out of the medium risk range.

[0010] In a specific implementation scheme, if the risk level is low risk, directly entering the normal interaction process includes: After outputting the risk score, if the score S is lower than the preset low risk threshold , the request is judged as a safe request. At this time, the request content is directly input into the response generation module, and the large language model is called to generate a standard response based on the current semantic content and context status.

[0011] In a specific embodiment, the risk scores are divided into levels based on a preset risk threshold, and the risk levels include high risk, medium risk and low risk, and further include: The double thresholds are set as high risk judgment threshold Low risk threshold , at any voice interaction round t, the risk interval difference variable is set as: ; After each round of interaction Make the following updates: ; in, 、 、 、 are the adjustment parameters for nonlinear regulation, score fluctuation weight, semantic fuzzy weight and symmetric offset feedback, respectively. Indicates the standard deviation of the risk scores output continuously in the most recent voice interaction window. Indicates the semantic ambiguity coefficient of the current speech input, represents the disturbance feedback term; After each round of execution of the update formula, a new risk buffer difference will be obtained , dynamically adjust the dual thresholds according to the following rules: ; .

[0012] In a second aspect, the present application provides a system for determining a voice response strategy based on dynamic risk assessment, which adopts the following technical solutions: A system for determining a voice response strategy based on dynamic risk assessment, comprising: The risk score output module is used to respond to the voice interaction request initiated by the user, obtain the user input and call the pre-trained large language model to output the risk score; A risk level classification module is used to classify the risk scores into levels based on a preset risk threshold, with the risk levels including high risk, medium risk and low risk; A high-risk determination module is used to immediately enter a strategic rejection response process if the risk level is high risk; A medium risk determination module is configured to, if the risk level is medium, enter a multi-round clarification dialogue process and continuously recalculate the risk score based on constructed clarification questions and the user's answers until the risk score is no longer in the medium risk range; The low risk determination module is used to directly enter the normal interaction process if the risk level is low risk.

[0013] In a third aspect, the present application provides an electronic device comprising a processor and a memory; a program is stored in the memory, and the program is loaded and executed by the processor to implement a method for determining a voice response strategy based on dynamic risk assessment as described in the first aspect.

[0014] In a fourth aspect, the present application provides a computer-readable storage medium, in which a program is stored. When the program is executed by a processor, it is used to implement a method for determining a voice response strategy based on dynamic risk assessment as described in the first aspect.

[0015] In summary, the beneficial effects of this application include at least: (1) Using a pre-trained large language model to accurately extract the semantic features of user requests, this approach deeply captures the potential risk characteristics of the input text, achieving refined and highly accurate risk scoring. Compared with traditional rule-based or simple model-based risk identification solutions, this solution can achieve a certain improvement in the accuracy of risk identification, significantly reducing the probability of misjudgment and missed judgment, and effectively solving the problem of rough risk grading in existing technologies.

[0016] (2) By designing a dual-threshold risk level assessment system, this system achieves a fine-grained classification of user requests (high, medium, and low risk) and dynamically adjusts the processing strategy in real time. When the dual-threshold dynamic strategy is adopted, the misjudgment rate of medium-risk requests can be improved to a certain extent, and the system can switch more flexibly between rejection and clarification, effectively solving the interaction rigidity and security vulnerabilities caused by traditional single-path decision-making.

[0017] (3) In medium-risk scenarios, the system uses a multi-round clarification dialogue and real-time risk reassessment linkage mechanism. By dynamically questioning the user's intentions and recalculating the risk score in real time, it ensures that the risk can be accurately judged after obtaining more supplementary information. After adopting the multi-round clarification and real-time risk reassessment mechanism, the accuracy of the system in security risk detection can be improved to a certain extent, greatly improving the consistency and security of the interactive process, and significantly solving the problem of the separation between the clarification link and the security assessment.

[0018] By introducing a multi-level risk classification mechanism, combining contextual semantic understanding with the risk scoring capabilities of a generative large language model, this system achieves refined risk identification, graded response execution, and strategic interaction management for voice input content within a unified process framework. This effectively addresses technical issues in existing voice systems, such as crude interaction security control, single response strategies, and fragmented user experience. Using the risk score S output by the large language model as a quantitative indicator, a set of static, pre-set dual-risk thresholds is set, allowing risk levels to be categorized into high, medium, and low risk levels, each corresponding to a different strategic response path. When the score falls into the high-risk range, the system immediately triggers a strategic rejection response process. Using the model to analyze user intent and semantic risk factors, the system outputs targeted, explanatory, and guiding rejection information, preventing invalid interactions or potentially violating content. When the score falls into the medium-risk range, the system initiates multiple rounds of clarification interactions, guiding users to gradually supplement or clarify their semantic expressions while maintaining contextual continuity. The risk score is reassessed after each round of interaction until a clear response path is determined. When the score falls into the low-risk range, the system directly outputs a standard voice response, maintaining a natural and efficient interaction. This approach enables the voice interaction system to flexibly switch response strategies when faced with different risk scenarios, ensuring real-time blocking of high-risk content while retaining room for optimization in guiding medium-risk semantics. This avoids the risk of misjudgment and interaction disruption associated with a one-size-fits-all approach. In particular, during medium-risk processing, multiple rounds of semantic clarification and large-scale model-assisted understanding enhance the system's ability to identify and resolve ambiguous expressions, semantically jumpy, or borderline requests, effectively improving the system's fault tolerance and adaptability.

[0019] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application and to implement it in accordance with the contents of the specification, the following is a detailed description of the preferred embodiments of the present application in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of a method for determining a voice response strategy based on dynamic risk assessment in an embodiment of the present application.

[0021] Figure 2 This is a schematic diagram of the overall process of the method for determining a voice response strategy based on dynamic risk assessment in an embodiment of the present application.

[0022] Figure 3 This is a structural block diagram of a system for determining a voice response strategy based on dynamic risk assessment in an embodiment of the present application.

[0023] Figure 4 This is a block diagram of an electronic device that determines a voice response strategy based on dynamic risk assessment in an embodiment of the present application. DETAILED DESCRIPTION

[0024] The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0025] Optionally, the present application uses the method for determining a voice response strategy based on dynamic risk assessment provided in each embodiment as an example for use in an electronic device, where the electronic device is a terminal or a server. The terminal can be a mobile phone, a computer, a tablet computer, etc. This embodiment does not limit the type of electronic device.

[0026] Reference Figure 1 , is a flow chart of a method for determining a voice response strategy based on dynamic risk assessment provided by an embodiment of the present application. The method includes at least the following steps: Step S101: In response to a voice interaction request initiated by a user, obtain user input and call a pre-trained large language model to output a risk score.

[0027] In step S101, in response to a user's voice interaction request, upon receiving the request, a pre-trained generative language model is invoked to perform semantic understanding and feature extraction on the input content, thereby completing a preliminary risk level assessment. The risk score S ranges from 0 to 1, with a higher value indicating a higher potential security risk of the request.

[0028] Specifically, the voice recognition module first transcribes the user's voice request into text, ensuring that the input content is easy to process. Subsequently, a pre-trained generative language model is used to perform semantic analysis on the text input, extracting semantic features, contextual information, and potential risk factors. This model, combining extensive linguistic knowledge with risk identification rules, comprehensively assesses the semantic content of the user's request and, through a built-in risk calculation mechanism, outputs a risk score S in real time. The score is strictly limited to the range of 0 to 1, quantifying the degree of security risk expressed in the input request.

[0029] It should be noted that the scoring calculation process is essentially within the scope of existing technology and has been widely used in current generative AI systems. Although each system may have different methods for extracting semantic features and setting weights for sensitive factors, the overall process follows the consistency principle of semantic understanding, feature extraction, and risk-weighted scoring.

[0030] Optionally, this application selects the Dongfeng Big Model developed by AISpeech as the big language model. The Dongfeng Big Model is based on the Transformer architecture and has been pre-trained on large-scale corpus. It has general semantic representation and natural language generation capabilities and is a mature existing technology model in this field. In implementation, the Dongfeng Big Model performs semantic understanding of user input and outputs corresponding semantic features. Subsequently, the risk calculation mechanism set in the model receives the semantic features and calculates them, thereby realizing the function of outputting the risk score S.

[0031] Step S102: Classify the risk scores based on a preset risk threshold, where the risk levels include high risk, medium risk, and low risk.

[0032] In step S102, the core is to use the risk score S output by the model according to the pre-set double threshold and Classification is performed, in which , thereby mapping the risk score into three levels of high risk, medium risk and low risk, which facilitates the refined diversion of subsequent interaction strategies.

[0033] Specifically, the risk score S is first compared with the two thresholds in sequence: When When When , it is judged as low risk. In the specific implementation, the score S and the threshold and The comparison judgment of is not performed in isolation, but is based on comprehensive factors such as the current interaction context, the semantic integrity of the user request, and historical scoring trends. In particular, when S is near the threshold boundary, the scoring result can be carefully judged. For example, if the score is slightly higher than , but historical requests show that the user frequently asks incomplete or vague questions, we can tend to keep the request in the medium risk range to trigger the necessary clarification process. However, user requests with strong contextual orientation and risk characteristics can trigger high-risk processes depending on the situation. This dynamic fine-tuning mechanism does not change the threshold itself, but rather fine-tunes the classification results through the joint judgment of multi-dimensional features in the case of marginal scores.

[0034] Step S103: If the risk level is high, immediately enter the strategic rejection response process.

[0035] In step S103, when the risk score S reaches or exceeds the preset high risk threshold When a request is identified as high risk, the strategic rejection response process is immediately initiated. At this stage, no further interactive clarification or risk reassessment is performed. Instead, the strategic rejection response process corresponding to high risk is immediately initiated to achieve immediate blocking and appropriate guidance for such high-risk requests.

[0036] Specifically, the rejection response process includes the following steps: First, based on the text content entered by the user and its contextual conversation information, a pre-trained generative large language model is invoked to perform in-depth semantic analysis of the request. During the analysis, specific features in the request that may trigger a high-risk determination are identified and extracted, such as whether the request involves content that violates laws and regulations, whether it contains keywords in sensitive areas, whether there are attempts at unauthorized access, and whether there are misleading or offensive expressions. Through this analysis, based on an understanding of user intent, the specific reasons why the request was determined to be high-risk are clearly defined.

[0037] After identifying risk factors, a large language model dynamically generates a context-appropriate rejection response based on the current context. This text not only clearly informs the user that their request has been rejected but also provides the specific justification for the rejection, such as "Your request involves access to sensitive areas and cannot be processed at this time" or "This type of issue falls within system limitations; we recommend adjusting your wording." Furthermore, to enhance user engagement and interactivity, the rejection response includes specific and actionable suggestions for revisions, such as guiding the user to rephrase their request, avoid sensitive language, or use more explicit language.

[0038] The strategic rejection responses generated based on this large language model ensure that high-risk requests are promptly blocked while providing valuable feedback to users. This reduces interaction security risks and prevents users from experiencing negative experiences due to rejection. Furthermore, the non-templated generation of rejection responses enhances the diversity and contextual relevance of response content, significantly improving the system's flexibility and expertise in handling complex semantic request scenarios.

[0039] Step S104: If the risk level is medium risk, enter a multi-round clarification dialogue process, and continuously recalculate the risk score based on the constructed clarification questions and the user's answers until the risk score is no longer in the medium risk range.

[0040] In step S104, if the risk score S is in the medium risk range, a rejection or direct response is not immediately generated. Instead, the process begins with multiple rounds of clarification interactions and dynamic risk reassessment. This process aims to gradually guide users to clarify their intentions for requests that are semantically ambiguous, incomplete, or contain potential sensitivity factors but do not yet pose a high risk. Based on this, the risk score is updated in real time and the decision branch is selected.

[0041] Specifically, when the current risk score S is detected to be in the medium risk range, a vague but guiding clarification question is first constructed as a response to the current user's request. This clarification question is generated based on the semantic features of the current input using a large language model. Its design aims to guide the user to provide more explicit and specific background information or behavioral intentions, thereby enabling the system to obtain more comprehensive context for risk reassessment. After the user provides additional information based on the clarification question, the pre-trained generative large language model is re-invoked to perform semantic understanding and feature extraction on the new input, updating and recalculating the risk score S. If the new risk score S rises to the high risk range, the clarification process is immediately terminated and the strategic rejection process is initiated, generating a rejection response containing risk explanations and recommendations. If the new risk score S falls to the low risk range and the current semantic features clearly indicate the user's intention, the clarification process ends and the normal response process begins. If the new risk score S remains in the medium risk range, the next round of clarification questions is generated, and the user provides additional information, followed by a reassessment of the score. This recalculation process repeats until the score moves out of the medium risk range.

[0042] In practice, this multi-round clarification and risk reassessment process maintains state and context continuity. Each round of clarification is modeled synthetically based on the previous semantic context and the latest input, ensuring that the model considers the continuity and evolution of the conversation history in its judgments, avoiding semantic misjudgments or score fluctuations. The construction and implementation of this multi-round clarification dialogue mechanism avoids blanket rejections of ambiguous or slightly sensitive requests, technically reducing the probability of false rejections and ensuring the fault tolerance and flexibility of the voice interaction system. Furthermore, by guiding users to gradually refine their input and implementing real-time risk score updates, the system dynamically adjusts risk assessment results, achieving refined risk handling and a graded response. Furthermore, this process enhances user engagement and feedback interpretability, allowing users to make corrections based on their understanding of the issues with their requests, thereby enhancing the collaborative and adaptive nature of human-computer interaction.

[0043] Step S105: If the risk level is low, directly enter the normal interaction process.

[0044] In step S105, when the risk score S is in the low risk range, it means that the current user request does not contain any identifiable security risks, sensitive information or compliance risks. There is no need to execute any rejection or clarification strategy based on security control. It directly enters the standard voice interaction process and outputs the regular response content.

[0045] Specifically, after the initial risk score is output, if the score S is lower than the low risk threshold preset by the system , the request is judged as a safe request. At this point, the subsequent clarification process or risk review mechanism is no longer triggered. Instead, the request content is directly input into the response generation module, and the large language model is called to generate a standard response based on the current semantic content and contextual state. The response content does not contain risk prompts, rejection guidance, or vague rhetorical questions. The service response fully follows the user's intention to ensure the naturalness and integrity of the interactive experience. By executing this step, when it is confirmed that the request risk is low, the processing path can be simplified, the response rhythm can be accelerated, the service efficiency can be improved, and the user experience can be avoided from being affected by excessive intervention. At the same time, the clear low-risk judgment boundary also ensures that the system only releases the response under the premise of no risk, building a safe and controllable voice interaction closed loop.

[0046] In addition, as a preferred embodiment, in order to improve the accuracy of risk level judgment and policy response flexibility in complex semantic interaction processes, the present application further proposes a risk buffer zone update mechanism based on dynamic adjustment of threshold differences. This mechanism introduces structurally innovative mathematical formulas to achieve adaptive updates of risk division threshold intervals, effectively enhancing the system's ability to respond to dynamic context changes.

[0047] Specifically, based on step S102, the dual thresholds on which the system risk level determination depends are respectively the high risk determination threshold and the Low risk threshold , at any voice interaction round t, the risk interval difference variable is set as: ; This difference is used to determine the width of the medium risk interval, thus affecting the system's tolerance and handling strategy for borderline risk requests. In order to enable the risk buffer to be dynamically adjusted over time, the system Make the following updates: ; in, 、 、 、 They are the adjustment parameters for nonlinear regulation, score fluctuation weight, semantic fuzzy weight and symmetric offset feedback, which can be set according to risk control level, training data or business strategy. Indicates the standard deviation of the risk scores output continuously in the most recent voice interaction window, used to characterize the volatility of user intent in the current interaction. This parameter can be calculated by counting the recent scores over a sliding window. Indicates the semantic ambiguity coefficient of the current speech input. This value is obtained by extracting features such as syntactic completeness, semantic certainty, and ambiguity from the user's text input using a large language model and mapping them through a preset semantic ambiguity function. The range is [0, 1]. Higher values ​​indicate more ambiguous expressions. The perturbation feedback term is calculated based on the degree of semantic intent shift between the previous and next rounds (e.g., keyword shift, context span increase, etc.). It is used to compensate for scoring instability caused by input jumps. A symmetric tension factor representing the threshold structure. This term is introduced to fine-tune the balance of the risk interval when the high and low risk thresholds are not symmetrically distributed around the center of the risk score interval.

[0048] After executing the update formula each round, a new risk buffer difference will be obtained , and then dynamically adjust the dual thresholds according to the following rules: ; ; In the design of the above formula, a dynamically adjusted risk buffer width variable is introduced and dynamically updated on a round-by-round basis through a formula that combines nonlinear mapping with multi-factor fusion. By introducing a differential variable and endowing it with the ability to dynamically change with interaction, the system can adjust the width of the medium-risk interval in real time based on the current semantic environment, user behavior volatility, and historical rating changes, thereby adapting to risk perception sensitivity in different contexts. When the interaction is stable and the semantics are clear, the medium-risk interval is narrowed, making the system respond more decisively. When the interaction is unstable or the user's expression is ambiguous, the medium-risk interval is appropriately widened, and the system enters a multi-round clarification process, improving fault tolerance.

[0049] In traditional risk assessment interactive systems, risk thresholds are often statically set and applicable to fixed scenarios. However, in real-world voice interaction applications, user behavior patterns, semantic features, and risk contexts are dynamic, making fixed thresholds difficult to consistently ensure accurate and flexible system responses. To address this issue, this solution introduces a dynamic dual-threshold update mechanism based on contextual state, interaction behavior, and system operation data. Risk tolerance varies across different application scenarios. For example, the risk assessment criteria for medical voice assistants and entertainment question-and-answer systems differ significantly. By dynamically updating the dual thresholds, the system automatically adapts to current application needs based on historical interaction characteristics, enabling fine-grained policy switching. Furthermore, real-time updates of risk thresholds help the system capture fine-tuning signals in user behavior patterns and request content distribution over the long term, maintaining high responsiveness to gray-zone risk scenarios and avoiding misjudgments (false rejections or false responses) caused by static classification.

[0050] In summary, combined with Figure 2This application introduces a multi-level risk classification mechanism, combining contextual semantic understanding with the risk scoring capabilities of a generative large language model. This allows for refined risk identification, graded response execution, and strategic interaction management of voice input content within a unified process framework, effectively addressing technical issues in existing voice systems, such as crude interaction security control, single response strategies, and fragmented user experience. Using the risk score S output by the large language model as a quantitative indicator, a set of static, preset dual-risk thresholds is set, allowing risk levels to be divided into high, medium, and low risk levels, each corresponding to a different strategic path. When the score falls into the high-risk range, the system immediately triggers a strategic rejection response process. By analyzing user intent and semantic risk factors through the model, the system outputs targeted, explanatory, and guiding rejection information to avoid invalid interactions or the generation of potentially illegal content. When the score falls into the medium-risk range, the system initiates multiple rounds of clarification interactions, guiding users to gradually supplement or clarify semantic expressions while maintaining contextual continuity. The risk score is reassessed after each round of interaction until a clear response path is determined. When the score falls into the low-risk range, the system directly outputs regular voice responses, maintaining a natural and efficient interaction. This approach enables the voice interaction system to flexibly switch response strategies when faced with different risk scenarios, ensuring real-time blocking of high-risk content while retaining room for optimization in guiding medium-risk semantics. This avoids the risk of misjudgment and interaction disruption associated with a one-size-fits-all approach. In particular, during medium-risk processing, multiple rounds of semantic clarification and large-scale model-assisted understanding enhance the system's ability to identify and resolve ambiguous expressions, semantically jumpy, or borderline requests, effectively improving the system's fault tolerance and adaptability.

[0051] Figure 3 This is a structural block diagram of a system for determining a voice response strategy based on dynamic risk assessment provided by an embodiment of the present application. The system includes at least the following modules: The risk score output module is used to respond to the voice interaction request initiated by the user, obtain the user input and call the pre-trained large language model to output the risk score; The risk classification module is used to classify risk scores into high risk, medium risk and low risk levels based on preset risk thresholds; High-risk determination module, used to immediately enter the strategic rejection response process if the risk level is high; The medium risk assessment module is used to enter a multi-round clarification dialogue process if the risk level is medium risk. The risk score is continuously recalculated based on the constructed clarification questions and the user's answers until the risk score is no longer in the medium risk range. The low-risk judgment module is used to directly enter the normal interaction process if the risk level is low.

[0052] For relevant details, please refer to the above method embodiment.

[0053] Figure 4 4 is a block diagram of an electronic device provided in one embodiment of the present application. The device includes at least a processor 401 and a memory 402.

[0054] Processor 401 may include one or more processing cores, such as a quad-core processor or an octa-core processor. Processor 401 may be implemented in hardware using at least one of the following: a DSP (Digital Signal Processing), an FPGA (Field-Programmable Gate Array), or a PLA (Programmable Logic Array). Processor 401 may also include a main processor and a coprocessor. The main processor is used to process data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 401 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing content displayed on the display screen. In some embodiments, processor 401 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0055] The memory 402 may include one or more computer-readable storage media, which may be non-transitory. The memory 402 may also include a high-speed random access memory and a non-volatile memory, such as one or more disk storage devices and flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 402 is used to store at least one instruction, which is used to be executed by the processor 401 to implement the method for determining a voice reply strategy based on dynamic risk assessment provided in the method embodiment of the present application.

[0056] In some embodiments, the electronic device may optionally include a peripheral device interface and at least one peripheral device. The processor 401, memory 402, and peripheral device interface may be connected via a bus or signal lines. Each peripheral device may be connected to the peripheral device interface via a bus, signal lines, or circuit boards. Illustratively, the peripheral devices include, but are not limited to, radio frequency circuitry, a touchscreen display, audio circuitry, and a power supply.

[0057] Of course, the electronic device may also include fewer or more components, which is not limited in this embodiment.

[0058] Optionally, the present application also provides a computer-readable storage medium, in which a program is stored. The program is loaded and executed by a processor to implement the method of determining a voice response strategy based on dynamic risk assessment in the above method embodiment.

[0059] Optionally, the present application also provides a computer product, which includes a computer-readable storage medium, in which a program is stored, and the program is loaded and executed by a processor to implement the method of determining a voice response strategy based on dynamic risk assessment in the above-mentioned method embodiment.

[0060] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0061] The above embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for determining a voice response strategy based on dynamic risk assessment, characterized in that: The method comprises: In response to a voice interaction request initiated by a user, the system obtains user input and calls a pre-trained large language model to output a risk score. Classifying the risk scores based on a preset risk threshold, with the risk levels including high risk, medium risk, and low risk; If the risk level is high, immediately enter the strategic rejection response process; If the risk level is medium, a multi-round clarification dialogue process is initiated, where the risk score is continuously recalculated based on the constructed clarification questions and the user's answers until the risk score is no longer in the medium risk range. If the risk level is low, directly enter the normal interaction process.

2. The method for determining a voice response strategy based on dynamic risk assessment according to claim 1, characterized in that: Obtaining user input and calling a pre-trained large language model to output a risk score includes: The user's voice request is transcribed into text through the speech recognition module, and the pre-trained generative large language model is called to perform semantic analysis on the text input. Semantic features, contextual information and potential risk factors are extracted to comprehensively evaluate the semantic content of the user's request, and the risk score S is output in real time through the built-in risk calculation mechanism.

3. The method for determining a voice response strategy based on dynamic risk assessment according to claim 2, characterized in that: The risk scores are divided into levels based on the preset risk threshold, and the risk levels include high risk, medium risk and low risk. The risk score S is compared with the pre-set double threshold and Compare in sequence: When When When .

4. The method for determining a voice response strategy based on dynamic risk assessment according to claim 1, characterized in that: If the risk level is high, immediately entering the strategic rejection response process includes: Based on the text content entered by the user and its contextual dialogue information, the pre-trained generative large language model is called to perform in-depth semantic analysis of the request. During the analysis process, the specific feature information that causes high-risk judgments in the request is identified and extracted. Based on the understanding of the user's intention, the analysis clearly defines the specific reasons why the request is judged as high-risk. The large language model generates a rejection response text based on the current context and simultaneously outputs the specific reasons for the rejection.

5. The method for determining a voice response strategy based on dynamic risk assessment according to claim 3, characterized in that: If the risk level is medium risk, a multi-round clarification dialogue process is entered, and the risk score is continuously recalculated based on the constructed clarification questions and the user's answers until the risk score is no longer in the medium risk range. When the current risk score S is detected to be in the medium risk range, the large language model analyzes the semantic features of the current input to construct a vague and guiding clarification question. After the user provides supplementary information based on the clarification question, the pre-trained generative large language model is re-called to perform semantic understanding and feature extraction on the new input, and the risk score S is updated and recalculated. If the new risk score S rises to the high-risk range, the clarification process will be terminated immediately and the strategic rejection process will be initiated, generating a rejection response that includes risk descriptions and recommendations. If the new risk score S is reduced to the low-risk range and the user's intention can be clarified based on the current semantic features, the clarification process ends and the normal response process begins. If the new risk score is still in the medium risk range, the next round of clarification questions will be generated, the user will provide additional information, the score will be evaluated again, and the score recalculation process will be repeated until the score jumps out of the medium risk range.

6. The method for determining a voice response strategy based on dynamic risk assessment according to claim 3, characterized in that: If the risk level is low risk, directly entering the normal interaction process includes: After outputting the risk score, if the score S is lower than the preset low risk threshold , the request is judged as a safe request. At this time, the request content is directly input into the response generation module, and the large language model is called to generate a standard response based on the current semantic content and context status.

7. The method for determining a voice response strategy based on dynamic risk assessment according to claim 3, characterized in that: The risk scores are divided into levels based on the preset risk threshold, and the risk levels include high risk, medium risk and low risk. The double thresholds are set as high risk judgment threshold Low risk threshold , at any voice interaction round t, the risk interval difference variable is set as: ； After each round of interaction Make the following updates: ； in, 、 、 、 are the adjustment parameters for nonlinear regulation, score fluctuation weight, semantic fuzzy weight and symmetric offset feedback, respectively. Indicates the standard deviation of the risk scores output continuously in the most recent voice interaction window. Indicates the semantic ambiguity coefficient of the current speech input, represents the disturbance feedback term; After each round of execution of the update formula, a new risk buffer difference will be obtained , dynamically adjust the dual thresholds according to the following rules: ; 。 8. A system for determining voice response strategies based on dynamic risk assessment, characterized in that: include: The risk score output module is used to respond to the voice interaction request initiated by the user, obtain the user input and call the pre-trained large language model to output the risk score; A risk level classification module is used to classify the risk scores into levels based on a preset risk threshold, with the risk levels including high risk, medium risk and low risk; A high-risk determination module is used to immediately enter a strategic rejection response process if the risk level is high risk; A medium risk determination module is configured to, if the risk level is medium, enter a multi-round clarification dialogue process and continuously recalculate the risk score based on constructed clarification questions and the user's answers until the risk score is no longer in the medium risk range; The low risk determination module is used to directly enter the normal interaction process if the risk level is low risk.

9. An electronic device, characterized in that: The device includes a processor and a memory; a program is stored in the memory, and the program is loaded and executed by the processor to implement a method for determining a voice response strategy based on dynamic risk assessment as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The storage medium stores a program, which, when executed by a processor, is used to implement a method for determining a voice response strategy based on dynamic risk assessment as described in any one of claims 1 to 7.

Images