Conversational application matching method, system, device and medium

Abstract

The application provides a dialogue application matching method, device and medium, which is used to solve the technical problem that the overall efficiency of content creation and workflow execution is reduced in the traditional scheme. The method comprises the following steps: traversing all function nodes in the canvas workflow, parsing fixed constant parameters and batch variable parameters therefrom, and generating variable input rules based on the batch variable parameters; combining the execution logic, the fixed constant parameters and the variable input rules to generate a standardized application; obtaining shared range information configured for the standardized application, and assigning corresponding operation permissions to target objects in the corresponding range based on the shared range information; in response to a task execution instruction of the target object for the standardized application, verifying the operation permission corresponding to the target object, and after the verification is passed, obtaining batch parameters input by the target object based on the variable input rules, combining the batch parameters with the fixed constant parameters, and executing the task in parallel according to the execution logic.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and provides a method, system, device and medium for conversational application matching. Background Technology

[0002] With the development of artificial intelligence technology, AI creation platforms have become an important carrier for daily content production and business processing. In existing AI creation platforms, the use and distribution of tools usually rely on a fixed-category application marketplace model.

[0003] However, existing technologies have significant limitations when dealing with users' complex creative needs. Specifically, when users need to perform specific content generation tasks, they must first locate the required generation tool by manually searching through a complex app store or by simply searching for keywords. After finding the appropriate tool, users still need to face a form-based configuration interface, manually breaking down their creative requirements and filling them in one by one into a large number of complex parameter configuration items.

[0004] This traditional interaction model cannot effectively understand the user's colloquial natural language creative intent, nor can it automatically convert the user's natural language description directly into the execution parameters required by the underlying application. This results in extremely redundant human-computer interaction steps during use, and the manual configuration of parameters is cumbersome and prone to errors, thus severely reducing the overall efficiency of content creation and workflow execution. Summary of the Invention

[0005] This application provides a conversational application matching method, system, device, and medium to solve the technical problem that traditional solutions reduce the overall efficiency of content creation and workflow execution.

[0006] To solve the above problems, the following technical solution is provided: A conversational application matching method, comprising: Acquire the natural language dialogue data to be processed input by the target object in the AI ​​creation platform; Semantic parsing is performed on the natural language dialogue data to extract structured intent features containing target generation parameters; In a preset application library, the target workflow application corresponding to the natural language dialogue data is determined by matching based on the structured intent features. The application library includes general system applications, shared tenant applications, and private user applications. The target generation parameters are extracted from the structured intent features, and the target generation parameters are mapped and filled into the variable input items corresponding to the target workflow application to obtain the target workflow application with configured parameters. In response to the generation task execution instruction for the target workflow application with the configured parameters, the corresponding generation task is executed.

[0007] A conversational application matching system, comprising: The user intent recognition module is used to acquire the natural language dialogue data to be processed input by the target object in the AI ​​creation platform, and to perform semantic parsing on the natural language dialogue data to extract structured intent features containing target generation parameters. An intelligent matching module is used to match the structured intent features in a preset application library to determine the target workflow application corresponding to the natural language dialogue data. The application library includes general system applications, tenant shared applications, and user private applications. The parameter automatic extraction and filling module is used to extract the target generation parameters from the structured intent features and map and fill the target generation parameters into the variable input items corresponding to the target workflow application to obtain the target workflow application with configured parameters. The dialogue interaction module is used to execute the corresponding generation task in response to the generation task execution instruction of the target workflow application with the configured parameters.

[0008] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described.

[0009] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described.

[0010] In the method provided in this embodiment, natural language dialogue data is acquired and structured intent features containing target generation parameters are extracted. Then, the target workflow application is matched within a full application library including system-wide, tenant-shared, and user-private applications. The generation parameters are automatically mapped and filled into the application input fields, and finally, the generation task is executed in response to the command. This establishes an end-to-end data-driven automated link. By introducing the acquisition and semantic parsing of natural language dialogue data, users can describe complex creative needs using everyday spoken language, reducing the technical threshold for tool search and invocation, and improving the directness of human-computer interaction. Matching within a full application library including tenant-shared and private applications overcomes the limitation of traditional conversational generation, which can only call fixed single model capabilities, allowing the reuse of customized workflow assets accumulated by enterprises and individuals. The automatic extraction and mapping of generation parameters based on structured intent features transforms the originally redundant manual form configuration process into a fully automated data transmission flow in the background, reducing parameter omission and error rates while improving the overall collaborative efficiency from requirement reception to task execution. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application 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.

[0012] Figure 1 This is a flowchart illustrating a conversational application matching method in one embodiment of this application; Figure 2 This is a schematic diagram of the structure of a conversational application matching system according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device according to one embodiment of this application. Detailed Implementation

[0013] 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, 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.

[0014] In one embodiment, a conversational application matching method is provided, the method comprising: Step S101: Obtain the natural language dialogue data to be processed input by the target object in the AI ​​creation platform; Step S102: Perform semantic parsing on the natural language dialogue data to extract structured intent features containing target generation parameters; Step S103: In a preset application library, match based on the structured intent features to determine the target workflow application corresponding to the natural language dialogue data, wherein the application library includes system general applications, tenant shared applications and user private applications; Step S104: Extract the target generation parameters from the structured intent features, and map and fill the target generation parameters into the variable input items corresponding to the target workflow application to obtain the target workflow application with configured parameters; Step S105: In response to the generation task execution instruction for the target workflow application with the configured parameters, execute the corresponding generation task.

[0015] Specifically, the aforementioned conversational application matching method can be applied to multimodal content generation, enterprise workflow automation orchestration, or multi-model collaborative scheduling in business application scenarios, without any specific limitations.

[0016] In step S101, the natural language dialogue data to be processed, input by the target user in the AI ​​creation platform, is acquired. The target user can be a registered user of the platform, a content creator, or a business operator within an organization; there are no specific limitations. Natural language dialogue data refers to text information or speech-to-text data input by the target user in a conversational, unstructured manner, directly conveying the target user's original creative needs or business instructions. For example, this dialogue data could be: "Generate 10 e-commerce women's clothing product images, 3:4 aspect ratio, 1080P resolution, with an Instagram-style minimalist and sophisticated feel." This natural language dialogue data can be acquired through real-time network collection via the text input component in the platform's front-end dialogue interface, or through audio input devices that collect speech and convert it into text data using speech recognition technology; there are no specific limitations. In practical applications, this step provides users with an interactive entry point that eliminates the need to perceive complex underlying menu categories.

[0017] In step S102, semantic parsing is performed on the natural language dialogue data to extract structured intent features containing target generation parameters. Semantic parsing refers to the process of using pre-trained natural language processing algorithms or intent recognition models to identify multi-dimensional semantic features and extract key information from unstructured dialogue text. In terms of data meaning, structured intent features refer to a data dictionary or feature vector set composed of multiple preset fields or key-value pairs that reflects the user's core creative needs. Specifically, they may include creation type features, business scenario features, core requirement descriptions, etc. Target generation parameters refer to quantitative indicators, physical control variables, or format constraint standards that directly determine the final content generation result. For example, in the aforementioned example, the extracted generation quantity parameter is "10 images," the size ratio parameter is "3:4," the resolution parameter is "1080P," and the style description parameter is "ins-style minimalist and sophisticated." Through this semantic parsing process, the system transforms vague and discrete everyday spoken expressions into structured control instructions with clear business boundaries and parameter definitions.

[0018] In step S103, based on the structured intent features, a matching process is performed in a preset application library to determine the target workflow application corresponding to the natural language dialogue data. The application library includes system-wide applications, tenant-shared applications, and user-private applications. The application library refers to the computing resource space at the system's underlying layer that centrally stores, manages, and schedules various AI generation models, business processing tools, or customized processes. To adapt to the needs of different security permissions and proprietary business scenarios, the application library is physically or logically divided into three levels: system-wide applications typically refer to basic generation tools or model API interfaces preset by the platform and open to all users across the network; tenant-shared applications typically refer to business flow applications customized and developed by developers within a specific enterprise or organizational tenant and shared only within the current enterprise or organizational structure; user-private applications are private tool assets built by the target individual based on their own habits and intended for personal use only. The target workflow application refers to the specific workflow sequence retrieved and matched from the aforementioned full application library that best satisfies the current structured intent features. This application can be a composite business flow composed of multiple single-model interfaces arranged according to a specific topological relationship. Through this matching across multiple library levels, the system can overcome the control barriers of a single model and directly adapt to and reuse workflow assets already accumulated within the enterprise.

[0019] In step S104, the target generation parameters are extracted from the structured intent features, and these parameters are mapped and filled into the variable input fields corresponding to the target workflow application, resulting in a target workflow application with configured parameters. Variable input fields refer to predefined data type input slots that the API interface or dynamic form relies on when the target workflow application starts running or initiates underlying network communication. Mapping and filling refers to establishing a data correspondence between the extracted target generation parameters and these variable input fields, and silently writing parameter values, text variables, or material pointers into the corresponding interface fields in the system background. For example, the quantity "10" is automatically assigned to the control variable of the image generation node, and the ratio "3:4" is automatically assigned to the format variable of the rendering node. A target workflow application with configured parameters refers to an instantiated workflow application object that has completed full parameter loading in the background and is in a ready state without further manual intervention. This process eliminates the need for users to manually fill out tedious forms a second time.

[0020] In step S105, in response to the generation task execution instruction for the target workflow application with configured parameters, the corresponding generation task is executed. The generation task execution instruction refers to the start command issued by the target object after confirming the application form or card with automatically matched and configured parameters by the system. It can be triggered by the target object physically clicking visual controls such as "Use Now" or "Submit Batch Task" on the front-end interactive interface, or it can be automatically issued by the model after recognizing affirmative intent descriptions in multi-turn dialogues; the specific method is not limited. Executing the corresponding generation task means that after receiving the instruction, the system's underlying task scheduling engine formally starts the various model inference processes or network interface calls within the target workflow application, passes in the loaded parameter data, thereby initiating the underlying distributed computing, and finally returning the content generation result on the front-end interface that highly matches the user's initial natural language requirements.

[0021] In the method provided in this embodiment, natural language dialogue data is acquired and structured intent features containing target generation parameters are extracted. Then, the target workflow application is matched within a full application library including system-wide, tenant-shared, and user-private applications. The generation parameters are automatically mapped and filled into the application input fields, and finally, the generation task is executed in response to the command. This feature processing logic establishes an end-to-end data-driven automated link. By introducing the acquisition and semantic parsing of natural language dialogue data, users can describe complex creative needs using everyday spoken language, reducing the technical threshold for tool search and invocation, and improving the directness of human-computer interaction. Matching within a full application library including tenant-shared and private applications overcomes the limitation of traditional conversational generation, which can only call fixed single model capabilities, allowing the reuse of customized workflow assets accumulated by enterprises and individuals. The automatic extraction and mapping of generation parameters based on structured intent features transforms the originally redundant manual form configuration process into a fully automated data transmission flow in the background, reducing parameter omission and error rates while improving the overall efficiency of content creation and workflow execution.

[0022] In one embodiment, step S120, namely, performing semantic parsing on the natural language dialogue data to extract structured intent features containing target generation parameters, specifically includes: Step S121: Input the natural language dialogue data into a pre-trained semantic parsing model for semantic parsing to obtain the creation type features, business scenario features, core requirement descriptions, and target generation parameters contained in the natural language dialogue data; Step S122: Store the creation type features, business scenario features, core requirement descriptions, and target generation parameters in a structured manner to generate the structured intent features.

[0023] Specifically, since the original text input by the target object often has the characteristics of semantic colloquialism, intertwined multi-dimensional demands and blurred information boundaries, after receiving the data, the system needs to convert these unstructured texts into a standard data format that the underlying computing engine can directly read and recognize.

[0024] In step S121, the system inputs the acquired natural language dialogue data into a pre-trained semantic parsing model. This pre-trained semantic parsing model refers to a text processing model with deep text feature extraction, named entity recognition, and intent classification capabilities. Specifically, it can be fine-tuned based on a general or industry-specific large language model. After receiving the dialogue text, the model performs multi-dimensional decoupling extraction of the word sequence in the hidden feature space, thereby selectively acquiring the creation type features, business scenario features, core requirement descriptions, and target generation parameters contained in the natural language dialogue data.

[0025] The creation type feature is used to identify the user's basic demand threshold or the physical modality of the generated product. Specifically, the model can identify whether the demand belongs to a core type such as image generation, video production, script creation, or audio generation. The business scenario feature is used to identify the vertical application environment in which the target content is ultimately implemented, such as specific business scenarios like e-commerce, short videos, comics, corporate new media, or marketing posters. The core demand description is used to extract the user's core creation goal, subjective effect requirements, or specific style semantics, such as "ins-style minimalist and high-end feel" in the previous example. The target generation parameters refer to quantifiable parameters that can be directly read by the underlying generation model or workflow interface and are used to hard control the content generation specifications, such as the number of images generated (e.g., "10 images"), aspect ratio (e.g., "3:4"), resolution (e.g., "1080P"), or specific quantifiable variables such as duration, model base, and export format.

[0026] As a preferred implementation, to further improve the completeness of intent recognition, the model can also simultaneously acquire implicit constraint features and operational intent features in the dialogue data during the extraction process. Constraint features are used to identify business-level limitations such as delivery time, compliance requirements, and brand guidelines; operational intent features are used to accurately determine the type of user action, identifying whether the user needs to directly generate content, edit existing materials, search for templates, or submit a customized service request, thus achieving accurate routing of user requests at the business action level from the initial interaction stage.

[0027] In step S122, the system structurally stores the creation type features, business scenario features, core requirement descriptions, and target generation parameters (as well as metadata such as possible constraint features and operational intent features) extracted from the model, generating the structured intent features. Technically, structured storage refers to writing the discrete features of each dimension into the system's temporary storage area or feature database according to a preset key-value mapping rule or a standardized data structure (such as JSON or XML format data). Through this structured reassembly, the discretely extracted business features and quantized control parameters are uniformly bound and encapsulated in the same highly cohesive intent feature data body.

[0028] In this embodiment, natural language dialogue data is input into a pre-trained semantic parsing model to deeply analyze and extract core dimensional features such as creation type and business scenario, along with target generation parameters. These features are then uniformly and structurally stored to generate structured intent features. By introducing a pre-trained semantic parsing model to analyze dialogue data from multiple dimensions, various business modalities and quantitative indicators interwoven in users' colloquial expressions can be accurately extracted and purified, thereby accurately revealing the user's true core intent. Furthermore, the extracted creation type features, business scenario features, core requirement descriptions, and generation parameters are structurally stored, transforming highly discrete and disordered natural language expressions into intent feature data volumes with standard data boundaries and fixed field specifications. This not only eliminates the shortcomings of keyword retrieval mode in understanding complex contexts but also provides deterministic data basis for subsequent steps to perform efficient multi-level filtering and matching in the full application library and to accurately map and fill generation parameters to underlying workflow variable input items, significantly improving the accuracy and data flow efficiency of the entire interaction matching chain.

[0029] In one embodiment, the step of matching the structured intent features in a preset application library to determine the target workflow application corresponding to the natural language dialogue data includes: Step S131: Based on the creation type features and business scenario features contained in the structured intent features, filter the applications in the application library and select applications with matching types as a candidate application set; Step S132: For each candidate application in the candidate application set, extract its corresponding functional description, scene tags, and keywords; Step S133: Calculate the vector similarity between the core requirement description contained in the structured intent feature and the functional description, scene tag and keyword corresponding to each candidate application, and extract a preliminary application set from the candidate application set based on the vector similarity; Step S134: Obtain the target object's historical usage habit data, industry data, application usage frequency data, and popular application data within the tenant to which the target object belongs; Step S135: Based on the historical usage data, the industry data, the application usage frequency data, and the popular application data, the applications in the preliminary application set are reordered, and the application with the best ranking is determined as the target workflow application.

[0030] Specifically, to achieve high-precision, low-latency application matching and addressing, after acquiring structured intent features, the system performs a progressive filtering and optimization sorting mechanism within a preset application library accessible to the user. Preferably, the matching and retrieval scope of the application library fully covers official general applications of the platform, enterprise applications shared within the current tenant, and private applications created by the user.

[0031] In step S131, based on the creation type features and business scenario features included in the structured intent features, the applications in the application library are filtered to select applications with matching types as a candidate application set. This filtering is logically a first-level coarse screening of type scenarios. Specifically, the system extracts creation type features (e.g., structured labels representing physical modal classifications such as image generation, video production, script creation, and audio generation) and business scenario features (e.g., scenario codes identifying specific vertical application areas such as e-commerce, short videos, comics, enterprise new media, and marketing posters) from the intent features, using them as hard filtering conditions and performing Boolean logic matching with the preset classifications and scenario labels of each application in the application library. By directly eliminating irrelevant applications, the volume of data to be processed is reduced, and a candidate application set with preliminary classification correspondence is output. This step is suitable for large-scale service environments with a massive number of coexisting applications. Its technical principle lies in quickly reducing the search space through lightweight attribute alignment, avoiding complex calculations performed directly on the entire library.

[0032] In step S132, for each candidate application in the candidate application set, its corresponding functional description, scene tags, and keywords are extracted. The functional description refers to a structured text description characterizing the core business logic, technical capabilities, and expected output of the candidate application; the scene tags are more granular vertical business tags that identify the most suitable application; and the keywords are a set of high-frequency words extracted from the application's feature text. This information is typically persistently stored as metadata in the system database when the application is released or listed, and is directly retrieved by the system after a candidate application is matched.

[0033] In step S133, the vector similarity between the core requirement description contained in the structured intent features and the corresponding functional description, scene tag, and keyword of each candidate application is calculated, and a preliminary application set is extracted from the candidate application set based on the vector similarity. This step belongs to the second layer of semantic similarity matching. In specific processing, the system uses a preset text encoder to transform the unstructured core requirement description (such as detailed descriptions of artistic style, visual effects, etc. in user-spoken expressions) into core requirement feature vectors in a high-dimensional semantic space; similarly, the functional descriptions, scene tags, and keywords corresponding to each candidate application are vectorized to obtain the corresponding application feature vectors. Subsequently, the system calculates the dot product or cosine distance between the core requirement feature vector and each application feature vector to obtain the corresponding vector similarity score. The system sorts the candidate applications in descending order based on the vector similarity and selects candidate applications whose similarity meets preset conditions (as an example, the top ten applications with the highest similarity can be selected, i.e., the Top 10 applications with the highest similarity) to form a preliminary application set. The technical principle behind this step is to deeply mine the semantic relationship between users' subjective effect demands and application functions by calculating the hidden semantic vector space, thus avoiding retrieval omissions caused by literal inconsistencies.

[0034] In step S134, the system acquires the target object's historical usage data, industry data, application usage frequency data, and popular application data within the target object's tenant. Historical usage data refers to behavioral tracking data such as frequently used application types or preferred parameter configurations within a preset time period. Industry data refers to the target object's registered industry affiliation (e.g., new media, e-commerce, short video) or the vertical industry characteristics identified by the system based on its profile. Application usage frequency data refers to the cumulative call frequency of each candidate application at the current user level. Popular application data refers to the set of shared applications frequently used and popular by other collaborating personnel within the target object's specific enterprise tenant or organizational structure. These multi-dimensional feature data are obtained in real-time from the tenant management database and user behavior logs by the system's personalized recommendation module.

[0035] In step S135, based on the historical usage data, the industry data, the application usage frequency data, and the popular application data, the applications in the initial application set are re-ranked, and the application with the best ranking is determined as the target workflow application. This step belongs to the third layer of personalized ranking. Specifically, the personalized recommendation module converts the acquired multi-dimensional behavioral features and industry features into weighted recommendation factors, which are then substituted into the ranking model to perform a secondary correction on the scores of each application in the initial application set. This process allows customized workflow applications that match the industry characteristics of the current target object, personal usage habits, and high frequency of use within its tenant to obtain higher ranking weights, thus being re-ranked in the initial application set. The system ultimately determines the application with the highest score and best ranking after the secondary ranking (i.e., the Top 1 best matching application) as the target workflow application. As a supplementary implementation, when outputting the final matching results, in addition to recommending the target workflow application as the best option, the system can also extract 2-3 other top-ranked applications as alternatives. These alternatives are then uniformly displayed to the target user in the form of application cards within the dialog interface. Each application card includes the application name, function description, matching score, and interactive elements such as a visual usage button, allowing the target user to intuitively view and access them. This design is suitable for enterprise-level multi-role collaborative work scenarios, aiming to achieve personalized and accurate recommendations.

[0036] By performing type and scenario filtering, semantic vector similarity matching, and personalized re-ranking of multi-dimensional habit data sequentially across a full application library encompassing system-wide, tenant-shared, and personal applications based on structured intent features, the application with the best ranking is identified as the target workflow application and presented in a card format. The first-layer type and scenario coarse filtering mechanism eliminates a large number of irrelevant applications at the outset with minimal computational overhead, significantly reducing the candidate base for subsequent complex calculations and substantially lowering the system's computational load and response latency. The second-layer semantic fine filtering based on vector similarity accurately captures the semantic tendencies in complex, colloquial descriptions of target objects, overcoming the technical shortcomings of traditional keyword retrieval, such as its inability to understand long texts and frequent search mismatches, ensuring the accuracy of the matching results in terms of business direction. The third-layer personalized re-ranking mechanism, integrating user industry, personal habits, and tenant popular data, effectively solves the monotonous problem of generic matching in traditional dialogue matching, enabling matching results to precisely match the unique creative preferences of different enterprises and individuals. This not only improves the final matching accuracy but also powerfully promotes the efficient reuse and promotion of high-quality custom workflow assets accumulated within the enterprise.

[0037] In one embodiment, the step of re-ranking the applications in the initial application set based on the historical usage habit data, the industry data, the application usage frequency data, and the popular application data, and determining the application with the best ranking as the target workflow application, specifically includes the following steps: Step S1351: Identify whether the target object meets the cold start conditions; Step S1352: If the cold start condition is met, then extract the preset scene matching weight based on the industry data of the target object; Step S1353: Based on the scenario matching weight, combined with the historical usage habit data, the application usage frequency data, and the popular application data, prioritize and rank the applications in the preliminary application set, and determine the application with the best ranking as the target workflow application.

[0038] Specifically, in order to achieve dynamic adaptation of application matching results in both time and object dimensions, the ranking mechanism needs to perform special scheduling for target objects that are newly added to the platform or lack behavioral trajectories, in order to avoid ranking distortion caused by data sparsity in traditional recommendations.

[0039] In step S1351, the system identifies whether the target object meets the cold start condition. The cold start condition refers to the target object having empty historical behavior tracking points, application call logs, or parameter configuration preference data stored within the system, or its accumulated number of valid interaction records being lower than the system's preset behavior statistics threshold. In practical application scenarios, this usually corresponds to newly registered platform users, or business operators newly assigned to a specific module without any historical call tendency. The system queries the target object's behavior feature matrix, and when it determines that its behavior density cannot meet the calculation requirements of the standard recommendation model, it confirms that the cold start condition is met.

[0040] In step S1352, if the cold start condition is met, the system extracts a preset scene matching weight from the system's configuration database based on the target object's industry data. The physical meaning of scene matching weight refers to a pre-defined quantitative feature vector or probability gain coefficient used to intervene in the recommendation priority of candidate applications for different vertical industry categories. The industry data can be the business domain tag actively selected by the target object when registering and opening an account (e.g., e-commerce, short video, comics, enterprise new media, etc.), or it can be the industry affiliation feature automatically inherited by the system based on the enterprise tenant attribute of the target object; no specific limitation is made. As an example, if the industry data obtained by the system is e-commerce, the extracted scene matching weight will assign a higher rearrangement gain coefficient to applications with functional tags such as product image generation, intelligent image editing, and marketing poster creation; if the industry data obtained is new media, the extracted scene matching weight will assign a higher rearrangement gain coefficient to applications with functional tags such as short video script creation and graphic poster generation. This process aims to quickly build a business baseline for new users without behavioral characteristics by introducing prior industry expert experience data.

[0041] In step S1353, based on the scenario matching weight, combined with the historical usage habit data, the application usage frequency data, and the popular application data, the applications in the initial application set are prioritized and ranked, and the application with the best ranking is determined as the target workflow application. In the specific calculation logic, since the target object has very little historical behavior at this time, the system uses the scenario matching weight as the dominant factor, and performs weighted matrix operations or joint collaborative filtering score calculations with a small amount of personal historical usage habit data, global application usage frequency data, and popular application data within the current user's enterprise tenant. Under this mechanism, although the new user does not have a record of frequently used applications, the system, under the spatial modulation of the scenario matching weight, can assign a very high merging weight to shared workflow applications or popular applications that are frequently used by other collaborators within the current enterprise tenant and highly match their industry attributes in the initial application set, so that they are re-ranked and prioritized for recommendation. The system finally determines the application with the highest score and the best ranking position after this cold start strategy re-ranking as the target workflow application for this dialogue, and presents it to the user in the form of an optimal card on the interactive interface.

[0042] By identifying whether a target object meets the cold start criteria, and extracting a pre-defined scenario matching weight based on its industry data when it does, the system then combines multi-dimensional features such as tenant popularity data and habitual data to perform priority recommendation ranking for the initial application set, thereby determining the target workflow application. First, this effectively solves the problem of traditional conversational matching algorithms failing to recommend applications or providing generic results due to the lack of historical data in the early stages of new user joining the platform. This allows the system to still provide accurate application recommendations that conform to industry practices even in the cold start scenario for new users. Second, by linking the new user's personal industry profile with the popular application data within their enterprise tenant through scenario matching weights, the system greatly improves the interactive experience during the cold start phase for new users and significantly enhances the level of automated reuse of enterprise-specific business workflows in multi-role collaborative scenarios.

[0043] In one embodiment, the step of extracting the target generation parameters from the structured intent features and mapping the target generation parameters to fill the variable input items corresponding to the target workflow application to obtain the target workflow application with configured parameters specifically includes the following steps: Step S141: Classify the target generation parameters into directly mappable parameters, text-based parameters, and multimodal material parameters; Step S142: Fill or associate the directly mappable parameters, the text-type parameters, and the multimodal material parameters into the variable input items corresponding to the target workflow application, respectively; Step S143: Generate a parameter configuration data table based on the completed variable input items, and determine the target workflow application after loading the parameter configuration data table as the target workflow application with the configured parameters.

[0044] Specifically, the target generated parameters often differ significantly in form and data structure. If a single string assignment method is used, it is very easy for the underlying API interface to encounter parsing errors due to data type mismatch during parameter validation. Therefore, before performing parameter mapping, the system needs to classify and isolate the parsed parameter set according to its format.

[0045] In step S141, the system categorizes the target generation parameters into directly mappable parameters, text-based parameters, and multimodal material parameters. Directly mappable parameters refer to quantifiable numerical or enumerable constraint variables with clearly defined data boundaries; examples include the number of generated images, image aspect ratios, screen resolution, or the specific duration of the generated video. Text-based parameters are long text structure variables that retain natural language features, such as user descriptions of the artistic style of the generated image, the core selling points of a specific product, or prompts provided to the underlying model. Multimodal material parameters refer to non-text file objects uploaded along with the text in the multimodal dialog box; these are typically parsed into storage address pointers or Uniform Resource Locators (URLs) pointing to specific image, audio, or video files during data flow. Through this classification operation, the system organizes heterogeneous parameter sources into data subsets adapted to different interface specifications.

[0046] In step S142, the directly mappable parameters, the text parameters, and the multimodal material parameters are respectively filled into or associated with the variable input items corresponding to the target workflow application. Specifically, for directly mappable parameters, the system strongly types-binds them to integer, floating-point, or enumeration fields of the application's underlying interface and directly fills them into the corresponding numerical input boxes on the front end; for text parameters, the system automatically writes them as string streams into the corresponding text boxes of the application. As a preferred configuration, these text boxes remain editable on the front end of the interface to support manual review and modification by the target object; for multimodal material parameters, the system does not directly transmit the large media file body, but instead associates the extracted storage address pointer with the preset media material input node of the workflow application, enabling cross-component calls to media data.

[0047] In step S143, a parameter configuration data table is generated based on the completed or associated variable input items, and the target workflow application loaded with the parameter configuration data table is identified as the target workflow application with configured parameters. The parameter configuration data table is an instantiated state object containing the assignment values ​​of all required and optional parameters of the current application. The system uniformly encapsulates all variables that have completed heterogeneous data binding to form this data table. Subsequently, the system loads the parameter configuration data table into the execution context of the target workflow application. At the presentation layer, the system automatically jumps to or renders and displays a dynamic data table (i.e., application card or form) containing the filled parameters in the current dialog interface. At this time, the system recognizes that the target workflow application has entered the data-ready pending execution state. The target object only needs to confirm that the parameters are correct and click submit to directly trigger the task generation, without manually filling in any initialization content.

[0048] By subdividing target generation parameters into directly mappable parameters, text-based parameters, and multimodal material parameters, and performing differentiated interface population or pointer association operations for different parameter types, the final result is encapsulated into a unified parameter configuration data table for loading into the target workflow application. Targeted classification and mapping of parameters based on data structure attributes effectively avoids runtime crashes caused by directly filling unstructured data into strongly typed interfaces, ensuring system stability in parameter transmission. Simultaneously, the differentiated population strategy enables smooth flow of multimodal data between the dialogue input domain and the workflow execution domain, automating the previously cumbersome manual parameter entry and material upload steps in the backend, significantly reducing user interaction costs and improving the overall efficiency of application configuration.

[0049] In one embodiment, after obtaining the target workflow application with the configured parameters, the method further includes: Step S161: Record the multi-turn dialogue context data corresponding to the natural language dialogue data; Step S162: Obtain the modification request data input by the target object, re-parse the modification request data in combination with the multi-turn dialogue context data, and extract the updated generation parameters and the updated intent features; Step S163: Determine whether the updated intent features exceed the capability range of the target workflow application with the configured parameters; Step S164: If the capability range is not exceeded, the target production parameters in the corresponding variable input item of the target workflow with the configured parameters are updated to the updated production parameters. Step S165: If the capability is exceeded, then the application library is re-matched based on the updated intent features to determine a new target workflow application; Step S166: Match the target generation parameters with the variable input items corresponding to the new target workflow application, extract the successfully matched parameters as reusable parameters, and map and fill the updated generation parameters and the reusable parameters into the variable input items corresponding to the new target workflow application.

[0050] Specifically, in real-world AI content creation scenarios, a single round of instruction input often fails to directly produce a perfectly matching result. The target audience typically needs to refine and fine-tune the generated content multiple times. To support this continuous iterative creation process, the system introduces a multi-round dialogue management and dynamic demand optimization mechanism.

[0051] In step S161, the system records the multi-turn dialogue context data corresponding to the natural language dialogue data. Structurally, the multi-turn dialogue context data refers to the historical interaction state stack persistently maintained by the system in memory cache or session database. This stack retains the creation type, business scenario, core entity elements, and historical parameter configurations already clearly defined by the target object in previous dialogues. The purpose of this step is to endow the system with memory capabilities, allowing the target object to avoid repeating already defined background information in subsequent interactions, only needing to explain the parts that need modification or adjustment.

[0052] In step S162, the system acquires the modification request data input by the target object, re-parses the modification request data in conjunction with the multi-turn dialogue context data, and extracts the updated generation parameters and updated intent features. Modification request data refers to the incremental natural language command input in the dialog box when the target object is dissatisfied with the current generation result or when its requirements change. As an example, modification request data could be "Change the image ratio to 1:1, change the style to Chinese trend, and change the quantity to 20 images." After receiving this data, the system does not treat it as an isolated command, but semantically concatenates and fuses it with the multi-turn dialogue context data, extracting the latest updated generation parameters (such as "1:1", "Chinese trend", "20 images") and the overall updated intent features through a pre-trained model.

[0053] In step S163, the system determines whether the updated intent feature exceeds the capability range of the target workflow application with configured parameters. Capability range refers to the modal boundaries and parameter adjustment thresholds supported by the currently used target workflow application at the underlying interface level. For example, if the current application is an image generation workflow, its capability range includes adjusting resolution, size, style, etc.; however, if the updated intent feature includes an operation request to "animate the image to generate a video," then this request exceeds the capability range of the current static image workflow in terms of physical execution logic.

[0054] Based on the above judgment, the system executes a traffic splitting strategy in steps S164 and S165: In step S164, if the capability range is not exceeded, the target generation parameters in the variable input items corresponding to the target workflow application with configured parameters are updated to the updated generation parameters. Under this path, the system determines that the current workflow is sufficient to support the new requirements, so it directly performs the local variable overwrite operation in the dynamic data table of the original application. The target object can directly see the parameters being updated in real time in the dialog interface, and after confirmation, the generation task can be re-executed without re-matching the application, resulting in the shortest operation flow path. In step S165, if the capability range is exceeded, the system re-matches the application library based on the updated intent features to determine a new target workflow application. Under this path, the system determines that the original tool has become invalid, thereby automatically triggering a new round of cross-modal or cross-scenario application library addressing and scheduling in the background to match an underlying application that can accept the new intent.

[0055] In step S166, the target generation parameters are matched with the variable input items corresponding to the new target workflow application. Successfully matched parameters are extracted as reusable parameters, and the updated generation parameters and the reusable parameters are mapped and filled into the variable input items corresponding to the new target workflow application. When the system switches between applications, historical data loss is likely to occur. To address this, the system extracts old parameters from the previous workflow (such as product main image materials, brand names, etc.), cross-compares them with the input slot specifications of the new application, and defines compatible parameters as reusable parameters for inheritance. Finally, the system merges the incremental update parameters just proposed by the target object and the inherited reusable parameters into the variable input items of the new application.

[0056] By recording multi-turn dialogue context data, when a modification request is received, update features are extracted based on the context. A two-way scheduling strategy of overwriting original application parameters or matching and reusing parameters with a new application is executed based on the current application's capability boundaries. This allows the system to continuously track and remember the user's incremental intent, avoiding repetitive input during multiple adjustments. Through capability range judgment, in-situ updates are implemented for minor parameter adjustments, significantly reducing the overhead and latency of reallocating computing resources. When a sudden change in requirements occurs, new application re-matching is implemented, inheriting reusable parameters, effectively preventing data gaps and operation resets caused by application switching. This solution significantly improves the naturalness and coherence of human-computer interaction, reduces the user's operational burden during continuous optimization of the generated results, and ensures that the generated results closely match the user's final business needs.

[0057] In one embodiment, step S120, namely, performing semantic parsing on the natural language dialogue data to extract structured intent features containing target generation parameters, specifically includes the following steps: Step S1201: Determine whether the natural language dialogue data meets the fuzzy requirement conditions; Step S1202: If the fuzzy requirement condition is met, generate multi-round question information and obtain supplementary dialogue data of the target object's response to the multi-round question information; fuse and parse the natural language dialogue data and the supplementary dialogue data to extract the structured intent features containing the target generation parameters; Step S1203: If the fuzzy requirement condition is not met, then the natural language dialogue data is directly subjected to semantic parsing to extract the structured intent features containing the target generation parameters.

[0058] Specifically, due to differences in the expression habits of different target objects, the text descriptions entered by some target objects when interacting with the platform for the first time are often too broad or abstract, making it impossible for the system to directly extract physical control variables that are compatible with the underlying API interface. To ensure the accuracy of application matching, the system first performs text completeness verification on the input unstructured data before performing deep parsing.

[0059] In step S1201, the system determines whether the natural language dialogue data meets the fuzzy requirement conditions. Fuzzy requirement conditions refer to situations where the input natural language dialogue data lacks key business dimension elements, or the text features of each intent dimension are too sparse, causing the system to be unable to generate effective matching retrieval conditions. In actual processing logic, as an example, when the target object only inputs vague text such as "help me create marketing content" in the dialog box, the system's internal pre-parsing operator, after performing preliminary keyword retrieval or shallow semantic scanning of the text, finds that its corresponding creation type features and business scenario features are blank, or the extracted intent confidence score is lower than the system's preset stability threshold. In this case, the system determines that the natural language dialogue data meets the fuzzy requirement conditions.

[0060] In step S1202, if the fuzzy requirement condition is met, the system automatically suspends the current application matching process and generates multi-round question information, which is then pushed to the front end. Multi-round question information refers to interactive text automatically generated by the system based on the currently missing intent dimension, used to guide the user to complete the information. For example, when both the business scenario and creation type are detected as missing, the system automatically generates targeted interactive questions such as, "What type of marketing content do you specifically need? Do you need new media images, marketing posters, or short video scripts?" After the target user sees the question on the front end interface and enters new response text, the system subsequently obtains supplementary dialogue data in response to the multi-round question information. Then, at the underlying level, the system performs text-level concatenation or vector-level feature fusion on the initially input natural language dialogue data and the subsequent multi-round obtained supplementary dialogue data, combining the multi-round interactive information into a fused dialogue stream with continuous contextual association. Subsequently, the system calls the pre-trained semantic parsing model to perform deep joint semantic parsing on the fused dialogue stream, thereby accurately extracting the structured intent features that include clear creation type features, business scenario features, and specific quantitative target generation parameters, which serve as the data basis for subsequent application addressing and filling.

[0061] In step S1203, if the fuzzy requirement condition is not met, it indicates that the target object has already provided relatively detailed business elements during the initial input (such as the aforementioned clear description of "help me generate 10 e-commerce women's clothing product images..."). In this case, the system determines that the initial input data has the completeness for direct addressing, so there is no need to trigger a multi-round follow-up mechanism. Instead, it directly calls the pre-trained semantic parsing model to perform one round of semantic parsing on the natural language dialogue data, directly extracting the structured intent features containing the target generation parameters, and transferring them to the subsequent application intelligent matching process, thereby reducing unnecessary interaction waiting time.

[0062] By introducing dynamic judgment of fuzzy demand conditions during semantic parsing, and generating multi-round question information to obtain supplementary dialogue data when fuzzy conditions are met, the initial dialogue and supplementary dialogue are then fused and parsed to extract structured intent features. This allows the system to gradually narrow down the broad and ambiguous colloquial expressions of the target object and transform them into standardized control variables with complete elements. This overcomes the shortcomings of existing technologies that are prone to matching failure, mismatch, or hard rejection of service when facing fuzzy input. It significantly improves the system's fault tolerance boundary and the proactive intelligence level of human-computer interaction when processing nondeterministic natural language input, and provides a reliable data contract closed loop guarantee for accurate addressing and adaptive parameter filling in the subsequent full application library.

[0063] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0064] In one embodiment, such as Figure 2 As shown, a conversational application matching system is provided, including: The user intent recognition module 210 is used to acquire the natural language dialogue data to be processed input by the target object in the AI ​​creation platform, and to perform semantic parsing on the natural language dialogue data to extract structured intent features containing target generation parameters. The application intelligent matching module 220 is used to match the structured intent features in a preset application library to determine the target workflow application corresponding to the natural language dialogue data. The application library includes system general applications, tenant shared applications and user private applications. The parameter automatic extraction and filling module 230 is used to extract the target generation parameters from the structured intent features and map and fill the target generation parameters into the variable input items corresponding to the target workflow application to obtain the target workflow application with configured parameters. The dialogue interaction module 240 is used to execute the corresponding generation task in response to the generation task execution instruction of the target workflow application with the configured parameters.

[0065] In one embodiment, such as Figure 3 As shown, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the conversational application matching method as described in any of the preceding claims.

[0066] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described conversational application matching method.

[0067] The details regarding the equipment and media can be found in the description of the aforementioned method embodiments, and will not be repeated here.

[0068] 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, and when executed, it can include the processes of the embodiments of the above methods. Furthermore, any references to memory, storage, user databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0069] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0070] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A matching method for conversational applications, characterized in that, include: Acquire the natural language dialogue data to be processed input by the target object in the AI ​​creation platform; Semantic parsing is performed on the natural language dialogue data to extract structured intent features containing target generation parameters; In a preset application library, the target workflow application corresponding to the natural language dialogue data is determined by matching based on the structured intent features. The application library includes general system applications, shared tenant applications, and private user applications. The target generation parameters are extracted from the structured intent features, and the target generation parameters are mapped and filled into the variable input items corresponding to the target workflow application to obtain the target workflow application with configured parameters. In response to the generation task execution instruction for the target workflow application with the configured parameters, the corresponding generation task is executed.

2. The method according to claim 1, characterized in that, The step of semantically parsing the natural language dialogue data and extracting structured intent features containing target generation parameters includes: The natural language dialogue data is input into a pre-trained semantic parsing model for semantic parsing to obtain the creation type features, business scenario features, core requirement descriptions, and target generation parameters contained in the natural language dialogue data. The creation type features, business scenario features, core requirement descriptions, and target generation parameters are stored in a structured manner to generate the structured intent features.

3. The method according to claim 2, characterized in that, The step of matching the structured intent features in a preset application library to determine the target workflow application corresponding to the natural language dialogue data includes: Based on the creation type features and business scenario features included in the structured intent features, the applications in the application library are filtered to select applications with matching types as a candidate application set. For each candidate application in the candidate application set, extract its corresponding functional description, scene tags, and keywords; Calculate the vector similarity between the core requirement description contained in the structured intent feature and the functional description, scene tag and keyword corresponding to each candidate application, and extract a preliminary application set from the candidate application set based on the vector similarity; Obtain historical usage data, industry data, application usage frequency data, and popular application data within the tenant to which the target object belongs; Based on the historical usage data, the industry data, the application usage frequency data, and the popular application data, the applications in the preliminary application set are reordered, and the application with the best ranking is determined as the target workflow application.

4. The method according to claim 3, characterized in that, The process involves re-ranking applications in the initial application set based on historical usage data, industry data, application usage frequency data, and popular application data, and determining the application with the best ranking as the target workflow application, including: Identify whether the target object meets the cold start conditions; If the cold start condition is met, then a preset scene matching weight is extracted based on the industry data of the target object. Based on the scenario matching weight, combined with the historical usage habit data, the application usage frequency data, and the popular application data, the applications in the preliminary application set are prioritized and ranked, and the application with the best ranking is determined as the target workflow application.

5. The method according to claim 1, characterized in that, The step of extracting the target generation parameters from the structured intent features and mapping and filling the target generation parameters into the variable input items corresponding to the target workflow application to obtain the target workflow application with configured parameters includes: The target generation parameters are classified into directly mappable parameters, text-based parameters, and multimodal material parameters. The directly mappable parameters, the text-type parameters, and the multimodal material parameters are respectively filled into or associated with the variable input items corresponding to the target workflow application; A parameter configuration data table is generated based on the variable input items that have been filled or associated, and the target workflow application after loading the parameter configuration data table is determined as the target workflow application with the configured parameters.

6. The method according to claim 1, characterized in that, After obtaining the target workflow application with the configured parameters, the method further includes: Record the multi-turn dialogue context data corresponding to the natural language dialogue data; Obtain the modification request data input by the target object, re-parse the modification request data in combination with the multi-turn dialogue context data, and extract the updated generation parameters and updated intent features; Determine whether the updated intent features exceed the capability range of the target workflow application with the configured parameters; If the target workflow with configured parameters is within the specified capacity range, the target generation parameters in the corresponding variable input item will be updated to the updated generation parameters. If the target exceeds the specified capability range, a new target workflow application is determined by re-matching the application library based on the updated intent features. The target generation parameters are matched with the variable input items corresponding to the new target workflow application. The successfully matched parameters are extracted as reusable parameters, and the updated generation parameters and the reusable parameters are mapped and filled into the variable input items corresponding to the new target workflow application.

7. The method according to claim 1, characterized in that, The step of semantically parsing the natural language dialogue data and extracting structured intent features containing target generation parameters includes: Determine whether the natural language dialogue data meets the fuzzy requirement conditions; If the fuzzy requirement conditions are met, multi-round question information is generated, and supplementary dialogue data of the target object's response to the multi-round question information is obtained; the natural language dialogue data and the supplementary dialogue data are fused and parsed to extract the structured intent features containing the target generation parameters; If the aforementioned fuzzy requirement conditions are not met, then the natural language dialogue data is directly subjected to semantic parsing to extract the structured intent features containing the target generation parameters.

8. A conversational application matching system, characterized in that, include: The user intent recognition module is used to acquire the natural language dialogue data to be processed input by the target object in the AI ​​creation platform, and to perform semantic parsing on the natural language dialogue data to extract structured intent features containing target generation parameters. An intelligent matching module is used to match the structured intent features in a preset application library to determine the target workflow application corresponding to the natural language dialogue data. The application library includes general system applications, tenant shared applications, and user private applications. The parameter automatic extraction and filling module is used to extract the target generation parameters from the structured intent features and map and fill the target generation parameters into the variable input items corresponding to the target workflow application to obtain the target workflow application with configured parameters. The dialogue interaction module is used to execute the corresponding generation task in response to the generation task execution instruction of the target workflow application with the configured parameters.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images