APP dialogue type service reaching method and system based on large model intention understanding

Abstract

The invention relates to the technical field of large model intention understanding, and discloses an APP dialogue type service reaching method and system based on large model intention understanding. The method comprises the steps of receiving a natural language text input by a user on an APP dialogue interface and performing large model semantic analysis to obtain service semantic data; inputting the business semantic data into an intention recognition model for business intention understanding to obtain business intention data; performing multi-agent cooperative process planning to obtain execution process data; carrying out interactive confirmation on the service parameters to obtain service execution parameters; transmitting the service execution parameters to corresponding service tool interfaces for calling execution to obtain a service tool calling result, and performing intelligent analysis and card rendering on the service tool calling result to obtain display card data. According to the method, the real business intention of the user is accurately understood, and the problems of execution efficiency and stability of a traditional single interface calling mode in a complex business scene are solved.

Description

Technical Field

[0001] This invention relates to the field of large model intent understanding technology, and in particular to a method and system for APP conversational business outreach based on large model intent understanding. Background Technology

[0002] In the import and export trade sector, business processes are complex and involve multiple stages, including customs declaration, foreign exchange settlement, logistics tracking, qualification review, and supply chain finance, each containing numerous service details. To cover these functions, related import and export trade apps often need to design a large number of functional modules and page entry points to meet users' operational needs at different stages. As functions continue to expand, the overall architecture of the app becomes lengthy and complex. Users often have to repeatedly jump between multiple levels of menus and pages to find the corresponding business entry point, resulting in cumbersome operation and a poor user experience. Even if existing apps provide search or navigation functions, they can only perform simple keyword matching and cannot truly understand the user's business intent and semantic expression. Traditional interaction methods are based on preset fixed process templates and static form filling, lacking intelligent intent understanding capabilities and unable to personalize process planning and parameter confirmation according to the user's specific business scenarios and needs. Summary of the Invention

[0003] The main objective of this invention is to provide a method and system for reaching users through conversational business based on large-scale model intent understanding. This invention achieves accurate understanding of users' true business intent and solves the problems of execution efficiency and stability in complex business scenarios of traditional single interface call methods.

[0004] To achieve the above objectives, this invention provides a method for conversational business outreach in apps based on large-scale model intent understanding, comprising the following steps: Receive natural language text input by the user in the APP dialogue interface, and perform large model semantic parsing on the natural language text to obtain business semantic data; The business semantic data is input into the intent recognition model to understand the business intent, thus obtaining business intent data; Based on the business intent data, multi-agent collaboration process planning is performed to obtain execution process data; Based on the execution process data, interactive confirmation of business parameters is performed to obtain business execution parameters; The business execution parameters are passed to the corresponding business tool interface for execution, the business tool call result is obtained, and the business tool call result is intelligently analyzed and card rendered to obtain display card data.

[0005] Optionally, in a first implementation of the first aspect of the present invention, receiving natural language text input by the user in the APP dialogue interface and performing large-scale semantic parsing on the natural language text to obtain business semantic data includes: It receives text or voice signals input by the user in the APP dialogue interface. When the input is a voice signal, it performs probability calculation and conversion from speech to text through the ASR acoustic model and language model to obtain natural language text. The natural language text is input into a pre-built dictionary of import and export trade terms for positive maximum matching to obtain term annotation text; The terminology-annotated text is input into a large language model for token encoding and multi-head attention mechanism calculation to obtain a semantic feature matrix; Based on the semantic feature matrix and the user's historical business records, contextual semantic enhancement calculations are performed to generate business semantic data.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the step of inputting the terminology-annotated text into a large language model for token encoding and multi-head attention mechanism calculation to obtain a semantic feature matrix includes: The terminology-annotated text is subjected to vocabulary-based token decomposition and vector mapping to obtain a token vector sequence; The token vector sequence is transformed linearly to generate a query matrix, a key matrix, and a value matrix, and the dot product attention score of the query matrix and the key matrix is ​​calculated. The attention output vector is obtained by multiplying the dot product attention score after softmax normalization with the Value matrix. The attention output vector is subjected to nonlinear transformation and layer normalization by a feedforward network, and then residually connected with the token vector sequence to obtain the semantic feature matrix.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the step of inputting the business semantic data into the intent recognition model for business intent understanding to obtain business intent data includes: The business semantic data is input into the intent recognition model to activate the intent recognizer and the business classifier, respectively. The intent recognizer performs multi-level semantic reasoning on the business semantic data to obtain intent recognition information. The intent recognition information is input into the business classifier, and the main classification is performed according to document query, business operation, business matching and information intelligence. Fine-grained sub-classification is performed under each main category to obtain business classification information. The intent recognition information and the business classification information are fused and standardized to obtain business intent data.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the step of performing multi-level semantic reasoning on the business semantic data through the intent recognizer to obtain intent recognition information includes: The business semantic data is input into the surface semantic analysis layer of the intent recognizer to identify operation instructions and obtain a surface intent vector. The surface intent vector is input into the deep semantic reasoning layer of the intent recognizer for contextual association analysis to obtain deep intent features; The deep intent features are input into the business knowledge matching layer of the intent recognizer for semantic graph mapping to obtain the knowledge matching result; The surface intent vector, the deep intent features, and the knowledge matching results are fused to generate intent recognition information.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the step of performing multi-agent collaborative process planning based on the business intent data to obtain execution process data includes: The business intent data is input into the intelligent agent scheduler for business complexity analysis to obtain a scheduling execution plan; Based on the aforementioned scheduling execution scheme and combined with the import and export trade business knowledge graph, constraint conditions are retrieved to obtain constraint analysis results; Based on the constraint analysis results, tool call sequences are generated and execution order is optimized to obtain a tool orchestration scheme; The tool orchestration scheme is converted into a standard workflow definition format, the monitoring parameters and checkpoint triggering conditions of the replay observer are configured, and the execution flow data is generated.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the step of generating a tool call sequence and optimizing the execution order based on the constraint analysis results to obtain a tool orchestration scheme includes: Based on the business type identifier and operational requirements in the constraint analysis results, a preset tool library is queried to obtain a set of candidate tools; Perform prerequisite condition checks and topological sorting on each tool node in the candidate tool set to obtain the initial call sequence; Based on the initial call sequence and the timeliness constraints and business priorities in the constraint analysis results, the execution order is rearranged and optimized to obtain the optimal execution sequence; Anomaly detection configuration and alternative scheme design are performed for each tool node in the optimal execution sequence to obtain the tool orchestration scheme.

[0011] Optionally, in a seventh implementation of the first aspect of the present invention, the interactive confirmation of business parameters based on the execution process data to obtain business execution parameters includes: Parse the parameter status data from the execution process data; For the missing parameter items in the parameter status data, perform an automatic completion operation to obtain parameter completion data; By combining the low-confidence parameters and completely missing required parameters in the parameter completion data, a targeted parameter confirmation request is sent to the user and interactive feedback data is received. The interactive feedback data and the parameter completion data are combined to form the business execution parameters.

[0012] Optionally, in the eighth implementation of the first aspect of the present invention, the step of passing the business execution parameters to the corresponding business tool interface for execution, obtaining the business tool call result, and performing intelligent analysis and card rendering on the business tool call result to obtain display card data includes: The business execution parameters are passed to the corresponding business tool interface for coordinated invocation, and the business tool invocation result is obtained. The results of the business tool calls are input into the replay observer for real-time execution monitoring and anomaly detection to obtain the monitoring and processing results. Based on the monitoring and processing results, business data integrity verification and logical consistency analysis are performed to obtain intelligent analysis data. Based on the intelligent analysis data, a visualization interface is constructed and cards are rendered to obtain the displayed card data.

[0013] This invention also provides an APP conversational business outreach system based on large model intent understanding, comprising: The semantic parsing module is used to receive natural language text input by the user in the APP dialogue interface, and perform large-scale model semantic parsing on the natural language text to obtain business semantic data. The business intent understanding module is used to input the business semantic data into the intent recognition model to understand the business intent and obtain business intent data. The process planning module is used to plan the multi-agent collaboration process based on the business intent data to obtain the execution process data. An interactive confirmation module is used to interactively confirm business parameters based on the execution process data to obtain business execution parameters; The execution module is used to pass the business execution parameters to the corresponding business tool interface for execution, obtain the business tool call result, and perform intelligent analysis and card rendering on the business tool call result to obtain display card data.

[0014] In summary, this invention, through token encoding and multi-head attention mechanisms in a large language model, can deeply understand the semantic features and contextual relationships in the natural language text input by the user. Combined with a dictionary of import and export trade terminology, it achieves accurate business entity identification, overcoming the limitations of traditional keyword matching and realizing an accurate understanding of the user's true business intent. Through a three-layer architecture consisting of a surface semantic analysis layer, a deep semantic reasoning layer, and a business knowledge matching layer, it can simultaneously identify the user's explicit operational instructions and implicit business objectives. Combined with semantic matching using an import and export trade business knowledge graph, it achieves accurate intent classification and fine-grained identification for complex business scenarios. Through an intelligent agent scheduler and a multi-agent collaboration mechanism, it can dynamically generate personalized workflows based on different business types, complexity, and constraints. Combining topological sorting algorithms and critical path methods to optimize the execution order, it overcomes the limitations of traditional fixed process templates and achieves intelligent business process orchestration. Through an intelligent inference module and personalized interaction strategies, it can automatically identify and complete missing business parameters and dynamically adjust the interaction method according to the user's business proficiency, achieving a technological breakthrough from static form filling to conversational dynamic information collection, significantly simplifying the user's operation process. By employing a replay observer and anomaly detection mechanism, the system enables parallel and serial coordinated invocation of multiple business tools, featuring intelligent retries, alternative service switching, and real-time monitoring capabilities. This addresses the efficiency and stability issues of traditional single-interface invocation methods in complex business scenarios. Leveraging the business understanding capabilities of a large model, the system performs integrity verification and logical consistency analysis on tool invocation results. Combined with the user's original intent, it generates targeted business insights and operational suggestions, which are then visualized through a card generation module. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the steps of an APP conversational business outreach method based on large model intent understanding in one embodiment of the present invention; Figure 2 This is a block diagram of an APP conversational business outreach system based on large model intent understanding in one embodiment of the present invention.

[0016] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Reference Figure 1This embodiment provides a method for conversational business outreach to apps based on large-scale model intent understanding, including the following steps: S1 receives natural language text input by the user in the APP dialogue interface, performs large-scale semantic parsing on the natural language text, and obtains business semantic data. The system receives user input data via a multimodal input module deployed in the app's dialogue interface. This data can be typed text or real-time recorded speech signals. When the user input is speech, the integrated Automatic Speech Recognition (ASR) module processes the speech signal in real time. The ASR module is composed of an acoustic model and a language model. The acoustic model extracts phoneme features and matches articulatory units based on the input waveform signal, while the language model models the language probability of word combinations within a time-series prediction framework. Through joint calculation of the audio frame sequence and the speech vocabulary probability space, it achieves the conversion from speech to text, outputting natural language text. A pre-built dictionary of import and export trade terms is invoked. This dictionary is constructed based on industry standard terms, common abbreviations, and key business process vocabulary. A forward maximum matching algorithm is used to segment and match consecutive word groups in the text. Preliminary annotation is performed on the original text, identifying terms in areas such as "customs declaration," "letter of credit," "customs clearance status," and "outbound instructions," generating term-annotated text with a term tag structure. The terminology-annotated text is input into a large language model built on the Transformer architecture for deep semantic analysis. The large language model performs token-level encoding on the input text, converting each word into an embedding vector. Through the collaborative action of multiple attention heads, a multi-head attention mechanism is used to globally model the word order context information, calculating a high-dimensional semantic feature matrix representing the semantic structure of the text content. Based on this semantic feature matrix and combined with the user's historical business interaction records in the app, including past order numbers, frequently used query types, and preferred data dimensions, context-enhanced reasoning is performed on the semantic feature matrix. By introducing user profiles and behavioral sequences, attention bias adjustments, entity weight amplification, and role semantic alignment are performed to generate business semantic data with user business context awareness capabilities.

[0019] S2, input the business semantic data into the intent recognition model to understand the business intent and obtain the business intent data; Specifically, the intent recognition module receives business semantic data and automatically triggers two collaborative sub-components: the intent recognizer and the business classifier. The intent recognizer performs deep semantic reasoning on the semantic data to extract the user's implicit operational intent, while the business classifier performs category classification based on the identified intent clues, categorizing the data according to the business scenario. The business semantic data is input into the intent recognizer, which utilizes a multi-layer semantic reasoning network built on a Transformer architecture. Through self-attention and semantic role embedding mechanisms, the input semantic vector is evolved and abstracted layer by layer. A domain knowledge graph is introduced into the deep reasoning network as an auxiliary information source, enabling multi-dimensional analysis of the semantic, behavioral, and logical layers of intent behind the user input, which implicitly includes business goals, operational tendencies, and tone. Structured intent recognition information is output, including target action, object of interest, execution method, and confidence score. The intent recognition information is passed as intermediate input to the business classifier. Based on the preset set of main intent categories, the business objectives are mapped to four primary classification structures: document query (covering static information queries such as customs declarations, clearance status, and order nodes), business operation (covering instruction-driven operations such as warehousing, outbound, status updates, and document submission), business matching (covering interactive matching services such as trade matching, offer push, and offer feedback), and information intelligence (covering information services such as trend analysis, policy dynamics, and market conditions). After completing the main category recognition, the business classifier performs fine-grained subclass determination on the current business intent under the main category. By constructing a multi-branch classification network and a multi-level Softmax structure, specific sub-intents are identified. For example, under the business operation category, it is subdivided into "initiating a remittance request," "confirming a container loading plan," and "submitting qualification materials," generating business classification information. Data fusion processing is performed on intent recognition information and business classification information. A multimodal fusion strategy is used to unify and standardize the output format. All semantic structures are described according to the JSON Schema structure to generate a unified business intent data object, which includes intent category, subclass number, confidence score, associated entity tag, required parameter fields and execution flow mapping identifier.

[0020] S3, based on business intent data, performs multi-agent collaboration process planning to obtain execution process data; It should be noted that the business intent data is input into the intelligent agent scheduler. The scheduler, based on the intent category, confidence score, necessary parameter list, and process mapping identifier within the business intent, combined with the current user's role, system load status, and historical task success rate, performs a multi-dimensional business complexity analysis. Based on the analysis results, it outputs a scheduling execution plan, indicating whether multi-agent collaboration is needed, which business tool interfaces should be activated, and whether an execution backtracking mechanism should be configured. Based on the target process nodes in the scheduling execution plan, the scheduler performs contextual retrieval by linking the business knowledge graph of the import and export trade field. It extracts node-level constraints on the involved business paths, including process dependency constraints (e.g., bill of lading status queries require order number parsing first), timeliness constraints (e.g., customs clearance applications must be submitted before the cut-off date), permission and role constraints (e.g., fund operations must be initiated by the finance department), and compliance constraints (e.g., customs declaration amounts cannot exceed the contract scope), and generates constraint analysis results. Based on the constraint analysis results, the embedded tool sequence planning module combines, filters, and prioritizes available tools. Under the premise of satisfying all business conditions, a tool call sequence is generated. Dependency graph modeling and topology sorting algorithms are used to globally optimize the call order between tools, resulting in a tool orchestration scheme with minimal redundancy, shortest execution time, and highest executability. The tool orchestration scheme is converted into a standardized workflow definition format that conforms to the executable specifications of the process automation execution engine. This format includes elements such as process node definitions, parameter flow rules, execution trigger conditions, and tool reference paths. Monitoring configurations managed by a replay observer are inserted at key process nodes. By setting status checkpoints, timeout trigger conditions, and anomaly response paths, the process is given observability and self-recovery capabilities. Structured process information, scheduling logic, and monitoring strategies are integrated and encapsulated into execution process data.

[0021] S4, based on the execution process data, interactively confirm the business parameters to obtain the business execution parameters; Specifically, the execution process data is structured and parsed to extract the input parameter definitions required for each process node. The status information of each parameter item is analyzed to identify confirmed parameters, automatically inferred parameters, parameters to be completed, and completely missing parameters, forming parameter status data. This data records the identification source, confidence score, format compliance, and whether it is a required field for each parameter item. Automatic completion is performed on parameter items marked as "missing but inferable" in the parameter status data. Completion methods include retrieving similar task parameters from the user's historical business records, performing entity mapping from the current context, or inferring possible values ​​through a business knowledge graph. A confidence score is calculated for each inference result, forming parameter completion data. From the parameter completion data, low-confidence parameters and completely missing but required fields are filtered out and enter the interactive confirmation stage. A semantically based interactive prompt template is generated, initiating a parameter confirmation request to the user and providing options for quick confirmation or modification. Simultaneously, formatted form interactive controls are generated for fields such as time, amount, and port code to improve interaction efficiency. All user feedback information forms interactive feedback data. The interactive feedback data and parameter completion data are merged. During the merging process, the data explicitly confirmed by the user is used first to cover the inference and filling results. All parameter items are checked for completeness, format, and type to generate business execution parameters that conform to the input specifications of the process nodes.

[0022] S5 passes the business execution parameters to the corresponding business tool interface for execution, obtains the business tool call results, and performs intelligent analysis and card rendering on the business tool call results to obtain the display card data.

[0023] Specifically, based on business execution parameters and according to the predefined tool orchestration scheme in the execution process data, parameters are injected into the corresponding business tool interfaces sequentially or in parallel. A coordination and invocation mechanism schedules various query, operation, matching, or intelligence-related service interface modules to complete various business operations such as customs declaration status query, logistics node tracking, goods outbound execution, or market analysis. The system records the start time, parameter content, return results, and invocation status of each tool call in real time, obtaining the business tool call results. These results are then input into a replay observer for real-time execution monitoring and anomaly detection. The replay observer tracks the status of each tool call process, monitoring for issues such as network timeouts, interface anomalies, misaligned return structures, or missing result fields. It also backtracks and compares the execution behavior of similar historical tasks to identify potential risks or abnormal patterns, outputting monitoring and processing results, including execution status codes, anomaly flags, risk scores, and response suggestions. Based on the monitoring results, integrity verification and logical consistency analysis are performed on the returned business data. Integrity verification focuses on confirming whether key fields are complete and whether the data format meets the target card rendering requirements. Logical consistency analysis uses a business knowledge graph and rule engine to determine whether the result content is logically closed and reasonable, such as whether the status progress jumps, whether the amount field is compliant, and whether the time series meets the sequential constraints. Anomalies or warnings during the analysis process are recorded in the intelligent analysis data. The intelligent analysis data is input into the visualization rendering engine, which automatically selects a matching display template based on the business intent type, user role, and result content. The card layout is generated through a component-based construction method, and the execution results are encapsulated into user-friendly display card data by combining visual elements such as status icons, data tables, timelines, and operation buttons. Interactive elements and jump links are also set for key data items.

[0024] In one example, the system receives natural language text input by the user in the app's dialogue interface, performs large-scale semantic parsing on the natural language text, and obtains business semantic data, including: It receives text or voice signals input by the user in the APP dialogue interface. When the input is a voice signal, it performs probability calculation and conversion from speech to text through the ASR acoustic model and language model to obtain natural language text. The natural language text is input into a pre-built dictionary of import and export trade terms and subjected to positive maximum matching to obtain term annotation text; The terminology-annotated text is input into a large language model for token encoding and multi-head attention mechanism calculation to obtain the semantic feature matrix; Contextual semantic enhancement calculations are performed based on the semantic feature matrix combined with the user's historical business records to generate business semantic data.

[0025] In this example, a dialogue interaction module deployed on the client's front-end interface receives user input, which can be either text directly typed by the user or voice signals collected through a microphone. When the user input is detected to be voice, an Automatic Speech Recognition (ASR) engine integrated locally or in the cloud is invoked. The ASR engine consists of a lower-level acoustic model and an upper-level language model working together. The acoustic model performs feature frame extraction, acoustic unit modeling, and temporal alignment based on the input speech waveform to calculate the observation probability distribution between the audio signal and the phoneme sequence. The language model, on the other hand, establishes a context-dependent model of word sequences based on an n-gram or Transformer structure to calculate the conditional probability of different word combinations at the semantic level. Through the joint decoding process of both, the maximum likelihood path is selected from multiple hypothetical paths as the final recognition result, completing the probabilistic conversion from speech signals to readable natural language text. Natural language text is input into a pre-built dictionary of import and export trade terms to perform term recognition and semantic annotation operations. The dictionary integrates high-frequency professional entity entries such as FOB, CIF, customs declaration, bill of lading, and letter of credit based on industry standard terms. By constructing a prefix tree structure based on string length, the algorithm for efficient execution of the forward maximum matching algorithm is achieved. During the traversal of the input text, the longest legal term fragment is matched first, and each matched term entity is marked with part-of-speech tag, term code and its business subclass attribute, generating term annotation text with a clear domain annotation structure. The terminology-annotated text is input into a large language model trained on it to perform token encoding and context modeling operations. The language model adopts a multi-layer decoding structure based on the Transformer architecture. The input text is segmented and mapped to a high-dimensional embedding space. Then, a multi-head attention mechanism is used to calculate the dependency strength between each word and its context. This allows the model to capture deep linguistic relationships such as sequence, causality, subject-predicate structure, and entity modification in different semantic subspaces. A set of semantic feature matrices reflecting the global semantic features of the text is generated by weighted integration of the vectors of each attention head. Each row of the matrix represents the semantic expressive power of a token in the context. Based on the semantic feature matrix and the user's historical business records, such as the last submitted document number, frequently queried field types, and previously followed business nodes, context enhancement fusion is performed. A context injection mechanism based on attention bias reconstruction is used to embed historical behaviors into the semantic space. By semantically strengthening historical target entities and adjusting related fields, the current semantic features are enhanced to represent the user's true intent, generating business semantic data with context relevance and business relevance.

[0026] In one example, the term-tagged text is input into a large language model for token encoding and multi-head attention calculation to obtain a semantic feature matrix, including: The terminology-annotated text is subjected to vocabulary-based token decomposition and vector mapping to obtain a token vector sequence. The token vector sequence is transformed linearly to generate a query matrix, a key matrix, and a value matrix, and the dot product attention score of the query matrix and the key matrix is ​​calculated. After performing softmax normalization on the dot product attention score, it is multiplied by the Value matrix to obtain the attention output vector; The attention output vector is processed by nonlinear transformation and layer normalization of the feedforward network, and then residually connected with the token vector sequence to obtain the semantic feature matrix.

[0027] In this example, the term-annotated text undergoes vocabulary-based token decomposition and vector mapping. A tokenizer segments the original term-annotated text according to the model vocabulary, mapping each semantic unit to a corresponding token. Each token is then mapped to a fixed-dimensional vector representation using an embedding matrix, generating a token vector sequence. Three independent linear transformation weight matrices are applied to the token vector sequence, generating three key intermediate representations: a query matrix, a key matrix, and a value matrix. The query matrix represents the attention request direction of the current token, the key matrix represents the semantic matching reference standard, and the value matrix represents the information content to be conveyed. An element-wise dot product is performed between the query matrix and the key matrix to obtain an attention score matrix, representing the attention intensity of any token in the sequence to all other tokens in the current context. The dot product result is then softmax normalized, so that each row represents the normalized attention probability of the current token to all context tokens. Finally, a matrix multiplication operation is performed between the normalized attention weight matrix and the value matrix to obtain an attention output vector, representing the semantic aggregation result of each token in the entire sentence context, demonstrating context-aware capabilities. The attention output vector is input into a feedforward neural network for nonlinear transformation. The feedforward network consists of two linear layers and an intermediate activation function, which is used to enhance semantic expression and reshape information. At the same time, to ensure the stability and generalization ability of network training, layer normalization is performed on the transformed output, and residual connection is used to add the initial token vector sequence and the processed attention result element by element to obtain the semantic feature matrix.

[0028] In one example, business semantic data is input into an intent recognition model for business intent understanding, resulting in business intent data, including: Input business semantic data into the intent recognition model to activate the intent recognizer and business classifier respectively; The intent recognition information is obtained by performing multi-level semantic reasoning on business semantic data through an intent recognizer. The intent recognition information is input into the business classifier, and the main classification is performed according to document query, business operation, business matching and information intelligence. Fine-grained sub-classification is then performed under each main category to obtain business classification information. The intent identification information and business classification information are fused and standardized in format to obtain business intent data.

[0029] In this example, business semantic data is input into the intent recognition model, which consists of two parallel sub-modules: an intent recognizer and a business classifier. Upon receiving the input, both functional modules are activated simultaneously, enabling intent parsing and business classification to work collaboratively and reinforce each other. The intent recognizer primarily performs multi-level reasoning on the user's implicit operational purpose and behavioral goals within the semantic data. By introducing a deep semantic reasoning network built on a multi-layered Transformer architecture, multi-granularity attention mechanisms are constructed at the word, phrase, and syntactic structure levels to map business semantic data into the intent semantic space. Combining this with a pre-trained set of intent label embedding vectors, the similarity between the current input and each standard intent is calculated layer by layer. A semantic alignment mechanism is used to identify the mapping relationship between direct commands, implicit intents, and ambiguous expressions in the user's expression, outputting structured intent recognition information containing fields such as intent type, intent confidence, target action, and intent trigger location. Intent recognition information is input into the business classifier as a semantic feature vector. The business classifier operates based on a multi-level classification network architecture. Its first layer is the main category recognition layer, which performs coarse classification based on four main categories defined in the import and export business field: document query (such as customs declaration, shipping schedule, order status), business operation (such as warehousing, outbound, settlement instructions), business matching (such as trade offer, offer matching), and information intelligence (such as market conditions, policy trends). The main category matching model is constructed by connecting intent with business process nodes in the knowledge graph to complete the initial classification from intent to process category. Based on the main category recognition, the sub-category fine classification stage begins. The classifier uses a weighted path recursive classification algorithm to determine the business sub-category based on specific terms, action modifiers, context features, and other information in the semantic features, and outputs fine-grained business labels and their confidence scores. For example, in the "business operation" category, it is specifically divided into multiple sub-operation labels such as "submit export documents", "update exchange rate data", and "initiate letter of credit application". The intent recognition information output by the intent recognizer and the business classification information output by the business classifier are structurally fused. During the fusion process, a semantic weight merging and label consistency detection mechanism is adopted to decouple semantic conflict items and form a unique interpretation path. At the same time, business intent data is generated according to a unified JSON Schema standard format, including fields such as intent type, main business class and subclass, parameter requirement list, process mapping identifier, priority weight and semantic confidence.

[0030] In one example, an intent recognizer performs multi-level semantic reasoning on business semantic data to obtain intent recognition information, including: The business semantic data is input into the surface semantic analysis layer of the intent recognizer to identify the operation command and obtain the surface intent vector. The surface intent vector is input into the deep semantic reasoning layer of the intent recognizer for contextual association analysis to obtain deep intent features. The deep intent features are input into the business knowledge matching layer of the intent recognizer for semantic graph mapping to obtain the knowledge matching results; The surface intent vector, deep intent features, and knowledge matching results are fused to generate intent recognition information.

[0031] In this example, business semantic data is input into the surface semantic analysis layer of the intent recognizer. This layer, jointly constructed with a shallow attention mechanism and a convolutional semantic perception network, performs rapid structural parsing of operational verbs, target entities, and modifier phrases in the input. It identifies the user's direct operational intent expressed on the linguistic surface, such as action types like "query," "submit," "cancel," "modify," and "push," and associates these with corresponding target objects such as "customs declaration," "letter of credit," and "order information" to generate a surface intent vector representing the explicit operational instructions in the user's input. This surface intent vector is then input into the deep semantic reasoning layer within the intent recognizer. The deep semantic reasoning layer, built on a multi-layer bidirectional Transformer encoder, uses a global context modeling mechanism to fuse and analyze the semantic consistency between the user's current input and their historical conversation records, business behavior trajectories, and user profile. Simultaneously, it performs context compensation and logical reasoning for ambiguous descriptions, omitted information, and semantic jumps in the expression, thereby uncovering potential business intents that the user did not directly express but can infer from the context. It outputs deep intent features, including contextual triggering factors, implicit targets, and dialogue dependency chains, enhancing the system's ability to understand unstructured expressions. The deep intent features are input into the business knowledge matching layer of the intent recognizer. This layer is built upon a knowledge graph representing the entire import / export trade process. Through graph embedding alignment and entity path retrieval techniques, the current semantic representation is matched with business nodes, process relationships, and entity semantics in the knowledge graph via multi-hop matching. This identifies the business process nodes, execution path identifiers, and upstream / downstream dependent entities that may correspond to the current intent in the knowledge graph, resulting in structured knowledge matching results, including graph matching paths, node labels, and process reference sequences. Semantic fusion and structural alignment are then performed on the surface intent vector, deep intent features, and knowledge matching results. An attention-weighted fusion mechanism is used to perform unified semantic mapping on the three input sources, and a unified intent recognition information vector is constructed through residual connections and layer normalization operations.

[0032] In one example, multi-agent collaboration process planning is performed based on business intent data to obtain execution process data, including: The business intent data is input into the intelligent agent scheduler for business complexity analysis to obtain a scheduling execution plan; Constraint analysis results were obtained by combining the scheduling execution scheme with the knowledge graph of import and export trade business. Based on the constraint analysis results, tool call sequences are generated and execution order is optimized to obtain a tool orchestration scheme; Convert the tool orchestration scheme into a standard workflow definition format, configure the monitoring parameters and checkpoint trigger conditions of the replay observer, and generate execution flow data.

[0033] In this example, business intent data is input into the agent scheduler. The scheduler integrates a task parsing engine and a complexity evaluation model. After receiving the business intent data, it extracts the business subclass tags, parameter completeness, operation type identifiers, priority information, and process mapping identifiers contained therein. Combined with user roles, historical execution frequency, and task path length, it uses a complexity evaluation function to comprehensively calculate the current task in multiple dimensions such as structural complexity, data dependency, and call chain depth to obtain the execution complexity level of the current business task in the agent system. Based on this, a scheduling execution plan is generated, which specifies the type of agent to be called, whether concurrency is allowed, whether task decomposition is required, and whether a process backtracking mechanism is enabled, among other scheduling details. Based on the scheduling and execution plan, the knowledge graph of import and export trade business is linked to perform constraint retrieval operations. This knowledge graph, built on a full lifecycle business model, covers the dependencies, triggering conditions, role permissions, document requirements, time limits, and compliance strategies of each process node, including customs declaration, foreign exchange settlement, logistics, quotations, and negotiations. The scheduler extracts all logical constraint information related to the current task objective through path matching and entity constraint association mechanisms within the graph, forming a structured constraint analysis result. This includes mandatory execution order, necessary preconditions, mutually exclusive nodes, data transfer requirements between nodes, and role operation permission restrictions. Based on the constraint analysis result, a tool call sequence is generated. Combining the existing tool capability registry and constraint information, a sequence generation algorithm based on dependency graph construction and topology sorting optimization is used to determine the call order of each business tool interface on the shortest dependency path. The execution order is optimized based on tool call time, execution failure rate, and data flow tightness, ensuring logical correctness while minimizing the total task execution time, thus generating a tool orchestration scheme. The tool orchestration scheme is converted into a standard workflow definition format that can be directly parsed by the process engine. The format adopts a graph-based workflow description language, recording the business tool name, input and output parameter list, front and back connection edge relationship, execution trigger conditions and failure jump mechanism corresponding to each process node. At the same time, status monitoring configuration managed by the replay observer is inserted at key nodes. The configuration content includes node status synchronization frequency, timeout limit, abnormal trigger action, result verification rules and error backtracking logic, thereby giving the workflow structure complete monitoring and recovery capabilities and generating execution process data.

[0034] In one example, based on the constraint analysis results, tool call sequences are generated and execution order is optimized to obtain a tool orchestration scheme, including: Based on the business type identifier and operational requirements in the constraint analysis results, a set of candidate tools is obtained by querying the preset tool library. Perform prerequisite condition checks and topological sorting on each tool node in the candidate tool set to obtain the initial call sequence; The execution order is rearranged and optimized based on the initial call sequence and the timeliness constraints and business priorities in the constraint analysis results to obtain the optimal execution sequence; Anomaly detection configuration and alternative scheme design are performed for each tool node in the optimal execution sequence to obtain the tool orchestration scheme.

[0035] In this example, a pre-defined tool library is efficiently indexed and queried based on the business type identifier and operation requirement fields from the constraint analysis results. The tool library is structured according to multiple dimensions such as operation domain, input / output type, permission level, and interface stability. During the matching process, the operation type, data processing mode, and business process node required by the current task are compared with the tool's meta-information using multi-dimensional features to extract a set of candidate tools that meet the basic calling conditions. Each tool node in the set contains meta-attribute information such as tool identifier, calling interface address, required parameter structure, output format description, and dependent node description. A precondition check is performed on each tool node in the candidate tool set, including whether the tool has complete input parameters, whether it depends on the execution results of other tools, and whether it is limited by role permissions or interface call frequency. A directed tool call graph is constructed based on the dependencies between nodes. After the graph structure is established, a topological sorting algorithm is used to parse the dependency graph and generate a logically feasible initial call sequence. Based on the initial call sequence, and combined with the timeliness constraints (e.g., some nodes need to be completed before the shipping date or qualification review needs to be completed before settlement) and business priority levels (e.g., emergency customs declaration, near-expiration customs clearance, etc. have higher weights) in the constraint analysis results, a global reordering optimization is performed on the initial call sequence. The optimization goal is to minimize the total process time, increase the probability of key nodes being executed first, and maximize resource utilization under resource competition conditions, thus obtaining the optimal execution sequence. For each tool node in the optimal execution sequence, an anomaly detection mechanism and backup plan design are configured. The anomaly detection configuration includes interface timeout monitoring, return value structure verification, status field validity judgment, and result consistency verification. Simultaneously, backup plans are designed for key nodes, such as automatically switching to a backup service address, switching to a low-precision fast channel, or triggering a manual review assistance mechanism after a call failure. Anomaly jump paths and retry limit are injected into the orchestration structure, resulting in a complete tool orchestration scheme with execution correctness, operational stability, and anomaly resilience.

[0036] In one example, interactive confirmation of business parameters is performed based on execution flow data to obtain business execution parameters, including: Parse parameter status data from the execution process data; Perform an auto-completion operation on missing parameter items in the parameter status data to obtain the parameter completion data; By combining low-confidence parameters and completely missing required parameters in the parameter completion data, a targeted parameter confirmation request is sent to the user and interactive feedback data is received. The interactive feedback data and parameter completion data are combined to form the business execution parameters.

[0037] In this example, the execution flow data is parsed. This data includes the tool interface configurations, parameter definitions, and data dependencies with preceding nodes for each node in the flow. By uniformly extracting the input fields from each node in the flow definition and identifying the parameter name, type requirements, whether it is a required field, default value settings, historical value sources, and source nodes for each field, parameter status data is constructed. This identifies which parameters have valid values, which are inferable, which have low-confidence inferences, and which are completely missing required parameters. After this parsing, the automatic completion stage begins. For fields marked as "inferable completion" in the parameter status data, automatic completion is performed based on information sources such as historical task databases, user profile data, commonly used templates, and context entity cache pools. Methods based on similarity matching, rule constraint judgment, or graph path mapping are used to automatically complete these parameter items. During the completion process, a confidence score is generated for each completed value, and its inference path and reference samples are recorded to constitute parameter completion data. For low-confidence parameters (such as those with confidence levels below the confidence threshold or multiple candidate values) and parameters that are completely missing but necessary for the current task, an interactive confirmation mechanism is implemented. This mechanism automatically generates multi-turn dialogue prompts through the user interface, including parameter requests in natural language, candidate value selection menus, contextual hints for ambiguous parameters, and auxiliary information such as the most recent input history. The complexity of the interactive expression is automatically adjusted based on user habits and business proficiency, and multiple interactive control types are supported to guide users in completing parameter supplementation and confirmation. All user responses are collected as interactive feedback data. Semantic-level merging is performed on the interactive feedback data and parameter completion data. For duplicate fields, a strategy prioritizing the most recent user input is adopted. If the user has not modified the automatically inferred value, the original completion result is retained. Simultaneously, validity markers and confidence scores are recalculated for all parameter items, and format standardization, field deduplication, value range validation, and boundary handling operations are completed to generate a set of business execution parameters.

[0038] In one example, business execution parameters are passed to the corresponding business tool interface for execution, the business tool call result is obtained, and intelligent analysis and card rendering are performed on the business tool call result to obtain the displayed card data, including: The business execution parameters are passed to the corresponding business tool interface for coordination and invocation, and the business tool invocation result is obtained. The results of business tool calls are input into a replay observer for real-time execution monitoring and anomaly detection, and the monitoring and processing results are obtained. Based on the monitoring and processing results, business data integrity verification and logical consistency analysis are performed to obtain intelligent analysis data. Based on the intelligent analysis data, a visualization interface is built and cards are rendered to obtain the data for displaying the cards.

[0039] In this example, business execution parameters are passed to the corresponding business tool interfaces. Based on the preset tool orchestration scheme, the interfaces are coordinated and called. During the coordination and calling process, the triggering conditions of each tool node are determined according to the execution order and data dependency chain in the process definition. If the conditions are met, the target tool is encapsulated and called through a unified tool calling engine. During the call, the input parameters are format-validated, value-domain-standardized, and authenticated to ensure that the parameters are highly matched with the format required by the target interface. Backend services are triggered synchronously or asynchronously to complete business tasks such as order status query, logistics node push, fund instruction execution, and document generation and uploading, and the business tool call results are obtained. The results of business tool calls are input into a replay observer for real-time execution monitoring and anomaly detection. The replay observer reviews each interface call execution by comparing tool call logs, return status codes, response times, and the structure of returned fields with expected values. It uses an embedded anomaly detection module to identify anomalies such as missing data, structural inconsistencies, abnormal business states, logical conflicts, or cross-interface inconsistencies. Simultaneously, it dynamically determines the anomaly level by combining the call context and historical execution trajectory, generating timestamped monitoring results. These results include structured fields such as interface success identifiers, anomaly tags, data validity markers, risk scores, and suggested recovery strategies. Based on the monitoring results, business data integrity verification and logical consistency analysis are performed. Integrity verification focuses on verifying whether all core fields of user concern are returned completely in the results, and whether there are null values, illegal encodings, or field truncation. Logical consistency analysis relies on a business knowledge graph and domain rule base to align the current tool results with the state chain of upstream and downstream processes, verifying whether time sequence, state transitions, and numerical calculations meet process closure conditions. The verification and analysis results are then uniformly constructed into intelligent analysis data. The intelligent analysis data is input into the visualization building module, which calls the corresponding business type card rendering template, performs component-based mapping processing on the results, and presents the structured analysis results to the user in an interactive, easy-to-understand, and highly adaptable visual way through various view elements such as charts, status bars, data lists, trend curves, and operation buttons. Based on the displayed content, auxiliary prompt text and suggestions for the next optional operation are generated to build the display card data.

[0040] Reference Figure 2 This embodiment provides an APP conversational business outreach system based on large model intent understanding, including: Semantic parsing module 1 is used to receive natural language text input by the user in the APP dialogue interface, and perform large-scale semantic parsing on the natural language text to obtain business semantic data. Business intent understanding module 2 is used to input business semantic data into the intent recognition model to understand business intent and obtain business intent data; Process planning module 3 is used to plan multi-agent collaboration processes based on business intent data to obtain execution process data; Interactive confirmation module 4 is used to interactively confirm business parameters based on execution process data to obtain business execution parameters; The execution module 5 is used to pass the business execution parameters to the corresponding business tool interface for execution, obtain the business tool call result, and perform intelligent analysis and card rendering on the business tool call result to obtain the display card data.

[0041] In this embodiment, the specific implementation of each unit in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0042] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, system, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, system, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, system, article, or method that includes that element.

[0043] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for conversational business outreach in apps based on large-scale model intent understanding, characterized in that: include: Receive natural language text input by the user in the APP dialogue interface, and perform large model semantic parsing on the natural language text to obtain business semantic data; The business semantic data is input into the intent recognition model to understand the business intent, thus obtaining business intent data; Based on the business intent data, multi-agent collaboration process planning is performed to obtain execution process data; Based on the execution process data, interactive confirmation of business parameters is performed to obtain business execution parameters; The business execution parameters are passed to the corresponding business tool interface for execution, the business tool call result is obtained, and the business tool call result is intelligently analyzed and card rendered to obtain display card data.

2. The APP conversational business outreach method based on large model intent understanding according to claim 1, characterized in that, The process involves receiving natural language text input by the user in the APP's dialogue interface, performing large-scale semantic parsing on the natural language text to obtain business semantic data, including: It receives text or voice signals input by the user in the APP dialogue interface. When the input is a voice signal, it performs probability calculation and conversion from speech to text through the ASR acoustic model and language model to obtain natural language text. The natural language text is input into a pre-built dictionary of import and export trade terms for positive maximum matching to obtain term annotation text; The terminology-annotated text is input into a large language model for token encoding and multi-head attention mechanism calculation to obtain a semantic feature matrix; Based on the semantic feature matrix and the user's historical business records, contextual semantic enhancement calculations are performed to generate business semantic data.

3. The APP conversational business outreach method based on large model intent understanding according to claim 2, characterized in that, The step involves inputting the terminology-annotated text into a large language model for token encoding and multi-head attention mechanism calculation to obtain a semantic feature matrix, including: The terminology-annotated text is subjected to vocabulary-based token decomposition and vector mapping to obtain a token vector sequence; The token vector sequence is transformed linearly to generate a query matrix, a key matrix, and a value matrix, and the dot product attention score of the query matrix and the key matrix is ​​calculated. The attention output vector is obtained by multiplying the dot product attention score after softmax normalization with the Value matrix. The attention output vector is subjected to nonlinear transformation and layer normalization by a feedforward network, and then residually connected with the token vector sequence to obtain the semantic feature matrix.

4. The APP conversational business outreach method based on large model intent understanding according to claim 1, characterized in that, The step of inputting the business semantic data into the intent recognition model for business intent understanding to obtain business intent data includes: The business semantic data is input into the intent recognition model to activate the intent recognizer and the business classifier, respectively. The intent recognizer performs multi-level semantic reasoning on the business semantic data to obtain intent recognition information. The intent recognition information is input into the business classifier, and the main classification is performed according to document query, business operation, business matching and information intelligence. Fine-grained sub-classification is performed under each main category to obtain business classification information. The intent recognition information and the business classification information are fused and standardized to obtain business intent data.

5. The APP conversational business outreach method based on large model intent understanding according to claim 4, characterized in that, The step of performing multi-level semantic reasoning on the business semantic data through the intent recognizer to obtain intent recognition information includes: The business semantic data is input into the surface semantic analysis layer of the intent recognizer to identify operation instructions and obtain a surface intent vector. The surface intent vector is input into the deep semantic reasoning layer of the intent recognizer for contextual association analysis to obtain deep intent features; The deep intent features are input into the business knowledge matching layer of the intent recognizer for semantic graph mapping to obtain the knowledge matching result; The surface intent vector, the deep intent features, and the knowledge matching results are fused to generate intent recognition information.

6. The APP conversational business outreach method based on large model intent understanding according to claim 1, characterized in that, The process planning for multi-agent collaboration based on the business intent data, to obtain execution process data, includes: The business intent data is input into the intelligent agent scheduler for business complexity analysis to obtain a scheduling execution plan; Based on the aforementioned scheduling execution scheme and combined with the import and export trade business knowledge graph, constraint conditions are retrieved to obtain constraint analysis results; Based on the constraint analysis results, tool call sequences are generated and execution order is optimized to obtain a tool orchestration scheme; The tool orchestration scheme is converted into a standard workflow definition format, the monitoring parameters and checkpoint triggering conditions of the replay observer are configured, and the execution flow data is generated.

7. The APP conversational business outreach method based on large model intent understanding according to claim 6, characterized in that, The step of generating tool call sequences and optimizing execution order based on the constraint analysis results to obtain a tool orchestration scheme includes: Based on the business type identifier and operational requirements in the constraint analysis results, a preset tool library is queried to obtain a set of candidate tools; Perform prerequisite condition checks and topological sorting on each tool node in the candidate tool set to obtain the initial call sequence; Based on the initial call sequence and the timeliness constraints and business priorities in the constraint analysis results, the execution order is rearranged and optimized to obtain the optimal execution sequence; Anomaly detection configuration and alternative scheme design are performed for each tool node in the optimal execution sequence to obtain the tool orchestration scheme.

8. The APP conversational business outreach method based on large model intent understanding according to claim 1, characterized in that, The interactive confirmation of business parameters based on the execution process data to obtain business execution parameters includes: Parse the parameter status data from the execution process data; For the missing parameter items in the parameter status data, perform an automatic completion operation to obtain parameter completion data; By combining the low-confidence parameters and completely missing required parameters in the parameter completion data, a targeted parameter confirmation request is sent to the user and interactive feedback data is received. The interactive feedback data and the parameter completion data are combined to form the business execution parameters.

9. The APP conversational business outreach method based on large model intent understanding according to claim 1, characterized in that, The process involves passing the business execution parameters to the corresponding business tool interface for execution, obtaining the business tool call result, and performing intelligent analysis and card rendering on the business tool call result to obtain display card data, including: The business execution parameters are passed to the corresponding business tool interface for coordinated invocation, and the business tool invocation result is obtained. The results of the business tool calls are input into the replay observer for real-time execution monitoring and anomaly detection to obtain the monitoring and processing results. Based on the monitoring and processing results, business data integrity verification and logical consistency analysis are performed to obtain intelligent analysis data. Based on the intelligent analysis data, a visualization interface is constructed and cards are rendered to obtain the displayed card data.

10. An APP conversational business outreach system based on large-scale model intent understanding, characterized in that, The steps for implementing the APP conversational business outreach method based on large model intent understanding as described in any one of claims 1 to 9 include: The semantic parsing module is used to receive natural language text input by the user in the APP dialogue interface, and perform large-scale model semantic parsing on the natural language text to obtain business semantic data. The business intent understanding module is used to input the business semantic data into the intent recognition model to understand the business intent and obtain business intent data. The process planning module is used to plan the multi-agent collaboration process based on the business intent data to obtain the execution process data. An interactive confirmation module is used to interactively confirm business parameters based on the execution process data to obtain business execution parameters; The execution module is used to pass the business execution parameters to the corresponding business tool interface for execution, obtain the business tool call result, and perform intelligent analysis and card rendering on the business tool call result to obtain display card data.

Images