A method and related device for intelligent outbound call dialogue control
By using the collaborative processing of task graphs and large models, the problems of inaccurate recognition and process control loss in intelligent outbound calling systems when faced with non-standard customer expressions were solved, achieving accuracy and stability of the dialogue process and improving recognition precision and result reliability.
Patent Information
- Application Number
- CN202611452883.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-18
- Publication Date
- 2026-11-10
AI Technical Summary
Existing intelligent outbound calling systems suffer from inaccurate recognition, process loss, limited adaptability, and unreliable results when dealing with non-standard customer expressions and cross-round semantics, making it difficult to reliably complete business processes.
By acquiring the task graph execution configuration and customer transcribed text, the system performs basic intent, slot and route recognition, combines historical conversations to determine the completeness of required entities, triggers multi-stage semantic enhancement as needed, and integrates adjudication to generate speech scripts, ensuring the accuracy and coherence of the dialogue process.
It improves the recognition accuracy, process control accuracy and result reliability of intelligent outbound dialogue, ensures the integrity of information collection and the stability of multi-round interaction during the dialogue, and reduces the latency loss caused by large model calls.
Smart Images

Figure CN123075128A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of intelligent outbound call dialogue control technology, and in particular to an intelligent outbound call dialogue control method and related apparatus. Background Technology
[0002] Outbound calling processes are typically driven by finite state machines or task graphs, with basic intent models and slot models responsible for understanding customer input. This approach is controllable, but in real calls, it frequently encounters issues such as pronoun errors, omissions, accent transcription errors, and multi-turn context dependencies. If the larger model is allowed to directly determine responses and transitions, problems arise such as skipping the current node, supplementing non-existent facts, or omitting required information.
[0003] Based on the existing implementation, the areas most prone to problems are: 1. The task graph and the large model work separately: the model does not know which events the current node can trigger, and the task graph cannot get the model's interpretation of cross-process semantics, so the result can only be hard-connected by a fixed threshold.
[0004] 2. The enhanced triggering conditions are too simple: calling it in every round will increase the waiting time, and calling it only when there is a complete rejection will not cover the cases of low confidence, candidate conflict, and only half of the entities have been collected.
[0005] 3. The generated script lacks business boundaries: The model may rewrite variables, risk disclosures, or process instructions, or it may not bring the topic back to the current node objective after answering the customer's interjection.
[0006] Therefore, how to overcome the shortcomings of existing technologies in intelligent outbound call interaction, such as inaccurate identification, process control failure, limited adaptability to scenarios, and unreliable results, is a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0007] The purpose of this application is to provide an intelligent outbound call dialogue control method and related apparatus to improve the accuracy of intelligent outbound call dialogue recognition, routing control accuracy and result reliability.
[0008] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an intelligent outbound call dialogue control method, including: Obtain the task graph operation configuration for this round, the client's transcribed text for this round, and historical sessions; the task graph operation configuration includes the task graph version, task graph namespace, set of allowed events for the current node, declaration slots for the current node, standard answer for the node, node type, large model capability switch, prompt word scenario configuration, node answer configuration, and model call timeout.
[0009] Based on the client's current transcribed text, basic intent, slots, and route recognition are performed to obtain the basic task answer; the basic task answer includes: the original current node, the previous node, basic candidate events, recognized slots, basic interface parameters, interface execution results, basic answer, basic route status, and basic intent; the basic route status is unique valid route, rejected, low confidence, or candidate conflict.
[0010] Based on the task graph runtime configuration and the basic task answers, determine the completeness of the required entities for the current node.
[0011] Based on the basic routing state and the completeness of the required entities of the current node, a first judgment result is obtained to determine whether to perform multi-stage semantic enhancement.
[0012] If the first judgment result is negative, then the final task answer for this round is determined based on the basic task answer. The final task answer includes: the final current node, the target node, the final intent, the valid slots, the interface parameters, and the answer type.
[0013] If the first judgment result is yes, then based on the task graph running configuration and the historical session, the corresponding large model multi-stage semantic enhancement is performed on the customer's current round text to obtain the enhancement result; the multi-stage semantic enhancement includes context question rewriting, process intent enhancement and entity processing; the enhancement result includes standardized questions, enhanced candidate events and enhanced candidate entities.
[0014] Under the constraints of the event whitelist and slot whitelist in the current task graph, the basic task answer and the enhanced result are fused and adjudicated to determine the final task answer for this round.
[0015] Based on the final task answer, a voice script is generated, and the final current node, the valid word slots, and the response content are written back to the session context as input for the next round of task graph execution.
[0016] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent outbound call dialogue control method described above.
[0017] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent outbound call dialogue control method described above.
[0018] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent outbound call dialogue control method described above.
[0019] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides an intelligent outbound call dialogue control method and related apparatus. By simultaneously acquiring the current task graph operation configuration, the customer's current transcribed text, and complete historical conversations, it provides complete and accurate process boundaries and contextual basis for each round of dialogue parsing. Through basic intent, slot, and route recognition of the customer's transcribed text, it obtains basic task answers, achieving preliminary structured parsing of the customer's original semantics. Based on the task graph operation configuration and basic task answers, it quantitatively solves the completeness of the required entities for the current node, achieving accurate determination of the node's information collection completeness. Combining the basic route status and entity completeness as dual conditions, it adaptively determines whether to initiate multi-stage semantic enhancement, enabling on-demand triggering of enhancement strategies. This avoids the latency loss caused by fully calling a large model and accurately covers abnormal recognition scenarios such as low confidence, candidate conflicts, and incomplete entity information. When enhancement is not needed, it directly outputs the current task based on reliable basic task answers. The nodes, intents, slots, and parameter results required for each round of dialogue are enhanced by combining task graph configuration and historical conversations to rewrite contextual questions, enhance process intents, and perform entity-specific processing. This outputs standardized questions, enhanced candidate events, and enhanced candidate entities, compensating for semantic defects caused by ambiguity in the original transcription, omission of references, and recognition bias. Strong constraints from task graph event whitelists and slot whitelists are used to fuse and adjudicate the basic recognition results and large model enhancement results, effectively filtering out model out-of-bounds outputs, illegal events, invalid entities, and abnormal parameters, ensuring that the dialogue process strictly conforms to business node specifications. Finally, based on the final task answer, a voice script is generated, and the conversation context of node status, valid slots, and response content is simultaneously written back, achieving coherent inheritance and closed-loop update of multi-round dialogue states. This comprehensively improves the semantic recognition accuracy, process jump standardization, entity collection completeness, and multi-round interaction stability of intelligent outbound dialogue. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is an application environment diagram of an intelligent outbound call dialogue control method according to an embodiment of this application; Figure 2 A flowchart illustrating an intelligent outbound call dialogue control method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the overall process of an intelligent outbound call dialogue control method provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0023] This application belongs to the fields of intelligent outbound calling, natural language processing, and dialogue flow control. It solves the problem of how the system can continue to stably complete the process when the business path has been defined in the task graph, but the customer's phrasing is not standard, cross-round omissions are made, or multiple meanings are expressed at once.
[0024] The task graph in the existing system is responsible for nodes, events, slots, interfaces, and transitions; the large model is only used as a controlled enhancement component. Each call includes the events allowed by the current node, the writable slots, and the target for this round. The returned result is only allowed to enter the task graph state after being verified by a whitelist.
[0025] Applicable scenarios include, but are not limited to, the following: 1. Securities, banking and insurance follow-up: When customers give vague answers or refer to previous texts, combine the content of the most recent rounds to determine their true intentions, but the model is not allowed to add product information or business conclusions on its own.
[0026] 2. Government notifications and service follow-ups: Identify questions, supplementary explanations, and refusal to express opinions from the public in the fixed processing flow, and return to the original node to collect the required information after answering the questions.
[0027] 3. Medical follow-up and appointment reminders: Extract entities such as date, symptoms, hospital or department from spoken language, restate and confirm key items, and only write them into subsequent interface parameters after confirmation.
[0028] 4. Marketing Lead Screening: When a customer's sentence simultaneously includes their level of interest, time commitment, and reason for rejection, break it down into multiple candidate leads and select the next step within the scope allowed by the task map.
[0029] 5. Online customer service and agent assistance: Use the same set of task graphs and enhanced orchestrations. When the model times out or the output is abnormal, the original recognition results and standard node scripts can be used directly.
[0030] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0031] The intelligent outbound call dialogue control method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the client's current round transcribed text to be processed to server 104. Server 104 receives the client's current round transcribed text and, for the client's current round transcribed text, obtains the current round task graph operation configuration and historical sessions. The task graph operation configuration includes the task graph version, task graph namespace, current node allowed event set, current node declared slot, node standard answer, node type, large model capability switch, prompt word scenario configuration, node answer configuration, and model call timeout. Based on the client's current round transcribed text, basic intent, slot, and route recognition are performed to obtain the basic task answer. The basic task answer includes: original current node, previous node, basic candidate events, identified slots, basic interface parameters, interface execution result, basic answer, basic route status, and basic intent. The basic route status is unique valid route, rejected, low confidence, or candidate conflict. Based on the task graph operation configuration and the basic task answer, the completeness of the required entities for the current node is determined. Based on the basic route status and the... The system determines whether to perform multi-stage semantic enhancement based on the completeness of the required entities at the current node, obtaining a first judgment result. If the first judgment result is negative, the final task answer for this round is determined based on the basic task answer. The final task answer includes: the final current node, the target node, the final intent, valid slots, interface parameters, and the answer type. If the first judgment result is positive, multi-stage semantic enhancement of the client's text for this round is performed based on the task graph running configuration and the historical session, obtaining an enhancement result. The multi-stage semantic enhancement includes contextual question rewriting, process intent enhancement, and entity processing. The enhancement result includes standardized questions, enhanced candidate events, and enhanced candidate entities. Under the constraints of the event whitelist and slot whitelist of the current task graph, the basic task answer and the enhancement result are fused and adjudicated to determine the final task answer for this round. Based on the final task answer, a voice script is generated, and the final current node, the valid slots, and the response content are written back to the session context as input for the next round of task graph running. The server 104 can feed back the obtained voice script to the terminal 102. In addition, in some embodiments, the intelligent outbound call dialogue control method can also be implemented by the server 104 or the terminal 102 separately. For example, the terminal 102 can directly process the current round of customer transcription text to be processed, or the server 104 can obtain the current round of customer transcription text to be processed from the data storage system and process it.
[0032] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.
[0033] In one exemplary embodiment, such as Figure 2 As shown, an intelligent outbound call dialogue control method is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 208: Step 201: Obtain the task graph operation configuration for this round, the client's transcribed text for this round, and historical conversations; the task graph operation configuration includes the task graph version, task graph namespace, set of allowed events for the current node, declaration slots for the current node, standard answer for the node, node type, large model capability switch, prompt word scenario configuration, node answer configuration, and model call timeout; the historical conversations are the conversations between the client and the robot in the most recent rounds.
[0034] Step 202: Based on the client's current transcribed text, perform basic intent, slot and route recognition to obtain the basic task answer; the basic task answer includes: original current node, previous node, basic candidate event, recognized slot, basic interface parameters, interface execution result, basic answer, basic route status and basic intent; the basic route status is unique valid route, rejected, low confidence or candidate conflict.
[0035] Step 203: Based on the task graph running configuration and the basic task answer, determine the completeness of the required entities for the current node.
[0036] Step 204: Based on the basic routing state and the completeness of the required entities of the current node, determine whether to perform multi-stage semantic enhancement to obtain the first judgment result.
[0037] Step 205: If the first judgment result is negative, then the final task answer for this round is determined based on the basic task answer; the final task answer includes: the final current node, the target node, the final intent, the valid slots, the interface parameters, and the answer type.
[0038] Step 206: If the first judgment result is yes, then based on the task graph running configuration and the historical session, perform corresponding large model multi-stage semantic enhancement on the customer's current round text to obtain the enhancement result; the multi-stage semantic enhancement includes context question rewriting, process intent enhancement and entity processing; the enhancement result includes standardized questions, enhanced candidate events and enhanced candidate entities.
[0039] Step 207: Under the constraints of the event whitelist and slot whitelist of the current task graph, perform a fusion decision on the basic task answer and the enhanced result to determine the final task answer for this round.
[0040] Step 208: Based on the final task answer, generate a voice script and write the final current node, the valid word slots, and the response content back to the conversation context as input for the next round of task graph execution.
[0041] By implementing steps 201 to 208 above, this application completes basic intent, slot, and route recognition based on task graph configuration, user-transcribed text, and historical conversations. It triggers multi-stage semantic enhancement as needed by combining route status and required entity completeness, and integrates the adjudication recognition results under event and slot whitelist constraints. In conventional scenarios, the basic recognition results are used directly to reduce the overhead of large model calls and shorten response latency. In fuzzy recognition scenarios, the recognition deviation is corrected through context rewriting, intent and entity processing to suppress illegal routes and entity misrecognition, generate voice scripts and update the conversation context. While ensuring that the intelligent outbound call dialogue strictly follows the business logic of the task graph, it also takes into account the dialogue processing efficiency, intent slot recognition accuracy, and conversation continuity.
[0042] As an optional implementation, this application completes the loading and construction of the task graph runtime configuration through the following steps, so that the various configuration parameters take effect in an orderly manner and constrain the model enhancement logic of this round of dialogue.
[0043] First, the task graph loads content according to the published graph identifier and corresponding task graph version, including nodes, connections between nodes, event monitoring and execution actions, intents, slots, node answers, and interface parameters. After the dialogue begins, the system obtains the same task graph runtime information based on the graph identifier and version returned by the dialogue platform, and establishes the correspondence between the current node, the previous node, the events allowed for the current node, the slots that the node should collect, and the node's standard answer. Subsequent intents, entities, and dialogues returned by the larger model are all bounded by this task graph runtime information; the creation of nodes, events, or slots that do not exist in the task graph is not allowed.
[0044] The large-scale model used in this application is a large language model that can receive text commands and multi-turn dialogue context, and output natural language or structured results. Its input includes the customer text after speech recognition, historical conversations, current task graph nodes, and node-related knowledge and entity information. The speech recognition model and speech synthesis model are deployed independently at the input and broadcast ends and are not part of the large-scale model in this step. If a multimodal model that can directly receive speech is used, its output results need to be uniformly converted into the same intent, entity, and utterance results as the text-based large-scale language model before entering the task graph to execute subsequent logic.
[0045] This application sets up multi-dimensional, scenario-based capabilities for large-scale models and centrally manages the activation of each capability through a large-scale model capability switch. These capabilities include contextual question rewriting, process intent enhancement, slot entity enhancement, entity self-clarification, entity information verification, entity result confirmation, and node intelligent response. Each capability uses independent prompts and output requirements. For example, process intent enhancement prompts carry candidate intents or events available at the current node, entity enhancement prompts carry the slot name and description declared by the current node, and node intelligent response prompts carry the current node's goal, standard answer, hit knowledge, and confirmed entities. This ensures that different capabilities only process the data needed for their current stage.
[0046] Dynamic reading and loading of prompts are accomplished based on the task graph namespace and prompt scenario configuration. Prompts are read according to the task graph's namespace and prompt scenario. First, prompts already in effect for that namespace are retrieved from the cache; if the cache is not found, the persistent configuration is queried, and the currently valid and highest-priority prompt is selected according to the scenario and then written back to the cache. The process intent scenario also normalizes optional inputs such as candidate similarity questions; if no valid candidates are available, empty candidate structures are not passed to the model to avoid invalid information interfering with the model's judgment.
[0047] The scenario-level configuration synchronously defines the model service operation rules, including whether the large model online service is enabled, the start / stop status of each enhancement capability, the maximum number of dialogue rounds allowed for intelligent reply, and the model call timeout. A global service switch is also configured; when the global large model service is not deployed or the corresponding scenario's capability switch is not enabled, the model enhancement call is skipped, and the reply is completed using the basic task answer and node standard answer from the task graph. Furthermore, the configuration of node answers enables speech polishing capabilities; by configuring node polishing identifiers and custom prompts, standardized answers can be personalized.
[0048] During the configuration preparation phase, the system centrally integrates the task graph version, task graph namespace, large model capability switch, prompt word scenario configuration, node answer configuration, and model call timeout, generating a configuration set specific to this round of dialogue. This configuration set runs through the entire process of context building, model calling, result verification, and answer generation, with each configuration parameter having a hierarchical constraint relationship: task graph-related parameters determine the overall business processing scope, large model capability switches control whether model capabilities are triggered, prompt word scenario configurations define the specific processing logic of the model, and model call timeout defines the abnormal exit conditions of the model. All parameters work together to complete the model enhancement constraints for this round of dialogue.
[0049] After the configuration is loaded, the system verifies the runtime conditions of the model capabilities required by the current node to ensure that the task graph runtime configuration is valid and usable. If the node has the intelligent response capability enabled but the large model online service switch is not enabled, no model call will be initiated; if the node has entity verification capability configured but no valid prompt words are matched, the basic slot parsing results will be retained, the task graph's native entity clarification logic will be executed, and finally, a complete and usable task graph runtime configuration adapted to this round of dialogue will be output.
[0050] Based on the task graph runtime configuration constructed above, the model enhancement processing for this round of dialogue has clear execution boundaries and constraints: 1. The task graph runtime configuration first determines the candidate events, slots, and answers allowed for the current node, giving the large model processing a clear input and output range. 2. Namespace configuration and prompt words are selected within the task graph scope and provided to specific enhancement capabilities along with parameters such as model service timeout and context turn number. 3. This stage ultimately obtains the task graph runtime information and the large model runtime configuration, which together serve as the basis for the next stage of parsing this round of dialogue and determining whether enhancement is triggered.
[0051] As an optional implementation, step 202, "based on the client's current transcription text, performing basic intent, slot, and route recognition to obtain the basic task answer," specifically includes: Step 2021: Write the customer's current round transcribed text into the conversation message according to the user role specification to complete the initialization of the current round of dialogue context. Then, based on the task graph running configuration, call the original dialogue platform to perform basic intent recognition, slot extraction and dialogue route recognition operations.
[0052] Step 2022: Based on the original recognition results returned by the dialogue platform, sequentially perform response status verification, domain robot type verification, and task-type robot result verification, filter invalid and abnormal recognition data, and after all verifications pass, parse and generate the basic task answer.
[0053] As an optional implementation, determining the final task answer for this round based on the basic task answer specifically includes: sending the basic task answer into task graph constraint fusion; under the constraints of the event whitelist and slot whitelist of the current task graph, verifying and selecting the answer for the basic task answer to determine the final task answer for this round. In this step, under the constraints of the event whitelist and slot whitelist of the current task graph, the system verifies the basic candidate events, identified slots, basic interface parameters, and basic routing status in the basic task answer, excluding events that do not belong to the allowed event set of the current node and entities that are not declared in the slots of the current node; after the verification is passed, the answer module selects the node standard answer, intent response, entity clarification message, or interface success / failure message according to the current node type and node answer configuration, thereby determining the final task answer for this round. The final task answer includes the final current node, target node, final intent, valid slots, interface parameters, and answer type.
[0054] As an optional implementation, based on the task graph runtime configuration and the historical sessions, the current round of text from the client is subjected to corresponding large-model multi-stage semantic enhancement to obtain the enhancement result, which specifically includes the following: Under the constraints of the large model capability switch in the task graph runtime configuration, selectively execute one or more of the following sub-capabilities: When the client's current text contains referential errors or omitting of sentences, the contextual question rewriting is performed, and the client's current text is semantically completed by combining the historical conversation to obtain a standard question.
[0055] When the basic routing state is denied, low confidence, or candidate conflict, the execution process intent is enhanced to obtain enhanced candidate events.
[0056] When the completeness representation of the required entity of the current node contains uncollected required slots, missing entity processing is performed to obtain enhanced candidate entities.
[0057] The canonical question, enhanced candidate event, and enhanced candidate entity together constitute the enhanced result.
[0058] As an optional implementation, the missing entity processing specifically includes the following steps 301 to 309: Step 301: When it is detected that there are uncollected required word slots in the original current node, the entity autonomously clarifies and outputs follow-up questions about the missing information to the calling customer.
[0059] Step 302: Obtain the reply text from the calling customer and extract the entity value. Perform format and semantic verification on the entity value according to the preset entity verification rules and generate the verification result.
[0060] Step 303: If the verification result is a verification failure, the corresponding verification failure branch in the task graph is triggered, and the original current node is maintained.
[0061] Step 304: If the verification result is successful, the candidate entity corresponding to the entity value is placed in the pending confirmation state, and an entity confirmation query is initiated to the calling customer.
[0062] Step 305: Obtain the caller's response regarding the entity to be confirmed.
[0063] Step 306: If the calling customer replies with confirmation, the candidate entity corresponding to the entity value is marked as valid and updated to the slot list.
[0064] Step 307: If the customer replies with a denial, determine whether the current entity collection round has reached the node's preset maximum allowed round.
[0065] Step 308: If the maximum allowed number of rounds has not been reached, return to step "obtain the reply text of the calling customer and extract the entity value, perform format verification and semantic verification on the entity value according to the preset entity verification rules, and generate the verification result".
[0066] Step 309: When the maximum allowed number of rounds has been reached, trigger the super-round branch corresponding to the task graph.
[0067] Entities that are in a pending confirmation state, have failed verification, or have been rejected by the customer will not be written into the slot list and will be prohibited from flowing into subsequent interface nodes.
[0068] Specifically, upon receiving the current customer conversation, the system first writes the customer's text into the conversation message according to user roles, and then calls the existing dialogue platform for basic recognition. The returned results need to be validated sequentially by response status, domain robot type, and task-oriented robot results. After successful validation, the system parses to obtain the basic task answer, which includes the current graph, current node status, routing status, intent, slots, interface parameters, and large model runtime statistics. For contextual question rewriting, the system reads one or more standardized questions after rewriting and records whether the rewriting was successful and the corresponding call status.
[0069] Then, it will load the most recent rounds of customer and robot messages, and supplement them with the current node, the previous node, the previous hop event, the current and historical slots, the interface return and hit knowledge to form the context required for enhanced processing.
[0070] Since the basic identification has yielded a unique valid route, the answer to the basic task is directly used as the answer to the final task.
[0071] When multi-stage semantic enhancement is required, such as process intent enhancement, slot enhancement, or question rewriting, or when there are still uncollected entities at the current node, the corresponding large model capabilities are invoked.
[0072] Contextual question rewriting; based on historical conversations, completes the references and omissions in the client's current expression to generate standard questions.
[0073] Enhanced flow intent: The process intent is determined by combining the standard question and the available flow intents at the current node. The enhanced flow intent returned by the dialogue platform also carries model call tracing information. Only results that can be found in the current task graph are parsed and used as candidate routing information in the task answer; candidates not declared in the task graph will not proceed to subsequent nodes.
[0074] Missing entity handling: The system determines uncollected items based on the difference between the slots to be collected at the current node and the slots already identified in this round. When uncollected items exist, the system combines the slot name, slot description, recent dialogue, and the current node's target into an entity self-clarification request, generating follow-up questions for the missing information. After identifying the entity, the system continues to read the entity verification results, keeping values that do not meet format or semantic requirements at the current node and providing verification statements. Entities requiring confirmation enter a result confirmation or secondary confirmation state; only confirmed values can enter the current slot list and the conversation slot list. If verification fails, the customer denies, or the allowed number of rounds for the node is exceeded, the corresponding rejection, round skipping, or re-collection of answers is used in the task graph.
[0075] This process addresses the technical problem in existing technologies where entity processing only reaches the extraction stage. Existing solutions cannot complete subsequent clarification, format validation, customer confirmation, and status activation processes for identified entities such as names, dates, and amounts, easily leading to the transmission of uncertain entities to the interface node. This application adds an entity validation sub-step to perform format validation on identified entities. When an entity is ambiguous or has an abnormal format, a clarification message is generated. After customer confirmation, the entity is marked as effective. Unconfirmed entities are marked as temporary entities and are prohibited from being directly transmitted to the interface node, thus preventing uncertain values from flowing into the interface node.
[0076] In another exemplary embodiment of this application, step 207, "under the constraints of the event whitelist and slot whitelist of the current task graph, perform a fusion decision on the basic task answer and the enhanced result to determine the final task answer for this round," can be replaced by steps 2071 to 2078: Step 2071: Compare the enhanced candidate event with the set of allowed events for the current node. Only if the enhanced candidate event belongs to the set of allowed events for the current node and the target node corresponding to the enhanced candidate event exists, the enhanced candidate event is determined as a valid event.
[0077] Step 2072: Compare the enhanced candidate entity with the current node declaration slot. Only when the entity name of the enhanced candidate entity belongs to the current node declaration slot and the entity value of the enhanced candidate entity meets the preset format verification rules or business verification rules, the enhanced candidate entity is determined as a valid entity.
[0078] Step 2073: Determine whether the basic candidate event and the valid event point to the same task graph event.
[0079] Step 2074: If the basic candidate event and the valid event point to the same task graph event, then the task graph event is determined as the triggering event for this round.
[0080] Step 2075: If the basic candidate event and the valid event point to different task graph events, and the basic routing state is a unique valid route, then the basic candidate event is determined as the trigger event for this round.
[0081] Step 2076: If the basic candidate event and the valid event point to different task graph events, and the basic routing state is rejection, low confidence, or candidate conflict, then the triggering event for this round is determined from the set of allowed events of the current node by combining the completeness of the required entity of the current node and the actual routing state of the task graph.
[0082] Step 2077: When the triggering event in this round cannot be uniquely determined, keep the original current node unchanged, and select clarification script, denial script, or re-collected script as the answer type for this round.
[0083] Step 2078: When the triggering event for this round is uniquely determined, based on the triggering event for this round, determine the final current node, target node, final intent, valid slots, interface parameters, and answer type for this round.
[0084] Specifically, the current node, the previous node, the basic routing status, the basic intent, the identified slots, and the interface execution results are first read from the answer to the basic task to form a pending decision status.
[0085] Subsequently, each enhanced candidate event and enhanced candidate entity returned by the large model is validated: the event must belong to the set of allowed events of the current node and the target node must exist; the entity name must belong to the declaration slot of the current node, and the entity value must also meet the format, enumeration, or business validation rules. Candidates that fail the validation are not written to the task graph state.
[0086] After verification, if the basic task answer and the enhanced result point to the same event, that event is adopted; if they are inconsistent but the basic route is uniquely valid, the basic route is maintained; if the basic route is rejected, has low confidence, or has candidate conflicts, the jump is determined in the allowed event based on the completeness of the required entities and the actual routing status of the task graph, and the final task answer for this round is determined. The final task answer includes the final current node, target node, final intent, valid slots, and interface parameters; the node standard answer, intent response, entity clarification script, or interface success / failure script is obtained according to the node type and node answer configuration. If a unique determination is still not possible, the current node is maintained and clarification, rejection, or re-collection of answers is selected.
[0087] Once a node is determined, a processing method can be selected based on the node type. Clarification nodes can use either a precise answer or an intelligent reply. Interface nodes select a success or failure answer based on the interface execution status. Termination nodes select a termination answer. When multiple intent replies exist, the answer list is formed by filtering out unmatched intents according to the pre-configured intent reply list in the final task answer. If the intelligent reply or polishing has already integrated intent knowledge and node objectives, the answer list is generated based on the intelligent reply or polishing results, without re-attaching the original node script.
[0088] In another exemplary embodiment of this application, step 208, "based on the final task answer, generating a voice script, and writing the final current node, the valid slots, and the response content back to the session context as input for the next round of task graph execution," can be replaced by steps 2081 to 2085: Step 2081: Based on the node answer configuration, determine whether to call the large model service to execute node response generation, and obtain the second judgment result.
[0089] Step 2082: If the second judgment result is yes, assemble the model request based on the final task answer, the node standard answer, the final hit knowledge, the final confirmed entity, the custom prompt content, and the historical session, and call the large model service to execute node response generation to obtain the node natural language answer; the node response generation supports intelligent reply mode and speech polishing mode. The final hit knowledge and the final confirmed entity are the final business knowledge and the effective entity of this round, respectively, obtained after the final task answer is determined in this round, and are used for the final response generation of the large model.
[0090] Step 2083: If the second judgment result is negative, then directly obtain the node standard answer as the node natural language answer.
[0091] Step 2084: Systemally replace the interface variables and slot variables in the natural language answer of the node, and parse and whitelist the TTS text and call control instructions in the replaced script to generate a voice script.
[0092] Step 2085: Write the final current node, the valid slots, and the response content back to the session context as input for the next round of task graph execution.
[0093] As an optional implementation, the intelligent outbound call dialogue control method further includes: When the large model service is not invoked, the model call times out, or the model returns an abnormal result, the node's standard answer is used as the output message.
[0094] When the final task answer parsing fails, the model returns an empty value, or the prompt words cannot be obtained, the basic task answer, the node standard script, the node rejection answer, or the super-rotation answer are used to continue execution.
[0095] When context question rewriting fails, the original customer question is retained for basic identification, blocking the abnormal spread of the enhanced link and avoiding the failure of the entire round of dialogue due to a single point of failure.
[0096] As an optional implementation, the large model service call runs in an independent execution task and waits for the result according to the timeout configured for the scenario; the model-generated content only replaces the natural language expression and does not change the nodes, intents, slots and interface execution results already determined by the task graph.
[0097] Specifically, when the node answer is configured in intelligent response mode (i.e., when intelligent response or polishing is required), the large model request is assembled using the target node, the node's standard answer, the final hit knowledge, the final confirmed entity, custom prompts, and historical conversations. When the node answer is configured in usePolish, the selected standard script is polished. In both cases, the model only generates natural language expressions and does not change the adjudicated nodes, events, slots, and interface results. Interface variables and slot variables are replaced in the order of the answer list. The speech script parser distinguishes between TTS text and call control instructions, retaining only instructions allowed by the task graph whitelist. The processed answer, node, intent, slot, interface parameters, and model call status together constitute the structured response of this round.
[0098] Furthermore, in scenarios requiring intelligent responses or dialogue refinement, a model request is assembled based on the target node, the node's standard answer, the final hit knowledge, the final confirmed entity, custom prompts, and historical conversations. This request then invokes the large model service to obtain the node's natural language answer. The large model service call runs independently in a task execution environment and waits for the call result according to the timeout configured for the scenario. If the call succeeds, the obtained response is used as the node's natural language answer. If the call times out, returns an exception, or the large model service is not enabled, the original standard answer is retained as the node's natural language answer. The content generated by the large model only replaces the natural language expression; it does not change the already determined nodes, intents, slots, or interface execution results in the task graph.
[0099] After obtaining the natural language answers from the nodes, a voice script parser is used to separate the TTS text from the call control commands, and interface variables and slot variables are replaced: interface variables are replaced according to the fields returned by the interface, and slot variables are taken from the current node slots or the session history slots; variables that cannot be obtained are deleted or left blank according to preset rules. Subsequently, the TTS text and call control commands are whitelisted and combined to generate a voice script that can be recognized by the call execution end, avoiding the direct broadcast of variable placeholders or non-business commands generated by large models to the customer.
[0100] The final task answer is assembled into a structured response, which includes node responses, intent responses, and a list of answers, along with the intent identified in this round, the current slot, the conversation slot, and the status of various large model calls. Any task graph execution anomalies output by the dialogue platform are transmitted along with the structured response for recording and troubleshooting after each call round. The content written to the robot's history messages is the response content after variable substitution and dialogue processing; the actual answer heard by the customer is read during the next round of context construction.
[0101] When the final task answer parsing fails, the model call times out, the model returns empty, or the prompt words cannot be obtained, the system does not write the incomplete enhancement results into the node state. Instead, it continues execution using the basic task answer, the node's standard dialogue, the node's rejection answer, or the super-round answer. When context question rewriting fails, the original customer question is retained for basic recognition, preventing the abnormal propagation of the enhancement link and avoiding single-point anomalies that could cause the entire round of dialogue to fail.
[0102] This yields a dialogue result that can be directly sent to the call execution end. This result records the current node, target node, identified and confirmed intent, valid slots, interface parameters, ordered natural language responses from nodes, voice script, and the status of various enhanced capabilities invoked in this round. The conversation module saves the nodes, valid slots, and responses from this round to the conversation context, serving as input for the next round of task graph execution.
[0103] This process addresses the technical problem of incomplete exception branches in existing technologies. Existing technologies tend to silently wait or treat exceptions as normal results and continue node jumps when model calls time out, return empty content, structure parsing fails, or unauthorized events are encountered. This application's embodiment adds an independent exception capture branch to the multi-stage semantic enhancement step. When model calls time out, return empty content, or structure parsing fails, enhancement is skipped, and the basic task answer is used directly for routing. When the enhancement result contains unauthorized events outside the whitelist, the unauthorized event is directly filtered during the fusion and adjudication stage and not included in the candidate events. Simultaneously, exception log marking and degradation strategies are added to avoid system blocking and erroneous jumps.
[0104] like Figure 3 The diagram illustrates an intelligent outbound call dialogue control method according to an embodiment of this application. First, customer input and basic recognition are performed: acquiring the current round task graph operation configuration, the customer's current round transcribed text, and historical conversations. The task graph operation configuration includes the task graph version, task graph namespace, current node allowed event set, current node declared slots, node standard answer, node type, large model capability switch, prompt word scenario configuration, node answer configuration, and model call timeout. Based on the customer's current round transcribed text, basic intent, slots, and route recognition are performed to obtain the basic task answer. The basic task answer includes: the original current node, the previous node, basic candidate events, identified slots, basic interface parameters, interface execution results, basic answer, basic route status, and basic intent. The basic route status is unique valid route, rejection, low confidence, or candidate conflict. Based on the task graph operation configuration and the basic task answer, the completeness of the required entities for the current node is determined.
[0105] The process then proceeds to the enhancement trigger judgment stage, which is based on rejection, low confidence, and entity missingness. If a high-confidence base result is selected, this branch skips the model; if multi-stage enhancement is performed, this step completes the closed-loop processing of context, intent, and entity. Specifically, based on the base routing state and the completeness of the required entities for the current node, a first judgment result is obtained. If the first judgment result is negative, the final task answer for this round is determined based on the base task answer. If the first judgment result is positive, the corresponding large-scale multi-stage semantic enhancement is performed on the customer's text for this round based on the task graph running configuration and the historical session, resulting in an enhancement result. The multi-stage semantic enhancement includes contextual question rewriting, process intent enhancement, and entity processing; the enhancement result includes standardized questions, enhanced candidate events, and enhanced candidate entities.
[0106] The two outputs are fed into the task graph constraint fusion, where constraints on events, slots, and target nodes are processed. Specifically, if no large-scale multi-stage semantic enhancement is performed, the basic task answer is fed into the task graph constraint fusion. Under the constraints of the event whitelist and slot whitelist of the current task graph, the basic task answer is validated and an answer is selected to determine the final task answer for this round. If large-scale multi-stage semantic enhancement is performed, under the constraints of the event whitelist and slot whitelist of the current task graph, the basic task answer and the enhancement result are fused and adjudicated to determine the final task answer for this round.
[0107] After fusion, the result enters the dialogue generation and verification process. This stage involves the system carrying the intelligent outbound call dialogue control method processing variables and instructions. Finally, state write-back is performed, simultaneously completing call link tracing. In case of timeout or unauthorized access exceptions, the basic result and standard dialogue are used as fallback, directly sent to the state write-back module to complete session updates and link tracing. Specifically, based on the final task answer, a voice script is generated, and the final current node, the valid slots, and the response content are written back to the session context.
[0108] In one exemplary embodiment, this application provides another intelligent outbound call dialogue control method. This application embodiment can advance the node response mode and incorporate it into the joint condition of the enhancement trigger judgment; in the enhancement trigger stage, that is, in combination with the node response mode, it is determined whether to start the large model enhancement to achieve unified arrangement of understanding and dialogue ability, specifically including the following steps 401 to 410: Step 401: Obtain the task graph operation configuration for this round, the client's transcribed text for this round, and historical sessions; the task graph operation configuration includes the task graph version, task graph namespace, set of allowed events for the current node, declaration slots for the current node, standard answer for the node, node type, large model capability switch, prompt word scenario configuration, node answer configuration, and model call timeout.
[0109] Step 402: Based on the client's current transcribed text, perform basic intent, slot and route recognition to obtain the basic task answer; the basic task answer includes: original current node, previous node, basic candidate event, recognized slot, basic interface parameters, interface execution result, basic answer, basic route status and basic intent; the basic route status is unique valid route, rejected, low confidence or candidate conflict.
[0110] Step 403: Based on the task graph running configuration and the basic task answer, determine the completeness of the required entities for the current node.
[0111] Step 404: Based on the rejection state, basic confidence level, candidate event conflict, completeness of the required entity in the current node, and node response mode, determine whether to trigger large model enhancement.
[0112] Step 405: If large model enhancement is not triggered, the basic task answer is directly fed into the task graph constraint fusion.
[0113] Step 406: If large-scale model enhancement is triggered, the corresponding large-scale model processing begins. Based on the client's current text, historical conversations, allowed events at the current node, slots to be collected, and the hit knowledge and confirmed entities selected during the identification of basic task answers, the process proceeds according to the large-scale model enhancement trigger result. This includes contextual question rewriting, process intent enhancement, entity extraction and self-clarification, entity verification and confirmation, resulting in the enhancement result. The enhancement result includes: standardized questions, enhanced candidate events, and enhanced candidate entities. Only events and slots declared in the task graph can be considered valid candidates.
[0114] Step 407: Compare the enhanced candidate events with the set of allowed events of the current node, compare the enhanced candidate entities with the declared slots and verification rules of the current node, and combine the basic routing state and the completeness of the required entities of the current node to make a fusion decision to obtain the final task answer determined in this round; the final task answer includes: the final current node, the target node, the final intent, the valid slots, the interface parameters, and the answer type; if it cannot be uniquely determined, obtain a clarification or rejection result to keep the current node.
[0115] Step 408: Based on the final task answer, node standard answer, final hit knowledge, final confirmed entity, and historical conversation, generate an intelligent response according to the node response pattern, or refine the selected script according to the `usePolish` configuration of the node answer to obtain the node's natural language answer. If the model service is not enabled, the call times out, returns an empty string, or the format is invalid, the original node standard answer, the rejected answer, or the over-rotation answer is obtained as a fallback result. When there are multiple intent responses, the answer list is formed by filtering out unmatched intents according to the order of the intent response list pre-configured in the final task answer; when the intelligent response or refinement has already integrated intent knowledge and node goals, the answer list is generated based on the intelligent response or refinement result, without repeatedly attaching the original node script.
[0116] Step 409: Based on the node's natural language answer and answer list, the interface variable and slot variable are replaced on the system side, and the TTS text and call control instructions are parsed and whitelisted to obtain a voice script arranged in playback order that can be recognized by the call execution terminal.
[0117] Step 410: Based on the spoken text and the final current node, target node, final intent, valid slots, interface parameters, and answer type determined in this round, assemble the final current node, target node, final intent, valid slots, interface parameters, answer list, voice script, and various large model call states into a structured dialogue result, and write the final node, valid slots, and actual responses into the session state to obtain the context required for the next round of task graph execution.
[0118] It should be noted that if any major model enhancement results in a timeout, empty return, structure parsing failure, or event overreach, the incomplete enhancement result will be discarded. Instead, the basic task answer, node standard script, rejection or over-rotation script will be used, and the final adopted result and abnormal status will be recorded together.
[0119] Specifically, the process involves: acquiring the published task graph, model capability switches, and prompt word configurations to form the task graph execution configuration; acquiring the customer's current transcribed text and historical conversations, performing basic intent, slot, and route recognition to obtain basic task answers; determining whether to trigger large model enhancement based on rejection, low confidence, candidate conflict, entity missing, and node response patterns to obtain standardized questions, candidate events, candidate entities, or intelligent dialogue; integrating the basic and enhanced results into the event and slot whitelists of the current task graph for fusion and adjudication to obtain the determined nodes, intents, valid slots, and answer types; performing variable substitution, voice command parsing, and exception fallback to obtain structured dialogue results that can be broadcast by the call execution end, and writing the final state back as the context for the next round.
[0120] The difference between this embodiment and the aforementioned basic embodiment is that the node response mode is added as a joint condition for enhanced triggering judgment. Even if the basic identification obtains a unique valid route and the required entities are complete, as long as the current node is configured in intelligent response or speech refinement mode, a multi-stage semantic enhancement branch can still be triggered; however, the multi-stage semantic enhancement module still only completes contextual question rewriting, process intent enhancement, and entity closure processing, and does not output response speech; intelligent response and speech refinement are still executed in subsequent step 408, realizing the orchestration and linkage of semantic understanding and speech capabilities.
[0121] This application adds on-demand large model processing capabilities to the existing task graph dialogue chain. The task graph is still responsible for defining the current node, triggerable events, target nodes, and word slots to be collected; the large model is responsible for handling issues such as cross-turn referencing, ambiguity of process intent, missing entity information, and naturalization of node phrasing. A dialogue first generates a task graph execution result from the basic recognition chain. Only when the execution result or node configuration meets the enhancement conditions will it enter the corresponding large model processing. The enhancement result then returns to the task graph to complete routing and answer selection.
[0122] The embodiments of this application have the following beneficial effects: 1. Reduce erroneous task graph jumps and unauthorized responses: Model candidates are limited to the current node event, slots, and target node. Even if the model provides additional suggestions, content outside the whitelist will not change the task graph state. After answering a customer's interrogation, the response still carries the current node's target, and the robot will not lose the necessary confirmation steps due to an open-ended generation.
[0123] 2. Limit large model calls to only those rounds that are truly necessary: High-confidence base results continue through the original process; calls are only made for rejection, low-confidence, ambiguity, and incomplete entities, avoiding the fixed delays caused by each round of model requests. Capabilities are broken down into stages; entity questions only call entity-related prompts, avoiding the need to repeatedly execute complete open-ended question-and-answer sessions.
[0124] 3. Key entities are transformed from identification results into usable business data: Entities only become effective after verification and necessary customer confirmation. Interface nodes cannot read unconfirmed values, reducing errors in date, amount, and organization name entries. If a customer denies the data, the original value and denial record are retained and returned to the collection step, avoiding the inability to trace why the inquiry was repeated after overwriting.
[0125] 4. Enhanced service ensures the original workflow continues to function even in the event of a failure: Timeouts, empty returns, and parsing failures all revert to the basic recognition, standard dialogue, or task graph rejection branch, preventing prolonged silence due to model unavailability. Enhancements are disabled by default in older scenarios; after upgrading, it's not necessary to completely modify the entire task graph at once, but rather to gradually enable them according to business needs.
[0126] 5. Facilitates pinpointing why a particular result is used in a given round: The system associates and records the call status and processing results of stages such as question rewriting, process intent, slot enhancement, entity verification, entity confirmation, and intelligent response with the current round of dialogue. When troubleshooting anomalies, this information can be used to determine whether the current round used the basic or enhanced result, and whether the model call timed out or returned an error. Prompt words are loaded from cached or persistent configurations according to namespaces. When a capability switch is not enabled, a valid prompt word does not exist, or a model call fails, the original result from the task graph continues to be used. This allows enhanced capabilities to be enabled gradually according to scenarios and facilitates pinpointing configuration issues specific to a particular scenario.
[0127] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 4As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data used in the process of implementing an intelligent outbound call dialogue control method. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an intelligent outbound call dialogue control method.
[0128] Figure 4 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0129] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0130] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0131] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.
[0132] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0133] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0134] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.
[0135] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for controlling intelligent outbound call dialogue, characterized in that, include: Obtain the task graph execution configuration for this round, the client's transcribed text for this round, and historical sessions; the task graph execution configuration includes the task graph version, task graph namespace, set of allowed events for the current node, declaration slots for the current node, standard answer for the node, node type, large model capability switch, prompt word scenario configuration, node answer configuration, and model call timeout; Based on the customer's current transcription text, perform basic intent, slot, and route recognition to obtain the answer to the basic task; The basic task answer includes: the original current node, the previous node, basic candidate events, identified slots, basic interface parameters, interface execution results, basic answer, basic routing status, and basic intent; the basic routing status is unique valid route, rejected, low confidence, or candidate conflict; Based on the task graph runtime configuration and the basic task answers, determine the completeness of the required entities for the current node; Based on the basic routing state and the completeness of the required entities of the current node, it is determined whether to perform multi-stage semantic enhancement, and a first judgment result is obtained; If the first judgment result is negative, then the final task answer for this round is determined based on the basic task answer; the final task answer includes: the final current node, the target node, the final intent, the valid slots, the interface parameters, and the answer type; If the first judgment result is yes, then based on the task graph running configuration and the historical session, the customer's current round of text is subjected to corresponding large model multi-stage semantic enhancement to obtain the enhancement result; the multi-stage semantic enhancement includes context question rewriting, process intent enhancement and entity processing; the enhancement result includes standardized questions, enhanced candidate events and enhanced candidate entities; Under the constraints of the event whitelist and slot whitelist in the current task graph, the basic task answer and the enhanced result are fused and adjudicated to determine the final task answer for this round; Based on the final task answer, a voice script is generated, and the final current node, the valid word slots, and the response content are written back to the session context as input for the next round of task graph execution.
2. The intelligent outbound call dialogue control method according to claim 1, characterized in that, Based on the task graph runtime configuration and the historical sessions, the current round of text from the customer is subjected to corresponding large-model multi-stage semantic enhancement to obtain the enhancement result, which specifically includes: Under the constraints of the large model capability switch in the task graph runtime configuration, selectively execute one or more of the following sub-capabilities: When the client's current text contains pronouns and omitting sentences, the contextual question rewriting is performed, and the client's current text is semantically completed by combining the historical conversation to obtain a standard question. When the basic routing state is rejection, low confidence, or candidate conflict, the execution process intent is enhanced to obtain enhanced candidate events. When the completeness representation of the required entity of the current node contains uncollected required slots, missing entity processing is performed to obtain enhanced candidate entities; The canonical question, enhanced candidate event, and enhanced candidate entity together constitute the enhanced result.
3. The intelligent outbound call dialogue control method according to claim 2, characterized in that, The missing entity processing specifically includes: When it is detected that there are uncollected required word slots in the original current node, the entity is triggered to clarify and output follow-up questions about the missing information to the calling customer; Obtain the reply text from the calling customer and extract the entity value. Perform format and semantic validation on the entity value according to the preset entity validation rules and generate the validation result. If the verification result is a verification failure, the corresponding verification failure branch in the task graph is triggered, and the original current node is maintained. If the verification result is successful, the candidate entity corresponding to the entity value is placed in the pending confirmation state, and an entity confirmation query is initiated to the calling customer. Obtain the caller's response regarding the entity to be confirmed; If the calling customer replies with confirmation, the candidate entity corresponding to the entity value is marked as valid and updated to the slot list; If the customer denies the call, determine whether the current entity collection round has reached the node's preset maximum allowed round; If the maximum allowed number of rounds has not been reached, return to the step "obtain the reply text of the calling customer and extract the entity value, perform format verification and semantic verification on the entity value according to the preset entity verification rules, and generate the verification result"; When the maximum allowed number of rounds has been reached, the corresponding super-round branch of the task graph is triggered; Entities that are in a pending confirmation state, have failed verification, or have been rejected by the customer will not be written into the slot list and will be prohibited from flowing into subsequent interface nodes.
4. The intelligent outbound call dialogue control method according to claim 1, characterized in that, Under the constraints of the event whitelist and slot whitelist in the current task graph, the basic task answer and the enhanced result are fused and adjudicated to determine the final task answer for this round, specifically including: The enhanced candidate event is compared with the set of allowed events of the current node. The enhanced candidate event is determined to be a valid event only if the enhanced candidate event belongs to the set of allowed events of the current node and the target node corresponding to the enhanced candidate event exists. The enhanced candidate entity is compared with the current node declaration slot. The enhanced candidate entity is determined as a valid entity only if the entity name of the enhanced candidate entity belongs to the current node declaration slot and the entity value of the enhanced candidate entity meets the preset format validation rules or business validation rules. Determine whether the basic candidate event and the valid event point to the same task graph event; If the basic candidate event and the valid event point to the same task graph event, then the task graph event is determined as the triggering event for this round; If the basic candidate event and the valid event point to different task graph events, and the basic routing state is a unique valid route, then the basic candidate event is determined as the triggering event for this round. If the basic candidate event and the valid event point to different task graph events, and the basic routing state is rejection, low confidence, or candidate conflict, then the triggering event for this round is determined from the set of allowed events of the current node by combining the completeness of the required entity of the current node and the actual routing state of the task graph. When the triggering event in this round cannot be uniquely determined, the original current node remains unchanged, and the clarification statement, the denial statement, or the re-collected statement is selected as the answer type for this round; When the triggering event for this round is uniquely determined, the final current node, target node, final intent, valid slots, interface parameters, and answer type for this round are determined based on the triggering event for this round.
5. The intelligent outbound call dialogue control method according to claim 1, characterized in that, Based on the final task answer, a speech script is generated, and the final current node, the valid slots, and the response content are written back to the conversation context as input for the next round of task graph execution. Specifically, this includes: Based on the node answer configuration, determine whether to call the large model service to execute node response generation, and obtain a second determination result; If the second judgment result is yes, the node natural language answer is obtained by assembling the model request based on the final task answer, the node standard answer, the final hit knowledge, the final confirmed entity, the custom prompt content, and the historical session, and calling the large model service to execute node response generation; the node response generation supports intelligent reply mode and speech polishing mode; the final hit knowledge and the final confirmed entity are the final business knowledge and the effective entity of this round obtained after the final task answer is determined, respectively, and are used for the final response generation of the large model; If the second judgment result is negative, then the node's standard answer is directly obtained as the node's natural language answer; The interface variables and slot variables in the natural language answer of the node are replaced by the system, and the TTS text and call control instructions in the replaced script are parsed and whitelisted to generate a voice script. The final current node, the valid slots, and the response content are written back to the session context as input for the next round of task graph execution.
6. The intelligent outbound call dialogue control method according to claim 5, characterized in that, The method further includes: When the large model service is not invoked, the model call times out, or the model returns an abnormal result, the node's standard answer is used as the output message; When the final task answer parsing fails, the model returns an empty value, or the prompt words cannot be obtained, the basic task answer, the node standard script, the node rejection answer, or the super-rotation answer are used to continue execution. When context question rewriting fails, the original customer question is retained for basic identification, blocking the abnormal spread of the enhanced link and avoiding the failure of the entire round of dialogue due to a single point of failure.
7. The intelligent outbound call dialogue control method according to claim 5, characterized in that, The large model service call runs in an independent execution task and waits for the result according to the timeout configured for the scenario; the model-generated content only replaces the natural language expression and does not change the nodes, intents, slots and interface execution results already determined by the task graph.
8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the intelligent outbound call dialogue control method according to any one of claims 1-7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent outbound call dialogue control method according to any one of claims 1-7.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent outbound call dialogue control method according to any one of claims 1-7.
