A method and system for voice dialogue response with slot constraints and business state machine

Abstract

The application discloses a voice dialogue response method and system of a business state machine and slot constraint, comprising: receiving a user natural voice, generating a candidate transcription text, a word-level transcription confidence and a semantic candidate result; determining a current state node in the business state machine, loading a slot constraint set, an allowed state transition set, a prohibited state transition set and an allowed response action set; updating a slot object according to the semantic candidate result and historical round information; generating a candidate response action; performing state transition legality verification, slot integrity verification, knowledge evidence consistency verification and risk compliance verification on the candidate response action; outputting a voice response action when the verification is passed, generating an abnormal supplementary question action when the verification is not passed, and writing state migration, slot change, evidence retrieval, response action and verification result into an audit record. The application can improve the process controllability, information integrity and response compliance of a task-oriented voice intelligent agent.

Description

Technical Field

[0001] This application relates to the field of intelligent voice dialogue and task-oriented intelligent agent response control technology, and more specifically, to a voice dialogue response method with business state machine and slot constraints. Background Technology

[0002] With the increasing applications of enterprise intelligent agents, voice customer service, voice office assistants, government affairs consultation robots, and property service voice agents, voice dialogue systems have gradually moved from single-turn question-and-answer to task-oriented, multi-turn, and executable business processing stages. Compared with text-based intelligent agents, voice-based intelligent agents need to achieve low-latency collaboration between real-time transcription, semantic understanding, state tracking, slot completion, knowledge retrieval, response generation, and voice broadcasting, and generate executable, interpretable, and auditable response actions under business process constraints.

[0003] In scenarios such as financial account opening, after-sales repair requests, government affairs consultations, and property services, users' voice expressions often exhibit characteristics such as colloquialism, omissions, repetitions, interruptions, error corrections, interjections, noise transcription errors, and unstable information order. Relying solely on large language models to directly generate responses can easily lead to problems such as uncontrolled state transitions, missing slot information, business node overstepping, conflicts between response content and the knowledge base, and unreasonable timing of follow-up questions.

[0004] Specifically, existing technologies have the following defects and shortcomings: Most task-oriented dialogue systems simply fill in the slots according to intents, lacking explicit verifiable business state machine constraints. This may lead to entering the next business stage when the data is incomplete, causing process skipping or premature triggering of business actions. When large language models directly generate voice responses, they lack joint constraints on state, slots, knowledge sources, and action types, which may result in responses that do not match the current business node or answers that are not supported by retrieved evidence. Traditional supplementary question strategies are usually based on slot missing triggers, without distinguishing between mandatory slots, conditional slots, derived slots, and mutually exclusive slots, leading to excessive supplementary questions or omissions of key supplementary questions, affecting task completion rate and dialogue efficiency. Voice scenarios may encounter situations such as speech recognition errors, user interruptions, and state rollbacks. Existing systems lack anomaly recovery mechanisms at the state machine level, which may lead to old slots being incorrectly retained, new slots failing to cover, or states being unable to roll back. Process-oriented voice agents usually lack pre-response verification and post-response recording mechanisms, making it difficult to prove whether a response conforms to process, permission, knowledge, and slot constraints, and making auditing and review difficult.

[0005] Therefore, there is a need for a response control method that can uniformly model user speech transcription results, business process status, slot constraints, knowledge retrieval evidence, and abnormal questioning strategies during multi-turn natural speech dialogues. This method would enable the speech agent to generate speech response actions while meeting the business state machine constraints and slot integrity requirements, and to perform controllable questioning, state rollback, and audit recording in abnormal situations. Summary of the Invention

[0006] The purpose of this application is to provide a voice dialogue response proxy method based on business state machine and slot constraints. By combining business state machine constraints, slot constraints, knowledge evidence verification, and abnormal question recovery, the method improves the process controllability, information integrity, and response compliance of task-oriented voice intelligent agents.

[0007] In a first aspect, embodiments of this application provide a voice dialogue response method based on a service state machine and slot constraints, comprising the following steps: Receive user's natural speech, perform speech recognition processing on the natural speech, and obtain candidate transcribed text and word-level transcription confidence; The candidate transcribed texts are subjected to intent recognition, slot extraction, and confidence labeling to obtain semantic candidate results; The current state node in the business state machine is determined based on the current session record. The state node includes state type, allowed intent set, required slot set, condition slot set, prohibited state transition set, allowed response action set, knowledge evidence policy, and anomaly recovery policy. Based on the current state node, load the slot constraint set, the allowed state transition set, the prohibited state transition set, and the allowed response action set; Based on the candidate transcribed text, the semantic candidate results, the word-level transcription confidence, and the historical round information, the slot object is updated according to the confidence threshold, source type, coverage rule, and state adaptation rule in the slot constraint set. Candidate response actions are generated based on the current state node, the updated slot object, and the knowledge retrieval results. The knowledge retrieval results are obtained based on the current state node, the user intent in the semantic candidate results, and the confirmed slot retrieval knowledge evidence. The candidate response action is subjected to state transition legality verification, slot integrity verification, knowledge evidence consistency verification, response action type legality verification, and risk compliance verification, respectively, to obtain state transition legality score, slot integrity score, knowledge evidence consistency score, response action type legality score, and risk compliance score. The above scores are weighted and summed according to preset weight coefficients to obtain the comprehensive verification score of the candidate response action. When the overall verification score is higher than the preset response threshold and there are no hard prohibitions, the candidate response action is output as a voice response action; when the overall verification score is not higher than the response threshold or there are hard prohibitions, an abnormal supplementary question action is generated. Write state transitions, slot changes, evidence retrieval, response actions, and verification results into the audit log; The hard prohibitions include at least one of the following: the target state belongs to the prohibited state transition set, the required slot is missing, the response is not bound to valid knowledge evidence, and the response action type is prohibited by the current state.

[0008] Preferably, the state types include initial state, information collection state, confirmation state, execution state, termination state, and abnormal state; the knowledge evidence strategy is used to limit the knowledge evidence required to generate explanatory, policy, commitment, or guidance responses; the abnormal recovery strategy is used to limit the triggering conditions for low-confidence clarification, slot conflict confirmation, state rollback, and transfer to manual processing.

[0009] Preferably, the slot constraint set includes constraints corresponding to each of the following: mandatory slots, conditional slots, mutually exclusive slots, derived slots, and sensitive slots; wherein, the mandatory slots are slots that must be collected and pass format validation before the current state node is advanced; the conditional slots are slots that are required to be filled due to business type, user identity, or the value of a preceding slot; the mutually exclusive slots are slots that cannot be true simultaneously; the derived slots are slots derived from confirmed slots according to rules or models; and the sensitive slots are slots involving identity, account, assets, or personal information.

[0010] Preferably, the state transition legality verification is used to determine whether the target state corresponding to the candidate response action belongs to the allowed state transition set and does not belong to the prohibited state transition set; the knowledge evidence consistency verification is used to determine whether the explanatory, policy, commitment, or guidance responses are bound to valid knowledge evidence that matches the current state node and the user question.

[0011] Preferably, the action of generating anomaly supplementary inquiries includes: for each slot to be supplemented or matter to be clarified, calculating the supplementary inquiry priority of the slot to be supplemented or matter to be clarified according to the following formula based on the business importance of the slot to be supplemented or matter to be clarified, the necessity for status advancement, the current confidence level, the risk level, and the dependence on the execution of subsequent slots or tools: ; in, Indicates the first Prioritize questions regarding pending slots or clarifications. Indicates the importance of the business. This indicates the necessity of advancing the state. Indicates the current confidence level. Indicates the risk level. Indicates the degree of dependence. , , , and These are weighting coefficients; the highest priority supplementary question is selected from the slots or matters to be supplemented or clarified that are suitable for the current dialogue state to generate a supplementary question action; wherein, suitable for the current dialogue state means that the supplementary question action generated for the slot or matter to be supplemented or clarified belongs to the supplementary question action type allowed by the set of allowed response actions of the current state node, and the content of the supplementary question action does not conflict with the confirmed slot; when the slot or matter to be supplemented or clarified with the highest priority supplementary question is not suitable for the current dialogue state, the second highest priority supplementary question is selected from the slots or matters to be supplemented or clarified that are suitable for the current dialogue state to generate a supplementary question action.

[0012] Preferably, when a user interrupts, corrects, cancels, or negates an expression, the current response action is paused, the user's intent and slot change are reassessed, and a status rollback, slot overwrite confirmation, or manual processing is performed.

[0013] Preferably, the updated slot object includes: when the newly extracted slot value is inconsistent with the historically confirmed slot value, generating a confirmation question and retaining the old value or adopting the new value according to the user's response; when the word-level transcription confidence is lower than the confidence threshold, generating a restatement confirmation or requesting the user to restate key information.

[0014] Preferably, the weighting coefficients of the comprehensive verification score and the weighting coefficients of the supplementary question priority are pre-configured according to the business scenario, or dynamically adjusted according to historical quality inspection data.

[0015] Preferably, the audit record includes voice round identifier, state transition trajectory, slot change content, retrieval evidence identifier, response action content, and verification result details.

[0016] Secondly, embodiments of this application provide a voice dialogue response system with service state machine and slot constraints, including: The voice input and transcription module is used to receive the user's natural voice, perform speech recognition processing on the natural voice, and obtain candidate transcribed text and word-level transcription confidence. The intent and slot recognition module is used to perform intent recognition, slot extraction and confidence labeling on the candidate transcribed text to obtain semantic candidate results. The business state machine module is used to determine the current state node in the business state machine based on the current session record. The state node includes state type, allowed intent set, required slot set, condition slot set, prohibited state transition set, allowed response action set, knowledge evidence strategy, and exception recovery strategy. The slot constraint module is used to load a slot constraint set, an allowed state transition set, a prohibited state transition set, and an allowed response action set based on the current state node; The slot object update module is used to update the slot object according to the confidence threshold, source type, coverage rule and state adaptation rule in the slot constraint set based on the candidate transcribed text, the semantic candidate result, the word-level transcription confidence and historical round information; The response action generation module is used to obtain knowledge retrieval results based on the current state node, the user intent in the semantic candidate results, and the confirmed slot retrieval knowledge evidence, and to generate candidate response actions based on the current state node, the updated slot object, and the knowledge retrieval results. The response verification module is used to perform state transition legality verification, slot integrity verification, knowledge evidence consistency verification, response action type legality verification, and risk compliance verification on the candidate response action, and obtain state transition legality score, slot integrity score, knowledge evidence consistency score, response action type legality score, and risk compliance score respectively. The above scores are weighted and summed according to preset weight coefficients to obtain the comprehensive verification score of the candidate response action. The response output and abnormal question completion module is used to output the candidate response action as a voice response action when the comprehensive verification score is higher than the preset response threshold and there are no hard prohibitions; and to generate an abnormal question completion action when the comprehensive verification score is not higher than the response threshold or there are hard prohibitions. The audit log module is used to write status transitions, slot changes, evidence retrieval, response actions, and verification results into the audit log. The hard prohibitions include at least one of the following: the target state belongs to the prohibited state transition set, the required slot is missing, the response is not bound to valid knowledge evidence, and the response action type is prohibited by the current state.

[0017] Compared with the solutions of the prior art, the present invention has at least the following beneficial technical effects: (1) By abstracting the business process into a computable and verifiable business state machine object containing state node identifiers, state types, allowed intent sets, required slot sets, condition slot sets, prohibited state transition sets, allowed response action sets, knowledge evidence strategies and anomaly recovery strategies, explicit constraints are imposed on state transitions and action types before voice response generation, avoiding voice agents from overstepping levels, making advance commitments or skipping confirmation steps, thus ensuring the strict execution of business process nodes from a mechanism perspective. (2) By setting five types of slot constraints, namely, mandatory slots, conditional slots, mutually exclusive slots, derived slots and sensitive slots, and adding slot value format, confidence threshold, source type, coverage rule and status adaptation rule, the completeness of information collection is improved, the invalid work orders or erroneous business submissions caused by missing information are reduced, and logical contradictions such as mutually exclusive slots being valid at the same time are avoided. (3) By conducting joint verification of five dimensions of response actions before voice broadcast, namely the legality of state transition, slot integrity, consistency of knowledge evidence, legality of response action type and risk compliance, and adopting a judgment method that combines weighted comprehensive verification score with hard prohibition items, the probability of hallucinatory answers and process errors is reduced, and it is ensured that all explanatory, policy, commitment and guidance responses are supported by effective knowledge evidence. (4) The priority of supplementary questions is calculated based on five factors: importance of slot business, necessity for state advancement, current confidence level, risk level and dependence. Supplementary questions are generated by selecting the supplementary items with the highest priority and suitable for the current dialogue state, which reduces invalid follow-up questions and duplicate confirmations, and improves task completion rate and dialogue efficiency. (5) By setting targeted supplementary questions, clarifications, rollbacks and recovery strategies for low confidence in speech recognition, user interruption, error correction, cancellation, negation, slot conflict and illegal jump, such as confirming to retain the old value or adopt the new value when there is a slot conflict, and pausing the current response and re-evaluating the intent when the user interrupts, the robustness of multi-turn speech dialogue in noisy and colloquial environments is improved. (6) By recording voice rounds, state transitions, slot changes, evidence retrieval, response actions and verification results, a complete dialogue audit log is formed, which facilitates quality inspection, review, training and quality evaluation, and allows each response action to be traced back to see if it complies with process, authority, knowledge and slot constraints; In summary, this application organically combines key technologies such as process constraints based on a computable business state machine, integrity constraints of five types of slots, five-dimensional joint verification and hard prohibition item determination, precise supplementary questioning based on five-factor priority, and anomaly recovery and full-process auditing to form a closed-loop response control system for task-oriented natural voice dialogue. In various business scenarios such as financial account opening, after-sales repair, government consultation, and property services, it can significantly improve the process controllability, information integrity, anomaly recovery capability, and response compliance of the voice intelligent agent, while taking into account both business execution efficiency and compliance audit requirements. Attached Figure Description

[0018] Exemplary embodiments of the present invention can be more fully understood by referring to the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain the present invention and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 A flowchart illustrating a voice dialogue response method based on a business state machine and slot constraints provided in an embodiment of this application; Figure 2 A schematic diagram of a financial account opening business state machine provided for an embodiment of this application; Figure 3 A schematic diagram illustrating a slot constraint and replacement process provided in an embodiment of this application; Figure 4 This application provides a schematic diagram of the structure of a voice dialogue response system with service state machine and slot constraints, as shown in the embodiments of this application. Figure 5 A schematic diagram of an electronic device provided for an exemplary embodiment of this application; Figure 6 This is a schematic diagram of a computer-readable medium provided for an exemplary embodiment of this application. Detailed Implementation

[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0021] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0022] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] It should be noted that, in the following embodiments, the voice dialogue response refers to an intelligent agent that interacts with the user in multiple rounds via voice and performs business process processing actions in a task-oriented business scenario; the business state machine refers to a state machine object that formalizes the business process into a computable and verifiable state machine object. Each state node includes not only the business name, but also allowed input types, transition states, slot constraints, response action constraints, knowledge evidence constraints, and exception handling strategies; and the slot refers to the information field related to business processing extracted from the user's expression.

[0024] This embodiment uses a financial account opening voice proxy scenario as an example to provide a detailed description of the voice dialogue response proxy method of this application. In this scenario, the user initiates a business account opening inquiry via telephone voice. The voice dialogue response is deployed on the financial institution's intelligent voice customer service platform and is connected to the business state machine rule base, slot constraint rule base, and knowledge base.

[0025] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a voice dialogue response method based on a service state machine and slot constraints provided in an embodiment of this application. The method in this embodiment is executed by a voice dialogue response processor and includes the following steps: S110, Receive user's natural speech, perform speech recognition processing on the natural speech, and obtain candidate transcribed text and word-level transcription confidence; Specifically, the user says "I want to open a business account" over the phone. The voice input and transcription processing segment and decodes the real-time voice stream to obtain the candidate transcribed text "I want to open a business account", the word-level transcription confidence of each word, and time segment information. The transcription result is not directly used as the final semantics, but enters the state constraint and slot verification process.

[0026] In an alternative implementation, the user's voice can also be recorded offline in its entirety, and the speech recognition processing can perform batch decoding of the entire recording. When speech recognition is unavailable or the recognition confidence is too low, the voice dialogue response agent can also receive text input by the user through the text channel and directly use the text as a candidate transcribed text to enter the subsequent processing flow, so as to ensure the continuity and availability of the dialogue flow.

[0027] S120, perform intent recognition, slot extraction and confidence labeling on the candidate transcribed text to obtain semantic candidate results; Specifically, the intent and slot recognition processing performs semantic parsing on the candidate transcribed text to obtain the intent recognition result (business account opening consultation), the slot candidate set (e.g., subject type, business name, processing method, etc.), and the confidence label of each slot candidate. At the same time, it identifies whether the user's expression contains error correction expression, negative expression, cancellation expression, or supplementary expression, and outputs the corresponding error correction signal or negative signal for subsequent anomaly recovery.

[0028] S130, determine the current state node in the service state machine based on the current session record. The state node includes state type, allowed intent set, required slot set, condition slot set, prohibited state transition set, allowed response action set, knowledge evidence strategy, and exception recovery strategy. Specifically, the voice dialogue response agent determines the current state node in the business state machine based on the dialogue state identifier in the current session record. To avoid this application being misunderstood as mere business process automation, this application formalizes the business process as a computable and verifiable business state machine object. Please refer to... Figure 2 , Figure 2This is a schematic diagram of a financial account opening business state machine provided in an embodiment of this application. In this embodiment, the user is making their first call, and the current status node is "Account Opening Type Confirmation". This status node contains the following fields: status node identifier (state_id, e.g., OPEN_TYPE_CONFIRM), status type (state_type, which can be one of the following types: start state, information collection state, confirmation state, execution state, end state, abnormal state, etc.), allowed intents (e.g., account opening consultation, account opening type selection, account cancellation), required slots, conditional slots, forbidden transitions (e.g., prohibiting the jump to the appointment processing state before the data verification fails), allowed actions (e.g., supplementary question, confirmation, explanation, transfer to human agent), evidence policy (e.g., policy-related responses must hit at least one valid knowledge clause), repair policy (e.g., clarification when confidence is low, confirmation of new values ​​when slot conflicts occur), and audit policy (audit_policy, limiting the range of fields that audit records must be written to).

[0029] In an alternative implementation, the business state machine can also be implemented in at least one of the following ways: (1) Implemented using a directed graph data structure, with state nodes as vertices of the graph, allowed state transitions stored as directed edges in the adjacency list, and prohibited state transition sets stored as negative edges or prohibited markers, and the validity of state transitions is verified by querying the vertex outgoing edge set; (2) Implemented using a rule engine, with state transition rules, slot constraint rules and response action constraint rules compiled into an executable rule set, and the rule engine completes matching and triggering at runtime; (3) Constructed using a historical session mining method, by statistically analyzing and clustering the state transition trajectories of historical successful sessions, automatically generating or correcting state transition rules, supplementary question priority weights and knowledge evidence strategies, thereby achieving continuous optimization of the business state machine.

[0030] S140, based on the current state node, load the slot constraint set, the allowed state transition set, the prohibited state transition set, and the allowed response action set; Specifically, the voice dialogue response agent loads the set of slot constraints required for enterprise account opening based on the "Account Opening Type Confirmation" status node: Mandatory slots include entity type, license information, agent's identity, account opening address, appointment method, and risk warning confirmation; conditional slots include enterprise license information that needs to be further collected when the entity type is an enterprise account; mutually exclusive slots include in-store and online processing, which cannot be used simultaneously; derived slots include standard address codes derived from the building, unit, and room number provided by the user; sensitive slots include business license number, unified social credit code, and agent's ID number. The mandatory, sensitive, and format requirements for the same slot can differ under different statuses. Simultaneously, it loads a set of allowed state transitions (e.g., "Account Opening Type Confirmation → Information Collection → Document Verification → Risk Warning Confirmation → Appointment Processing → End") and a set of prohibited state transitions (e.g., prohibiting jump to the appointment processing status before license information is confirmed), as well as a set of allowed response actions.

[0031] In an alternative implementation, derived slots can be derived using either preset rules, such as constructing a standard address code from building, unit, and room numbers, or using a model, such as normalizing non-standard addresses to standard address codes using an entity linking model. In an extended embodiment, sensitive slots can also be anonymized. The complete content of sensitive slots is not repeated in voice broadcasts, and they are stored in an anonymized form in audit records, for example, by replacing several middle characters with a mask, thus reducing the risk of sensitive information leakage.

[0032] S150, based on the candidate transcribed text, the semantic candidate result, the word-level transcription confidence, and the historical round information, update the slot object according to the confidence threshold, source type, coverage rule, and state adaptation rule in the slot constraint set; Specifically, the voice dialogue response agent merges and updates the candidate slots extracted in the current round with the previously confirmed slots: for each new candidate slot, it determines whether its confidence level reaches the confidence threshold, records its source type (speech recognition result, user confirmation, or historical round), decides whether to cover historical values ​​according to coverage rules, and determines whether the slot value is suitable for the current state node according to state adaptation rules. When the newly extracted slot value is inconsistent with the previously confirmed slot value, a confirmation question is generated, asking the user to confirm whether to retain the old value or adopt the new value; when the word-level transcription confidence level is lower than the confidence threshold, a restatement confirmation or a question requesting the user to restate key information is generated. For example, if the subject type confirmed by the user in the first round of voice is "enterprise household", this slot is written to the slot object with the user confirmation as the source, triggering the mandatory requirement of the conditional slot of enterprise certificate information.

[0033] S160, generate candidate response actions based on the current state node, the updated slot object and the knowledge retrieval results, wherein the knowledge retrieval results are obtained based on the current state node, the user intent in the semantic candidate results and the confirmed slot retrieval knowledge evidence; Specifically, the knowledge retrieval process searches for knowledge evidence in the knowledge base based on the current state node, user intent, and confirmed slots. This knowledge evidence must match the state, business type, and user permissions, and each piece of knowledge evidence has an evidence source identifier and an evidence score. For example, when a user asks "How long does it take to open a business account?", the retrieval process retrieves knowledge evidence explaining the processing time. The response action generation process generates candidate response actions based on the current state node, the updated slot object, and the knowledge retrieval results. Candidate response actions are prohibited from being directly announced or executed before passing verification.

[0034] In an alternative implementation, different sets of allowed response actions and knowledge evidence ranges can be loaded based on user identity and permissions. For example, ordinary users are not allowed to trigger internal approval-type response actions, and permission descriptions or manual actions are generated when permissions are insufficient.

[0035] S170, the candidate response action is subjected to state transition legality verification, slot integrity verification, knowledge evidence consistency verification, response action type legality verification and risk compliance verification, respectively to obtain state transition legality score, slot integrity score, knowledge evidence consistency score, response action type legality score and risk compliance score, and the above scores are weighted and summed according to preset weight coefficients to obtain the comprehensive verification score of the candidate response action. Specifically, the state transition legality check is used to determine whether the target state corresponding to the candidate response action belongs to the allowed state transition set of the current state node and does not belong to the prohibited state transition set; the slot integrity check is used to determine whether the required slots and condition slots corresponding to the current state node meet the requirements for value, format, confidence level, and source; the knowledge evidence consistency check is used to determine whether the explanatory, policy, commitment, or guidance responses are bound to valid knowledge evidence that matches the current state node and the user question; the response action type legality check is used to determine whether the current state node allows the generation of this type of response action, for example, prohibiting the execution of tools or the submission of business documents when the data has not been verified; and the risk compliance check is used to perform permission and compliance checks when the response involves identity, finance, government affairs, or sensitive information.

[0036] The voice dialogue response agent calculates the comprehensive verification score of the candidate response actions according to the following formula (1): (1); in, To calculate the overall verification score, This represents the score for the legality of the state transition. This indicates the slot integrity score. The score represents the consistency of knowledge evidence. This indicates the score for the legality of the response action type. Indicates the risk compliance score; , , , and These are weighting coefficients configured based on business scenarios or dynamically adjusted based on historical quality inspection data. Verification is only considered successful when the overall verification score exceeds the preset response threshold and there are no hard prohibitions. In this embodiment, hard prohibitions include: the target state of the candidate response action belongs to the prohibited state transition set, a required slot is missing, explanatory or policy-related responses are not bound to valid knowledge evidence, and the current state node does not allow the generation of this type of response action, etc.

[0037] For example, if a user asks "How long does it take to open a business account?", the candidate response action is linked to the knowledge evidence of the processing time specified in the search results. , , , and The scores for each dimension are as follows: =1.0、 =0.9、 =1.0、 =1.0、 =0.9, at this time the comprehensive verification score V=0.94 is obtained, which is higher than the preset response threshold (e.g. 0.75) and there are no hard prohibitions. The verification is passed and the description type response action can be output.

[0038] In an alternative implementation, the comprehensive verification score can also be calculated or determined in at least one of the following ways: First, using a weighted product method, multiplying the scores of each dimension by their respective weight coefficients to obtain the comprehensive verification score; second, using an item-by-item threshold determination method, where verification is considered passed when the scores of each dimension are all higher than their respective individual thresholds and there are no hard prohibited items; third, using a minimum value determination method, where the minimum score of each dimension is taken as the comprehensive verification score. Wherein, the weight coefficients... to The threshold can be pre-configured by business scenarios or dynamically adjusted based on historical quality inspection data. For example, based on the distribution of verification dimensions corresponding to responses marked as unqualified in historical sessions, the weight coefficient of the corresponding dimension can be increased in reverse. In an extended scheme of this embodiment, the confidence threshold and the response threshold can also be dynamically adjusted based on the historical dialogue quality and quality inspection results. When the dialogue quality is good, the threshold can be appropriately relaxed to improve efficiency, and when there are many quality inspection problems, the threshold can be tightened to improve compliance.

[0039] S180, when the comprehensive verification score is higher than the preset response threshold and there are no hard prohibitions, the candidate response action is output as a voice response action; when the comprehensive verification score is not higher than the response threshold or there are hard prohibitions, an abnormal supplementary question action is generated. Specifically, if a user only provides the company name but not the required licenses and permits, the voice dialogue response agent will not directly enter the appointment processing state. The slot integrity check will detect that the required slot "licenses and permits information" is missing, constituting a hard prohibition, and the check will fail. In this case, no processing response will be output; instead, the system will enter the error correction process. Please refer to [link / reference]. Figure 3 , Figure 3 This is a schematic diagram of a slot constraint and refill process provided in an embodiment of this application. The abnormal refill handling generates a refill action based on the refill priority, which is calculated according to the following formula (2): (2); in, Indicates the first Prioritize questions regarding pending slots or clarifications. Indicates the importance of the business. This indicates the necessity of advancing the state. Indicates the current confidence level. Indicates the risk level. This indicates the dependency of this slot on subsequent slots or tool execution; , , , and These are the weighting coefficients.

[0040] Specifically, voice dialogue response agent selection The highest priority question that is most suitable for the current dialogue state is asked as a follow-up question. "Suitable for the current dialogue state" means that the follow-up question action generated for the question to be asked belongs to the allowed follow-up question action type of the current state node's allowed response action set, and the content of the follow-up question action does not conflict with the confirmed slot. When the highest priority question to be asked is not suitable for the current dialogue state, the next highest priority question to be asked that is suitable for the current dialogue state is selected to generate the follow-up question action. For example, if a user has just denied offline processing, a follow-up question involving the processing method conflicts with the confirmed slot and is not suitable for the current dialogue state. The voice dialogue response agent skips this question and selects the next highest priority question to be asked as a follow-up question.

[0041] For example, for the "Certificate and License Information" slot, set , , , , , , =0.95, =0.2, =0.7, 0.85, calculated For the "Account Opening Address" slot, the values ​​for each factor are: =0.8、 =0.7、 =0.9、 =0.2、 =0.4, calculated as follows =0.5; therefore, priority is given to generating a supplementary question for "Certificate Information": "Have you prepared your business license or unified social credit code?"

[0042] It should be noted that different supplementary questions and recovery actions are generated for different types of anomalies: For missing slot anomalies, i.e., the required slot is empty or below the confidence threshold, a single supplementary question is generated for the highest priority slot; for conflict anomalies, i.e., the new slot is inconsistent with the historically confirmed slot, a confirmation supplementary question is generated, asking the user to confirm whether to retain the old value or adopt the new value; for low confidence anomalies, i.e., the confidence of speech recognition or semantic understanding is low, a restatement confirmation is generated or the user is asked to repeat the key information; for illegal transition anomalies, i.e., the candidate state does not belong to the allowed transition set, a supplementary question or explanation action is given that is allowed in the current state; for insufficient evidence anomalies, i.e. the retrieved evidence is below the threshold or does not match the question, supplementary retrieval, conservative answer, or transfer to human intervention is performed; for user interruption anomalies, i.e., interruption, error correction, or cancellation of expression is detected, the current response action is paused, the user's intent and slot change are reassessed, and state rollback, slot overwrite confirmation, or transfer to human intervention is performed.

[0043] In an alternative implementation, the target for supplementary questioning can be determined in at least one of the following ways: First, using a hard rule chain approach, first determine whether there is a hard-to-fill slot; if so, directly select that slot for supplementary questioning, and then judge sequentially according to factors such as business importance and dependency; Second, using a hierarchical calculation approach, first layer according to the type of exception, for example, first handle illegal jump exceptions and low confidence exceptions, then handle slot missing exceptions, and then calculate the priority within the same layer according to formula (2); Third, using a learning approach, using the success rate of user responses after supplementary questioning in historical dialogues to evaluate the priority of supplementary questioning. to Training and adjustments are performed. In one extended solution of this embodiment, an upper limit for the number of follow-up questions can be set for each session. If the same follow-up question is asked continuously for a preset number of rounds and still cannot be confirmed, the process is automatically transferred to a human agent to avoid endless follow-up questions affecting the user experience. For follow-up questions with a risk level higher than the preset level, the process can be transferred directly to a human agent without follow-up questions.

[0044] S190, write state transitions, slot changes, evidence retrieval, response actions and verification results into the audit log; Specifically, for explanatory responses that pass verification, the voice output processing synthesizes the response text into speech and broadcasts it to the user. Subsequently, the audit log processing writes the voice round identifier, state transition trajectory (e.g., from "Account Type Confirmation" to "Information Collection"), slot change content (e.g., new slots and their source, confidence level), retrieval evidence identifier, response action content, and verification result details into the dialogue audit log for subsequent quality inspection, review, training, and quality evaluation. If a follow-up question is executed, the follow-up question action and verification failure items are also written into the audit log.

[0045] This embodiment achieves closed-loop control of "identification—tracking—constraint—update—generation—verification—supplementary question—auditing" through the organic connection of the above steps: the outputs of S110 and S120 serve as inputs to S130 to S150. The slot object updated in S150 and the knowledge retrieval results in S160 jointly drive the generation of candidate response actions. S170 performs joint verification on the candidate response actions. If the verification passes, it enters S180 to output the voice response action and executes the audit record in S190. If the verification fails, it generates a supplementary question action through S180. When the user's next round of voice arrives, it re-enters S110, thereby completing the response control of multi-round voice dialogue under the premise of meeting the business state machine constraints and slot integrity requirements.

[0046] To further illustrate the above method, this application will further elaborate on the solution by combining three scenarios: after-sales repair, government affairs consultation, and property services.

[0047] This embodiment uses an after-sales repair voice proxy scenario as an example. A user expresses "The air conditioner is broken, please come to repair it this afternoon" via voice in the after-sales hotline. In S110, the user's voice is transcribed using speech recognition. In S120, the candidate transcribed text is semantically parsed to identify the repair intent and the device type slot "air conditioner". In S130, the business state machine determines the current state node as the information collection state based on the current session record (state node identifier, for example, REPAIR_INFO_COLLECT). The required slots for this state node include the repair type, repair address, contact information, and fault description. Allowed intents include repair, supplementing address, supplementing contact information, canceling repair, etc. In S140, the corresponding slot constraint set is loaded, where the fault description is a conditional slot when the repair type is air conditioner. In S150, the slot object is updated, writing the device type "air conditioner" into the slot object based on the source of this round of speech recognition. In S160, the repair process rules are retrieved based on the repair intention and candidate response actions are generated; in S170, the slot integrity check finds that the repair address, contact person, and specific fault phenomenon are missing, constituting a hard prohibition item, and the check fails. The business state machine determines that the current state is the information collection state and prohibits directly entering the work order submission state; in S180, the priority of supplementing questions for each missing slot is calculated according to formula (2). Since the repair address has the highest necessity for state advancement and the highest dependence on subsequent work order execution, the question "Where is the specific address that needs on-site repair?" is supplemented first. After the address is confirmed, the fault phenomenon and contact information are supplemented; in S190, after all necessary slots are satisfied, the confirmation state is entered, the repair information is restated for user confirmation, and the confirmation result and state transition are written to the audit log. This embodiment reflects the blocking effect of the business state machine on the overstepping of business nodes and the control effect of the supplementing question priority on the supplementing question order.

[0048] This embodiment uses a government affairs consultation voice proxy scenario as an example. A user asks, "Can I still apply for this subsidy?" The intent and slot recognition process obtains the consultation intent, but slots for conditions such as subsidy type, applicant identity, region, and application time are missing. The business state machine determines the current state to be a policy consultation state. The knowledge evidence strategy for this state node requires that policy-related responses must be bound to at least one valid policy clause, and questions involving case-specific approval conclusions are not allowed to directly provide a promise-based answer. The voice dialogue response proxy searches the policy knowledge base for the corresponding subsidy policy clause based on the applicant's identity and region. If the search evidence is insufficient or the user's question involves case-specific approval conclusions, the knowledge evidence consistency check fails, and the abnormal supplementary question and recovery processing generates a supplementary clarification action (e.g., "What type of subsidy did you apply for?") or transfers the request to a human agent. For consultations with bound valid policy clauses, after the risk compliance check passes, a policy explanation-type response is output, prompting that the window review conclusion shall prevail, avoiding unfounded promise-based answers. This embodiment demonstrates the constraining effect of knowledge evidence consistency checks and risk compliance checks on promise-based answers in government affairs scenarios.

[0049] This embodiment uses a property service voice agent scenario as an example. A user says, "My house is leaking, please send someone immediately." The intent and slot recognition process obtains an urgent repair intent, and the business state machine enters the emergency information collection state. If the address slot is missing, the slot integrity check fails, and the abnormal supplementary inquiry and recovery processing prioritizes address supplementation; if a historical address exists but is not confirmed in this round of voice, a confirmation supplementary inquiry is generated, such as "Is it still the address registered in Building X, Unit X, Room X?" After the necessary slots are met, a work order submission action is generated. Before work order submission, it must pass state transition legality checks (the current state allows the work order submission action and the target state does not belong to the prohibited state transition set) and risk compliance checks. After work order submission, the work order summary, state transition, slot source, and response verification results are written to the audit log for easy dispatching, quality inspection, and review by the property management platform. This embodiment demonstrates the synergistic effect of supplementary inquiry priority, confirmation supplementary inquiry, and audit records in emergency scenarios.

[0050] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the structure of a voice dialogue response system 20 with business state machine and slot constraints provided in this application embodiment. This embodiment further explains the aforementioned method embodiment in conjunction with the detailed working process of each module of the system. The voice dialogue response system includes a voice input and transcription module, an intent and slot recognition module, a business state machine module, a slot constraint module, a slot object update module, a response action generation module, a response verification module, a response output and exception handling module, and an audit log module. Each module corresponds one-to-one with the nine steps in the aforementioned method embodiment, as detailed below: The voice input and transcription module 210 is used to receive the user's natural speech, perform speech recognition processing on the natural speech, and obtain candidate transcribed text and word-level transcription confidence, corresponding to method step S110; The intent and slot recognition module 220 is used to perform intent recognition, slot extraction and confidence labeling on the candidate transcribed text to obtain semantic candidate results, and to identify error correction expressions, negative expressions, cancellation expressions and supplementary expressions, and output corresponding error correction signals or negative signals, corresponding to method step S120; The business state machine module 230 is used to maintain business process nodes, state transition rules, allowed action sets, prohibited action sets and rollback rules, and to determine the current state node in the business state machine based on the current session record, corresponding to method step S130; The slot constraint module 240 is used to load a slot constraint set, an allowed state transition set, a prohibited state transition set, and an allowed response action set based on the current state node, corresponding to method step S140; The slot object update module 250 is used to update the slot object according to the confidence threshold, source type, coverage rule and state adaptation rule in the slot constraint set based on the candidate transcribed text, the semantic candidate result, the word-level transcription confidence and historical round information, corresponding to method step S150; The response action generation module 260 is used to retrieve knowledge base, frequently asked questions, process rules, business form fields and policy clauses based on the current state node, intent and confirmed slot, obtain knowledge retrieval results, and generate candidate response actions based on the current state node, updated slot object and knowledge retrieval results, corresponding to method step S160; The response verification module 270 is used to perform state transition legality verification, slot integrity verification, knowledge evidence consistency verification, response action type legality verification and risk compliance verification on the candidate response action. The above scores are weighted and summed according to formula (1) to obtain the comprehensive verification score of the candidate response action. The verification is determined to be passed when the comprehensive verification score is higher than the preset response threshold and there are no hard prohibitions. Otherwise, the verification is determined to be failed, corresponding to method step S170. The response output and abnormal question-filling module 280 is used to output the candidate response action as a voice response action when the comprehensive verification score is higher than the preset response threshold and there are no hard prohibitions; when the comprehensive verification score is not higher than the response threshold or there are hard prohibitions, it generates an abnormal question-filling action; and when it detects user interruption, error correction, cancellation or negation, it pauses the current response action, re-evaluates the user intent and slot change, and performs state rollback, slot coverage confirmation or transfer to manual processing, corresponding to method step S180. The audit log module 290 is used to write state transitions, slot changes, evidence retrieval, response actions and verification results into the audit log to form a dialogue audit log, corresponding to method step S190.

[0051] It should be noted that this embodiment, through the collaborative work of the above modules, realizes a complete control chain from voice input to response output and then to audit recording. Each module has clearly defined responsibilities and interfaces, facilitating deployment and operation in systems such as voice customer service platforms, intelligent office assistants, government affairs consultation robots, and property service voice agents. Among them, the voice input and transcription module and the response output and anomaly correction module interface with the telephone voice channel or the network voice channel, respectively, while the remaining modules run on the processor of the voice dialogue response agent.

[0052] Please refer to Figure 5 This illustrates a schematic diagram of an electronic device provided by some embodiments of this application. For example... Figure 5 As shown, the electronic device 30 includes: a processor 300, a memory 301, a bus 302, and a communication interface 303. The processor 300, the communication interface 303, and the memory 301 are connected via the bus 302. The memory 301 stores a computer program that can run on the processor 300. When the processor 300 runs the computer program, it executes the method described above in this application.

[0053] The memory 301 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 303 (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0054] Bus 302 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory 301 is used to store programs. After receiving an execution instruction, the processor 300 executes the program. The methods disclosed in any of the foregoing embodiments of this application can be applied to the processor 300, or implemented by the processor 300.

[0055] The processor 300 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 300 or by instructions in software form. The processor 300 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 301. The processor 300 reads the information in memory 301 and, in conjunction with its hardware, completes the steps of the above method.

[0056] The electronic device and the method provided in this application are based on the same inventive concept and have the same beneficial effects as the methods used, operated, or implemented therein. The electronic device can be a smartphone, tablet computer, laptop computer, desktop computer, intelligent voice customer service terminal, workstation, or edge computing device, or any other device with voice interaction and data processing capabilities.

[0057] This application also provides a computer-readable medium corresponding to the method provided in the foregoing embodiments. Please refer to... Figure 6 The computer-readable storage medium shown is an optical disc 40, on which a computer program (i.e., a program product) is stored, which, when run by a processor, performs the aforementioned method.

[0058] The computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. In a specific application scenario, the computer-readable storage medium may be built into a conference terminal or mobile terminal as part of the firmware; or it may be a standalone software product.

[0059] The computer-readable storage medium provided in the above embodiments of this application and the method provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0060] This embodiment also provides a computer program product, which includes a computer program. When executed by a processor, the computer program implements the method described in any of the above embodiments.

[0061] The computer program product may include any combination of one or more program modules.

[0062] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0063] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0064] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0065] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0066] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application.

Claims

1. A voice dialogue response method based on a business state machine and slot constraints, characterized in that, Includes the following steps: Receive user's natural speech, perform speech recognition processing on the natural speech, and obtain candidate transcribed text and word-level transcription confidence; The candidate transcribed texts are subjected to intent recognition, slot extraction, and confidence labeling to obtain semantic candidate results; The current state node in the business state machine is determined based on the current session record. The state node includes state type, allowed intent set, required slot set, condition slot set, prohibited state transition set, allowed response action set, knowledge evidence policy, and anomaly recovery policy. Based on the current state node, load the slot constraint set, the allowed state transition set, the prohibited state transition set, and the allowed response action set; Based on the candidate transcribed text, the semantic candidate results, the word-level transcription confidence, and the historical round information, the slot object is updated according to the confidence threshold, source type, coverage rule, and state adaptation rule in the slot constraint set. Candidate response actions are generated based on the current state node, the updated slot object, and the knowledge retrieval results. The knowledge retrieval results are obtained based on the current state node, the user intent in the semantic candidate results, and the confirmed slot retrieval knowledge evidence. The candidate response action is subjected to state transition legality verification, slot integrity verification, knowledge evidence consistency verification, response action type legality verification, and risk compliance verification, respectively, to obtain state transition legality score, slot integrity score, knowledge evidence consistency score, response action type legality score, and risk compliance score. The above scores are weighted and summed according to preset weight coefficients to obtain the comprehensive verification score of the candidate response action. When the comprehensive verification score is higher than the preset response threshold and there are no hard prohibitions, the candidate response action is output as a voice response action; When the comprehensive verification score is not higher than the response threshold or there is a hard prohibition item, an abnormal supplementary question action is generated; Write state transitions, slot changes, evidence retrieval, response actions, and verification results into the audit log; The hard prohibitions include at least one of the following: the target state belongs to the prohibited state transition set, the required slot is missing, the response is not bound to valid knowledge evidence, and the response action type is prohibited by the current state.

2. The method according to claim 1, characterized in that, The state types include initial state, information collection state, confirmation state, execution state, termination state, and abnormal state; the knowledge evidence strategy is used to limit the knowledge evidence required to generate explanatory, policy, commitment, or guidance responses; the abnormal recovery strategy is used to limit the triggering conditions for low-confidence clarification, slot conflict confirmation, state rollback, and manual processing.

3. The method according to claim 1, characterized in that, The slot constraint set includes constraints corresponding to each of the following: mandatory slots, conditional slots, mutually exclusive slots, derived slots, and sensitive slots. The mandatory slots are those that must be collected and pass format validation before the current state node is advanced. The conditional slots are those triggered by the business type, user identity, or the value of a preceding slot. The mutually exclusive slots are multiple slots that cannot be valid simultaneously. The derived slots are those derived from confirmed slots according to rules or models. The sensitive slots are those involving identity, account, asset, or personal information.

4. The method according to claim 1, characterized in that, The state transition legality check is used to determine whether the target state corresponding to the candidate response action belongs to the allowed state transition set and does not belong to the prohibited state transition set; the knowledge evidence consistency check is used to determine whether the explanation, policy, commitment or guidance response is bound to valid knowledge evidence that matches the current state node and the user question.

5. The method according to claim 1, characterized in that, The generation of abnormal inquiry actions includes: for each slot to be filled or matter to be clarified, calculating the inquiry priority of the slot to be filled or matter to be clarified according to the following formula based on the business importance of the slot to be filled or matter to be clarified, the necessity for status advancement, the current confidence level, the risk level, and the dependence on the execution of subsequent slots or tools: ； in, Indicates the first Prioritize questions regarding pending slots or clarifications. Indicates the importance of the business. This indicates the necessity of advancing the state. Indicates the current confidence level. Indicates the risk level. Indicates the degree of dependence. , , , and These are weighting coefficients; the highest priority supplementary question is selected from the slots or matters to be supplemented or clarified that are suitable for the current dialogue state to generate a supplementary question action; wherein, suitable for the current dialogue state means that the supplementary question action generated for the slot or matter to be supplemented or clarified belongs to the supplementary question action type allowed by the set of allowed response actions of the current state node, and the content of the supplementary question action does not conflict with the confirmed slot; when the slot or matter to be supplemented or clarified with the highest priority supplementary question is not suitable for the current dialogue state, the second highest priority supplementary question is selected from the slots or matters to be supplemented or clarified that are suitable for the current dialogue state to generate a supplementary question action.

6. The method according to claim 1, characterized in that, When a user interrupts, corrects, cancels, or negates an expression, the current response action is paused, the user's intent and slot changes are reassessed, and a status rollback, slot overwrite confirmation, or manual processing is performed.

7. The method according to claim 1, characterized in that, The updated slot object includes: when the newly extracted slot value is inconsistent with the historically confirmed slot value, generating a confirmation question and retaining the old value or adopting the new value according to the user's response; when the word-level transcription confidence is lower than the confidence threshold, generating a restatement confirmation or requesting the user to restate the key information.

8. The method according to claim 5, characterized in that, The weighting coefficients of the comprehensive verification score and the weighting coefficients of the supplementary question priority are both pre-configured according to the business scenario or dynamically adjusted according to historical quality inspection data.

9. The method according to claim 1, characterized in that, The audit log includes voice round identifiers, state transition trajectories, slot change information, retrieval evidence identifiers, response action content, and verification result details.

10. A voice dialogue response system with business state machine and slot constraints, characterized in that, include: The voice input and transcription module is used to receive the user's natural voice, perform speech recognition processing on the natural voice, and obtain candidate transcribed text and word-level transcription confidence. The intent and slot recognition module is used to perform intent recognition, slot extraction and confidence labeling on the candidate transcribed text to obtain semantic candidate results. The business state machine module is used to determine the current state node in the business state machine based on the current session record. The state node includes state type, allowed intent set, required slot set, condition slot set, prohibited state transition set, allowed response action set, knowledge evidence strategy, and exception recovery strategy. The slot constraint module is used to load a slot constraint set, an allowed state transition set, a prohibited state transition set, and an allowed response action set based on the current state node; The slot object update module is used to update the slot object according to the confidence threshold, source type, coverage rule and state adaptation rule in the slot constraint set based on the candidate transcribed text, the semantic candidate result, the word-level transcription confidence and historical round information; The response action generation module is used to obtain knowledge retrieval results based on the current state node, the user intent in the semantic candidate results, and the confirmed slot retrieval knowledge evidence, and to generate candidate response actions based on the current state node, the updated slot object, and the knowledge retrieval results. The response verification module is used to perform state transition legality verification, slot integrity verification, knowledge evidence consistency verification, response action type legality verification, and risk compliance verification on the candidate response action, and obtain state transition legality score, slot integrity score, knowledge evidence consistency score, response action type legality score, and risk compliance score respectively. The above scores are weighted and summed according to preset weight coefficients to obtain the comprehensive verification score of the candidate response action. The response output and abnormal question filling module is used to output the candidate response action as a voice response action when the comprehensive verification score is higher than the preset response threshold and there are no hard prohibition items. When the comprehensive verification score is not higher than the response threshold or there is a hard prohibition item, an abnormal supplementary question action is generated; The audit log module is used to write status transitions, slot changes, evidence retrieval, response actions, and verification results into the audit log. The hard prohibitions include at least one of the following: the target state belongs to the prohibited state transition set, the required slot is missing, the response is not bound to valid knowledge evidence, and the response action type is prohibited by the current state.

Images