Intelligent sorting and selection method for business travel application form based on interactive content scenario

Abstract

This invention belongs to the field of data processing technology and provides an intelligent sorting and selection method for business travel application forms in interactive content scenarios. The method includes: receiving user input statements in a business travel dialogue session; performing multi-level intent classification and parsing on the input statements; extracting multi-level slot parameters; and marking the input statement as a trigger event when the parsing result indicates a change in at least one core matching parameter. Based on the multi-level slot parameters, a dynamic context is determined, and a context snapshot is generated. Based on the context snapshot, a set of candidate business travel application forms is retrieved from a database, and all candidate business travel application forms are dynamically sorted to generate a priority sorting list. Based on the priority sorting list, the preferred business travel application form is determined, a business travel resource query is performed, and the business travel resource query results and natural language response content are generated and returned to the user. This solution improves the accuracy and response efficiency of application form matching and optimizes the experience of interactive business travel dialogues.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent sorting and selection method for business travel application forms based on interactive content scenarios. Background Technology

[0002] With the widespread adoption of digital office solutions in enterprises, business travel management systems have become core tools for corporate travel planning and itinerary management. Intelligent chatbots based on large-scale models, offering a natural interactive experience, are widely integrated into these systems to provide employees with intelligent services such as travel itinerary planning, transportation, and hotel bookings. In practical applications, employees often have multiple pending business travel requests. Chatbots are needed to match requests from these requests to those that meet the current travel planning requirements, and then use these matched requests to complete subsequent business travel resource searches and bookings. Therefore, accurate and efficient matching of requests is crucial for improving the interactive experience and service efficiency of business travel management systems.

[0003] In related technologies, existing systems have exposed many problems in business travel dialogue matching scenarios with multiple application forms: First, the sorting model of existing systems is static or updates slowly. Once the initial application form matching is completed, it is unable to dynamically perceive continuous changes in the user's intent within the same conversation flow and to quickly reorder the application forms within a single conversation, resulting in the recommendation results not following the changes in user intent in real time. Second, when the user's intent changes slightly or complexly, the system's attention is easily distracted by a large amount of historically irrelevant information, making it unable to effectively distinguish between historical intent and the latest intent, and thus unable to accurately identify and activate the application form that best meets the current needs from multiple pending travel application forms. Third, because the system cannot automatically complete the dynamic matching and switching of application forms, employees need to manually search and select the target document among multiple application forms, and may even need to restart the dialogue flow to complete the application form switching, which seriously disrupts the continuity of dialogue interaction, significantly increases the operational costs of employees' itinerary planning, and reduces the overall efficiency of business travel management.

[0004] This demonstrates that existing business travel management systems suffer from technical problems such as dynamic response failures, low matching accuracy, and cumbersome user operations in scenarios involving multiple application forms. Summary of the Invention

[0005] This invention provides an intelligent sorting and selection method for business travel application forms based on interactive content scenarios, in order to solve the shortcomings of existing business travel management systems in scenarios involving multiple application forms, such as dynamic response failure, low matching accuracy, and cumbersome user operations.

[0006] This invention provides an intelligent sorting and selection method for business travel application forms in an interactive content scenario, comprising: The system receives input statements from users in business travel conversations, performs multi-level intent classification and parsing on the input statements, extracts multi-level slot parameters, and marks the input statements as trigger events when the parsing result indicates that at least one core matching parameter has changed. Based on the multi-level slot parameters, the dynamic context of the triggering event is determined, and based on the dynamic context, a context snapshot focusing on the user's latest intent is generated; Based on the aforementioned context snapshot, a set of candidate business trip application forms for the user is retrieved from a pre-established database, and all candidate business trip application forms in the set are dynamically sorted to generate a priority sorting list for the application forms. The system determines the preferred business trip application based on the application priority ranking list, performs a business travel resource query based on the preferred business trip application, and generates and returns the business travel resource query results and a natural language response containing application matching change instructions to the user.

[0007] According to the intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided by the present invention, the input statement is subjected to multi-level intent classification and parsing, including: The input statement is decomposed into a structured form using a large language model to obtain intent classification results that include first-level intent scenarios, second-level intent classifications, and third-level intent classifications. Extract the slot parameters corresponding to the first-level intent scene, the second-level intent category, and the third-level intent category, and perform slot parameter completion operation to obtain multi-level slot parameters; The intent classification result and the multi-level slot parameters are used as the parsing result.

[0008] According to the intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided by the present invention, the dynamic context of the triggering event is determined based on the multi-level slot parameters, and a context snapshot focusing on the user's latest intent is generated based on the dynamic context, including: The multi-level slot parameters are updated to the pre-built dynamic context storage unit to obtain a dynamic context that is updated in real time with user input; The dynamic context is optimized by using conflict-aware dynamic KL constraints and a dual-region temporal attention mechanism to generate a context snapshot that focuses on the user's latest intent.

[0009] The intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided by the present invention optimizes the dynamic context through conflict-aware dynamic KL constraints and a dual-region temporal attention mechanism, generating a context snapshot focusing on the user's latest intent, including: The dynamic context is processed into layers to obtain the layered context; The layered context is processed using a dual-region temporal attention mechanism to obtain focused context features; The focused context features are subjected to semantic vector encoding and intent conflict degree quantization to obtain a conflict degree quantization value. Based on the quantified value of the conflict level, dynamic KL constraint adjustment with conflict awareness is performed to obtain the optimized context feature set; The optimized context feature set is structured and integrated, and a snapshot is generated to obtain a context snapshot that focuses on the user's latest intent.

[0010] According to the intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided by the present invention, the dynamic context is processed in layers to obtain a layered context, including: The information contained in the dynamic context is divided into an invariant anchor region and a variable state region to obtain a hierarchical context; The fixed anchor area is used to store core information that does not need to change with the dialogue, while the variable state area is used to store dynamically changing information.

[0011] According to the intelligent sorting and selection method for business travel application forms in interactive content scenarios provided by the present invention, the layered context is processed by a dual-region temporal attention mechanism to obtain focused context features, including: A zero bias is applied to the invariant anchor region to assign a constant highest attention weight to all core information within the invariant anchor region, thus obtaining the attention configuration result of the invariant anchor region; A negative bias proportional to the time distance is applied to the variable state region to perform soft forgetting on the dynamic change information of the variable state region, thereby obtaining the attention configuration result of the variable state region. Based on the attention configuration results of the invariant anchor region and the attention configuration results of the variable state region, attention feature fusion is performed to obtain focused context features.

[0012] According to the intelligent sorting and selection method for business travel application forms in interactive content scenarios provided by the present invention, the focused contextual features are subjected to semantic vector encoding and intent conflict degree quantification to obtain a conflict degree quantification value, including: Based on the focused context features, the user's current round instruction is semantically encoded to generate a current instruction semantic vector, and the key information of the previous round matching result in the variable state area is semantically encoded to generate a historical state semantic vector. Calculate the maximum cosine similarity between the current instruction semantic vector and all historical state semantic vectors, and calculate the decay temperature based on the maximum cosine similarity. The decay temperature is used as a quantification value for the degree of conflict.

[0013] According to the intelligent sorting and selection method for business travel application forms in interactive content scenarios provided by the present invention, based on the quantified value of conflict degree, dynamic KL constraint adjustment with conflict awareness is performed to obtain an optimized context feature set, including: Based on the quantified value of the conflict level, the KL penalty coefficient is dynamically updated to obtain the adjusted KL penalty coefficient. The adjusted KL penalty coefficient is applied to the focused context features to perform context feature constraint optimization, resulting in an optimized context feature set.

[0014] According to the intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided by the present invention, all candidate business travel application forms in the candidate business travel application form set are dynamically sorted to generate an application form priority sorting list, including: Determine the basic characteristics that match the current user's needs for each candidate business trip application form in the candidate business trip application form set; Based on the context snapshot, the candidate business trip application set, and the basic features, a dynamic sorting algorithm is used to globally re-evaluate and calculate the matching degree of all candidate business trip applications, generating a priority sorting list of application applications.

[0015] According to the intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided by the present invention, the database is updated in real time with the basic data and restriction data of the business travel application forms; The limiting conditions data include the user's department's business trip booking permissions, the city of the business trip, and the matching rules for hotel bookings.

[0016] This invention provides an intelligent sorting and selection method for business travel application forms in interactive content scenarios. Through multi-level intent classification and parsing, it accurately captures changes in core user needs and triggers dynamic processing. Relying on a contextual snapshot focusing on the user's latest intent, it achieves accurate retrieval and global dynamic sorting of candidate business travel application forms. Then, based on the best-matched preferred application form, it completes the business travel resource query and synchronously provides feedback on matching changes. This effectively overcomes the technical shortcomings of traditional business travel management systems, such as rigid static matching and the inertia of large-scale, long-dialogue states. It achieves real-time synchronization between application form sorting results and changes in user intent, significantly improving the accuracy and response efficiency of application form matching. Simultaneously, it eliminates the need for users to manually search for or switch application forms, simplifying the business travel itinerary planning process, reducing user operation time and communication costs, significantly optimizing the experience of interactive business travel dialogue, and improving the overall intelligence and efficiency of enterprise business travel management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the system functional module collaboration principle upon which the intelligent sorting and selection method for business travel application forms based on interactive content scenarios in this embodiment of the invention relies; Figure 3 This is the core working logic diagram of the dynamic sorting engine in this embodiment of the invention; Figure 4 This is a schematic diagram of the actual business travel dialogue interaction interface in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] The following is combined with Figures 1 to 4 This invention describes the detailed scheme of the intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided by embodiments of the present invention.

[0021] like Figure 1 and Figure 2 As shown in the figure, the intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided by this invention mainly includes the following steps: Step 110: Receive the user's input statement in the business travel dialogue session, perform multi-level intent classification and parsing on the input statement, extract multi-level slot parameters, and mark the input statement as a trigger event when the parsing result contains at least one core matching parameter change.

[0022] In this embodiment, the dialogue interaction and listening module can receive the user's input statements in the business travel dialogue session. The input statements can be initial itinerary planning requirements or subsequent intention change instructions, and support multiple input forms such as text input and voice-to-text conversion.

[0023] In practical applications, the Large Language Model (LLM) can be used to perform structured decomposition and intent parsing of input statements, with the core implementation being intent classification and slot parameter extraction and completion.

[0024] It is understood that this embodiment can determine whether the core matching parameters such as business trip destination, business trip time, business trip city, and travel type have changed based on the parsing result. If the determination result is yes, the input statement is marked as a trigger event and a signal to start the dynamic sorting process is sent to the core workflow engine. If the determination result is no, it is processed according to the normal business travel dialogue logic, and the parsing result is only updated to the dynamic context storage unit without triggering dynamic sorting.

[0025] Step 120: Determine the dynamic context of the triggering event based on the multi-level slot parameters, and generate a context snapshot focusing on the user's latest intent based on the dynamic context.

[0026] It should be noted that this embodiment can use the synergistic effect of conflict-aware dynamic KL constraints and dual-region temporal attention mechanism to perform hierarchical processing, attention weight configuration, conflict degree quantification, constraint adjustment and structured integration of dynamic context, filter out interference from historical irrelevant information, focus on the user's latest intent, and obtain a context snapshot.

[0027] Step 130: Based on the context snapshot, retrieve the user's candidate business trip application set from the pre-established database, and dynamically sort all candidate business trip application forms in the candidate business trip application set to generate an application form priority sorting list.

[0028] In this embodiment, dynamic sorting of candidate application forms is the core step in achieving accurate matching of application forms. The core step is to complete the retrieval of candidate application forms and global dynamic sorting, and generate a priority sorting list of application forms, providing effective data basis for the final query output.

[0029] Step 140: Determine the preferred business trip application based on the application priority sorting list, perform a business travel resource query based on the preferred business trip application, generate and return the business travel resource query results and natural language response content containing application matching change instructions to the user.

[0030] In this embodiment, result feedback and interaction are the final stages of the dynamic sorting process. The core tasks are to determine the preferred application form, query business travel resources, and provide result feedback, thereby achieving real-time interaction with the user.

[0031] The intelligent sorting and selection method for business travel application forms based on interactive content scenarios provided in this embodiment is applicable to the real-time intelligent matching and dynamic sorting of multiple business travel application forms in interactive business travel dialogue scenarios. Relying on large-model intent recognition technology, multi-level intent classification design, conflict-aware dynamic KL constraints, and dual-region time attention mechanism, it constructs a dialogue-driven, real-time response closed-loop technical process, realizing the second-level synchronization of application form matching results with the user's latest intent. It solves the technical problems of static matching rigidity, large-model long dialogue state inertia, and dynamic response failure in existing business travel management systems, significantly improving the efficiency of business travel planning and interactive experience.

[0032] In business travel conversations, enabling application form matching to follow changes in user intent in real time requires a basic closed-loop process: real-time dialogue interaction, intent-driven re-evaluation and re-ranking based on the latest context, and immediate result refresh and feedback. The most crucial and indispensable step is to use the intent-identified trigger event as a trigger to initiate a completely new, global recalculation of the application form ranking, rather than performing localized adjustments or simple filtering of the existing results. Figure 3 This demonstrates the core working logic of the dynamic sorting engine.

[0033] In one embodiment, multi-level intent classification and parsing of the input statement is performed, specifically including: First, the input statement is decomposed into a structured form using a large language model to obtain intent classification results that include first-level intent scenarios, second-level intent classifications, and third-level intent classifications.

[0034] The first-level intent scenario determines the core task scenario and usually remains stable during the dialogue. The second-level intent classification refines the scenario type and is the hub for dynamic responses. The third-level intent classification aggregates specific business parameters and is the core basis for ranking calculations. This embodiment, through a multi-level intent classification design, structurally decomposes complex open-domain, multi-turn dialogues into a predefined business hierarchy of scenario, operation, and parameters. It transforms the fuzzy natural language semantic understanding problem into a problem of accurate classification and filling within a limited, known label space, significantly improving the accuracy of intent recognition.

[0035] Then, the slot parameters corresponding to the first-level intent scene, the second-level intent category, and the third-level intent category are extracted, and the slot parameter completion operation is performed to obtain multi-level slot parameters.

[0036] In practical applications, the parameter slot configuration table for the corresponding business scenario can be called to extract the slot parameters corresponding to each level of intent from the input statement. If there are missing, ambiguous or incomplete parameters, they can be completed based on parameter completion technology combined with business travel rules and dialogue context. The slot parameters are then standardized and validated to obtain standardized multi-level slot parameters.

[0037] Finally, the intent classification results and multi-level slot parameters are used as the parsing results.

[0038] Table 1 below illustrates, for example, the first-level intent scenarios, second-level intent classifications, and third-level intent classifications, as well as the multi-level slot parameters, corresponding to different input statements.

[0039] Table 1. Intent classification results and multi-level slot parameters for different input statements.

[0040] In one embodiment, the dynamic context of the triggering event is determined based on multi-level slot parameters, and a context snapshot focusing on the user's latest intent is generated based on the dynamic context, specifically including: First, the multi-level slot parameters are updated to the pre-built dynamic context storage unit to obtain a dynamic context that is updated in real time with user input.

[0041] In this embodiment, standardized multi-level slot parameters can be updated to a pre-built dynamic context storage unit. This dynamic context storage unit is used to maintain a set of intent parameters that are updated in real time with user input during a single dialogue session. After the update, a dynamic context matching the user's current input is obtained.

[0042] Then, the dynamic context is optimized by using conflict-aware dynamic KL constraints and a dual-region temporal attention mechanism to generate a context snapshot that focuses on the user's latest intent.

[0043] In one specific implementation, dynamic context is optimized using conflict-aware dynamic KL constraints and a dual-region temporal attention mechanism to generate a context snapshot focusing on the user's latest intent, including: The first step is to perform layered processing on the dynamic context to obtain the layered context.

[0044] In this embodiment, the dynamic context is processed into layers to obtain a layered context, specifically including: The information contained in the dynamic context is divided into an invariant anchor region and a variable state region to obtain a hierarchical context.

[0045] The unchanging anchor area stores core information that does not need to change with the conversation, including corporate travel policies, core security compliance rules, and basic system prompts. The variable state area stores dynamically changing information, including the full user conversation history of the current session, historical intent parsing results, historical application form matching status and slot parameters, etc. The variable state area information is timestamped according to the conversation round, completing structured preprocessing.

[0046] The second step is to process the layered context using a dual-region temporal attention mechanism to obtain focused context features.

[0047] This embodiment designs differentiated attention weight configuration rules for the two regions of the context after layering, realizes weight locking for the invariant anchor region and soft forgetting of historical information in the variable state region, and finally completes attention feature fusion to obtain focused context features.

[0048] Specifically, a dual-region temporal attention mechanism is applied to the hierarchical context to obtain focused context features, including: On the one hand, a zero bias is applied to the invariant anchor region to give all core information within the invariant anchor region a constant highest attention weight, thus obtaining the attention configuration result of the invariant anchor region.

[0049] Understandably, applying zero bias to the invariant anchor zone can assign a constant highest attention weight to all core information within the zone, thereby ensuring that core information such as corporate travel policies and compliance rules are not attenuated or blocked at any stage of the conversation.

[0050] On the other hand, a negative bias proportional to the time distance is applied to the variable state region to perform soft forgetting of the dynamic change information of the variable state region, thus obtaining the attention configuration result of the variable state region.

[0051] In practical applications, a negative bias proportional to the time distance can be applied to the variable state region. By combining the time distance of dialogue rounds / tokens, a learnable decay intensity coefficient, and a fixed temperature parameter, attention scores can be automatically reduced for distant and outdated historical dialogue information, while high attention weights can be assigned to the most recent rounds of user input, thereby achieving soft forgetting of invalid historical information.

[0052] This means that when a user says they want to go to Beijing, features related to Shanghai in the old context are masked, allowing the model to eliminate historical interference and accurately pinpoint and extract the latest intent.

[0053] In this embodiment, the attention weights can be directly modified using a bias matrix, as defined below: (1) in, B i, j The bias matrix, To query location i AND key position j The time distance between them This is a scaling factor for conflict sensitivity. A fixed temperature parameter is used to control the smoothness of the decay.

[0054] Finally, based on the attention configuration results of the invariant anchor region and the attention configuration results of the variable state region, attention feature fusion is performed to obtain focused context features.

[0055] In practical applications, the attention configuration results of the two regions mentioned above can be normalized using the Softmax function, and the information of the high-weight invariant anchor region, the latest user input information, and the effective dialogue information of the recent turns can be fused to obtain focused contextual features.

[0056] The third step is to perform semantic vector encoding and intent conflict degree quantification on the focused context features to obtain the conflict degree quantification value.

[0057] In this embodiment, semantic vector encoding and intent conflict level quantization are performed on focused context features to obtain a conflict level quantization value, specifically including: First, based on the focused context features, the user's current round instruction is semantically encoded to generate the current instruction semantic vector, and the key information of the previous round matching result in the variable state area is semantically encoded to generate the historical state semantic vector.

[0058] In practical applications, pre-trained models such as SBERT can be used to semantically encode the user's current round of instructions to generate a semantic vector for the current instruction. At the same time, key information such as the summary information of the previous round of matching results and historical key instructions in the variable state region can be semantically encoded to generate a historical state semantic vector.

[0059] Then, the maximum cosine similarity between the current instruction semantic vector and all historical state semantic vectors is calculated, and the decay temperature is calculated based on the maximum cosine similarity.

[0060] It is understandable that the maximum cosine similarity between the current instruction semantic vector and all historical state semantic vectors can inversely represent the degree of intent conflict. The lower the maximum cosine similarity, the stronger the degree of intent conflict. Combined with two adjustable hyperparameters, the conflict sensitivity scaling factor and the basic decay temperature, the decay temperature can be calculated in real time.

[0061] Finally, the decay temperature is used as a quantification of the degree of conflict.

[0062] The fourth step is to perform dynamic KL constraint adjustment based on the conflict degree quantification value to obtain the optimized context feature set.

[0063] In this embodiment, based on the quantified conflict level, dynamic KL constraint adjustment for conflict awareness is performed to obtain an optimized context feature set, specifically including: First, based on the quantified value of the conflict level, the KL penalty coefficient is dynamically updated to obtain the adjusted KL penalty coefficient.

[0064] Specifically, the core of conflict-aware dynamic KL constraint is to design the KL penalty coefficient β during training as a dynamic variable. β ( t Specifically, it is expressed as follows: (2) in, τ The stronger the conflict, the better the temperature reduction. τ The smaller, β ( t The faster the decay, the lighter the penalty for the model deviating from history, thus encouraging it to update the state boldly.

[0065] The core of this mechanism is the dynamic calculation of the decay temperature, which in turn controls the rate of constraint relaxation. The dynamic decay temperature can be expressed as follows: (3) in, τ ( u T () represents the dynamic decay temperature. For the user's instruction in round T, such as "No, go to Beijing"; For the current instruction The semantic vector, i.e., the semantic vector of the current instruction; The semantic vector of the summary or key user instructions of the previous round of matching results in historical round i is the historical state semantic vector. For example, if the previous round of matching results was a Shanghai application form, its summary vector may encode information such as Shanghai and early next week. The maximum cosine similarity between the current instruction semantic vector and all historical state semantic vectors is used in business travel scenarios to quantify the degree of conflict between the user's new intent and the historical matching state. The conflict sensitivity scaling factor is an adjustable hyperparameter. The base decay temperature.

[0066] Then, the adjusted KL penalty coefficient is applied to the focused context features to perform context feature constraint optimization, resulting in the optimized context feature set.

[0067] In practical applications, the difference between the predicted distribution based on historical context and the predicted distribution based on the latest instruction can be used by the KL divergence calculation model to dynamically redistribute the weights of the feature set, thereby further enhancing the decision weight of the latest intent features and weakening the weights of historical features that conflict with the latest intent. If obvious semantic conflicts are detected, the decision authority of conflicting historical features is directly blocked, resulting in an optimized context feature set.

[0068] The fifth step involves structurally integrating the optimized context feature set and generating a snapshot to obtain a context snapshot that focuses on the user's latest intent.

[0069] In this embodiment, the core information of the optimized context feature set can be extracted and restructured. The invariant anchor zone compliance rules, the complete slot parameters of the user's latest intent, and the conflict-free near-turn dialogue auxiliary information can be extracted and structured and integrated according to the business needs of matching business travel application forms to generate a context snapshot focusing on the user's latest intent. This context snapshot is a read-only standardized data body and is the only input basis for subsequent application form retrieval and sorting.

[0070] In this embodiment, based on a focused contextual snapshot, the application form real-time retrieval and evaluation module can retrieve all valid candidate business trip application forms under a user's name from a pre-established database, forming a candidate business trip application form set. The database updates the basic data and restriction data of the business trip application forms in real time. The basic data includes the application form number, business trip date, business trip city, and business trip participant, while the restriction data includes the business trip booking permissions of the user's department, the matching rules for business trip city and hotel booking, and travel expense over-limit restrictions.

[0071] In one embodiment, all candidate business trip application forms in the candidate business trip application form set are dynamically sorted to generate an application form priority sorting list, specifically including: First, determine the basic characteristics that match the current user's needs for each candidate business trip application in the candidate business trip application set.

[0072] In this embodiment, for each candidate application form in the candidate business trip application form set, the basic features that match the current user's needs can be calculated by combining the latest user intent parameters in the focused context snapshot. Specifically, the basic features include quantitative indicators such as destination matching degree, time matching degree, and business trip type matching degree.

[0073] Then, based on the context snapshot, the candidate business trip application set, and basic features, a dynamic sorting algorithm is used to globally re-evaluate and calculate the matching degree of all candidate business trip applications, generating a priority sorting list of application applications.

[0074] In this embodiment, the core workflow engine can call the dynamic sorting calculation module, which inputs the focused context snapshot, the candidate business trip application set, and the basic features of each application into the module. The dynamic sorting algorithm performs a global re-evaluation and matching degree calculation on all candidate business trip applications. Instead of performing local repairs or simple filtering on the original matching results, it completes a full re-evaluation of all applications based on the user's latest intent. Finally, it sorts the applications from high to low according to the matching degree to generate a priority sorting list.

[0075] In the results feedback and interaction phase, the core workflow engine can select the application with the highest matching degree from the application priority ranking list as the preferred business trip application and lock the unique identifier of the application. Then, it can send the unique identifier of the preferred business trip application to the downstream business travel resources and services module. Based on the business trip information of the preferred application and combined with the restrictive data in the database, this module can complete the real-time query of business travel resources such as air tickets, hotels, trains, and car rentals to obtain business travel resource query results. Finally, based on the matching results of the preferred application, it can generate natural language response content containing the application matching change explanation. Then, through the dialogue interaction and listening module, the business travel resource query results and natural language response content are synchronously fed back to the user, thus completing a complete dynamic sorting response closed loop.

[0076] Figure 4 This shows the actual state of the business travel dialogue interface, such as... Figure 4 As shown, in the first dialogue, the user first enters "Help me book a train to Xi'an for August 10th." The system immediately responds, "An application form has been quickly matched for you. Please confirm whether you want to plan your trip," and displays key information of the matched business trip application form, including the business trip dates of September 26th to 27th, 2025, the city of business trip (Shanghai), the traveler, and the application form number. In the second dialogue, the user enters, "I'm going to Shanghai for a business trip the day after tomorrow. Please help me plan my trip." The system simultaneously responds, "An application form has been intelligently matched for you. Please confirm," and displays the same Shanghai business trip application form, while providing interactive options to view more applications, search only, and confirm.

[0077] After the feedback is completed, the current business travel dialogue session can be kept active. The dialogue interaction and listening module continues to listen for the user's subsequent input. If a new input statement containing changes in core parameters is received, the above process of trigger event recognition, focused context snapshot generation, dynamic sorting of candidate application forms, result feedback and interaction is repeated to realize real-time dynamic updates of application form sorting.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A method for intelligent sorting and selecting business travel application forms based on interactive content scenarios, characterized in that, include: The system receives input statements from users in business travel conversations, performs multi-level intent classification and parsing on the input statements, extracts multi-level slot parameters, and marks the input statements as trigger events when the parsing result indicates that at least one core matching parameter has changed. Based on the multi-level slot parameters, the dynamic context of the triggering event is determined, and based on the dynamic context, a context snapshot focusing on the user's latest intent is generated; Based on the aforementioned context snapshot, a set of candidate business trip application forms for the user is retrieved from a pre-established database, and all candidate business trip application forms in the set are dynamically sorted to generate a priority sorting list for the application forms. The system determines the preferred business trip application based on the application priority ranking list, performs a business travel resource query based on the preferred business trip application, and generates and returns the business travel resource query results and a natural language response containing application matching change instructions to the user.

2. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 1, characterized in that, The input statement is subjected to multi-level intent classification and parsing, including: The input statement is decomposed into a structured form using a large language model to obtain intent classification results that include first-level intent scenarios, second-level intent classifications, and third-level intent classifications. Extract the slot parameters corresponding to the first-level intent scene, the second-level intent category, and the third-level intent category, and perform slot parameter completion operation to obtain multi-level slot parameters; The intent classification result and the multi-level slot parameters are used as the parsing result.

3. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 1, characterized in that, Based on the multi-level slot parameters, the dynamic context of the triggering event is determined, and based on the dynamic context, a context snapshot focusing on the user's latest intent is generated, including: The multi-level slot parameters are updated to the pre-built dynamic context storage unit to obtain a dynamic context that is updated in real time with user input; The dynamic context is optimized by using conflict-aware dynamic KL constraints and a dual-region temporal attention mechanism to generate a context snapshot that focuses on the user's latest intent.

4. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 3, characterized in that, The dynamic context is optimized using conflict-aware dynamic KL constraints and a dual-region temporal attention mechanism to generate a context snapshot focused on the user's latest intent, including: The dynamic context is processed into layers to obtain the layered context; The layered context is processed using a dual-region temporal attention mechanism to obtain focused context features; The focused context features are subjected to semantic vector encoding and intent conflict degree quantization to obtain a conflict degree quantization value. Based on the quantified value of the conflict level, dynamic KL constraint adjustment with conflict awareness is performed to obtain the optimized context feature set; The optimized context feature set is structured and integrated, and a snapshot is generated to obtain a context snapshot that focuses on the user's latest intent.

5. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 4, characterized in that, The dynamic context is processed into layers to obtain a layered context, including: The information contained in the dynamic context is divided into an invariant anchor region and a variable state region to obtain a hierarchical context; The fixed anchor area is used to store core information that does not need to change with the dialogue, while the variable state area is used to store dynamically changing information.

6. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 5, characterized in that, The hierarchical context is processed using a dual-region temporal attention mechanism to obtain focused contextual features, including: A zero bias is applied to the invariant anchor region to assign a constant highest attention weight to all core information within the invariant anchor region, thus obtaining the attention configuration result of the invariant anchor region; A negative bias proportional to the time distance is applied to the variable state region to perform soft forgetting on the dynamic change information of the variable state region, thereby obtaining the attention configuration result of the variable state region. Based on the attention configuration results of the invariant anchor region and the attention configuration results of the variable state region, attention feature fusion is performed to obtain focused context features.

7. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 5, characterized in that, The focused contextual features are subjected to semantic vector encoding and intent conflict quantization to obtain conflict quantization values, including: Based on the focused context features, the user's current round instruction is semantically encoded to generate a current instruction semantic vector, and the key information of the previous round matching result in the variable state area is semantically encoded to generate a historical state semantic vector. Calculate the maximum cosine similarity between the current instruction semantic vector and all historical state semantic vectors, and calculate the decay temperature based on the maximum cosine similarity. The decay temperature is used as a quantification value for the degree of conflict.

8. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 4, characterized in that, Based on the quantified conflict level, dynamic KL constraint adjustment with conflict awareness is performed to obtain an optimized context feature set, including: Based on the quantified value of the conflict level, the KL penalty coefficient is dynamically updated to obtain the adjusted KL penalty coefficient. The adjusted KL penalty coefficient is applied to the focused context features to perform context feature constraint optimization, resulting in an optimized context feature set.

9. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 1, characterized in that, Dynamically sort all candidate business trip application forms in the candidate business trip application form set to generate a priority ranking list of application forms, including: Determine the basic characteristics that match the current user's needs for each candidate business trip application form in the candidate business trip application form set; Based on the context snapshot, the candidate business trip application set, and the basic features, a dynamic sorting algorithm is used to globally re-evaluate and calculate the matching degree of all candidate business trip applications, generating a priority sorting list of application applications.

10. The intelligent sorting and selection method for business travel application forms based on interactive content scenarios according to claim 1, characterized in that, The database is updated in real time with the basic data and restriction data of the business trip application form; The limiting conditions data include the user's department's business trip booking permissions, the city of the business trip, and the matching rules for hotel bookings.

Images