Travel information processing method and device, storage medium and program product

Abstract

The embodiment of the invention provides a travel information processing method and device, a storage medium and a program product. In the scheme, based on semantic scene intelligent perception and preference weight dynamic allocation, the de-duplication intensity gear of each dimension can be adaptively determined, the macroscopic performance target is accurately conducted to the preference dimension hierarchy, and a hierarchical and parameterized control link is formed. And secondly, through the synergistic effect of a degradation mechanism and an upgrade mechanism, when candidate results are insufficient, intelligent circle expansion and amount compensation are performed from low to high according to preference weights, and when the results are excessive, dimension-by-dimension tightening is performed according to the preference weights, so that the number of the results is rapidly adjusted to the optimal interval while the core preference is not weakened. More importantly, according to the scheme, a traditional duplicate removal mode with a fixed threshold value is abandoned, so that a duplicate removal strategy can be continuously evolved according to a real-time scene and user preference, finally, on the premise of guaranteeing the accuracy of journey recommendation, the diversity and individuation degree of recommendation results are remarkably improved, and accurate and rich specialized journey selection is provided for the user.

Description

Technical Field

[0001] This application relates to the field of travel information recommendation technology based on artificial intelligence, and in particular to a method, device, storage medium and program product for processing travel information. Background Technology

[0002] In its early days, the travel industry relied primarily on travel agencies and paper maps for itinerary recommendations, requiring users to visit physical locations for consultations and bookings. With the development of internet technology, the travel industry has undergone a significant transformation from offline to online. Various travel platforms provide convenient travel services to users through websites or mobile applications. Specifically, users need to provide precise keywords to the travel platform, which then analyzes user needs based on these keywords and makes targeted recommendations accordingly.

[0003] However, users often have vague travel ideas and cannot provide precise keywords, such as "want to go to a place with exotic charm but not too commercialized." In such cases, travel platforms may not be able to make accurate recommendations. To address this, travel platforms can obtain users' historical preference information and multimodal demand information, and combine this with the user's current multimodal demand information to conduct a comprehensive demand analysis, and then recommend travel itineraries based on the results of this comprehensive demand analysis.

[0004] However, travel itinerary recommendations based on comprehensive needs analysis often result in a large number of homogeneous options and insufficient personalization. Users need to spend a lot of time filtering, and the final selected solution may not be optimal, potentially causing users to miss out on unique options (such as those with local cultural and customs features) or itineraries that better suit their actual needs. Although travel platforms usually prioritize and deduplicate recommended itineraries based on the matching degree between comprehensive needs analysis results and the itineraries to help users choose the best option, the current matching degree calculation and deduplication methods still have limitations and cannot guarantee both the accuracy and diversity of the recommended results. Summary of the Invention

[0005] This application provides a method, device, storage medium, and program product for processing trip information, which abandons the traditional fixed threshold deduplication mode and enables the deduplication strategy to continuously evolve according to real-time scenarios and user preferences. Ultimately, while ensuring the accuracy of trip recommendations, it significantly improves the diversity and personalization of recommendation results, providing users with accurate and rich personalized trip options. This application provides a method for processing travel information, including: acquiring a user's target demand preference information; inputting the target demand preference information into an information recommendation model, and performing the following operations in the information recommendation model: parsing the target demand preference information to obtain multiple preference dimensions and preference description information for each preference dimension; for any preference dimension, selecting travel information that matches the preference dimension and its preference description information from a diversified travel information database to obtain an initial travel list corresponding to the preference dimension; determining the target deduplication intensity level corresponding to each preference dimension based on preset multiple deduplication intensity levels, combined with the current semantic scene information and the preference weights of each preference dimension, wherein each preference dimension's target deduplication level has its own target similarity threshold, and the current semantic scene information refers to the set of context features used to dynamically determine each preference dimension; and deduplicating the initial travel list of the corresponding preference dimension based on the target similarity thresholds of each preference dimension. The process involves several steps: First, a deduplicated itinerary list is obtained for each preference dimension. Then, the intersection of these deduplicated itineraries is performed to obtain a candidate itinerary list. If the number of itineraries in the candidate itinerary list is less than a first threshold, a downgrade mechanism is initiated, discarding the intersection constraints of each preference dimension in ascending order of preference weight until the number of itineraries in the candidate itinerary list is greater than or equal to the first threshold, thus obtaining the target itinerary list. The intersection constraint means that the candidate itineraries must meet the adaptation conditions of the corresponding preference dimension. Second, an upgrade mechanism is initiated, adding additional intersection constraints to each preference dimension in descending order of preference weight until the number of itineraries in the candidate itinerary list is less than or equal to the second threshold, thus obtaining the target itinerary list. The additional intersection constraint means that the candidate itineraries must also meet the refined adaptation conditions under the corresponding preference dimension, and the second threshold must be greater than or equal to the first threshold.

[0006] This application also provides an electronic device, including: a memory and a processor; the memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program to implement the steps in the above method.

[0007] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, causes the processor to perform the steps in the above-described method.

[0008] This application also provides a computer program product, including: a computer program that, when executed by a processor of an electronic device, causes the processor to perform the steps in the above-described method.

[0009] The technical solution provided in this application, by constructing a bidirectional constraint optimization mechanism driven by multi-dimensional adaptive adjustment of deduplication intensity and preference weights, achieves a paradigm shift from "single-dimensional static deduplication" to "multi-dimensional dynamic deduplication," overcoming the fundamental drawback of traditional recommendation systems that struggle to balance accuracy and diversity. Specifically, firstly, based on semantic scene intelligent perception and dynamic allocation of preference weights, the deduplication intensity level of each dimension can be adaptively determined, accurately transmitting macro-performance targets to the preference dimension level, forming a parameterized control link of "global diversity target → dimension-level deduplication constraint → bidirectional adjustment strategy." Secondly, through the synergistic effect of degradation and upgrade mechanisms, when there are insufficient candidate results, the circle is intelligently expanded and supplemented according to preference weights; when there are too many results, the circle is tightened dimension by dimension according to preference weights, ensuring that the core preference is not weakened while quickly adjusting the number of results to the optimal range. Crucially, this solution abandons the traditional fixed-threshold deduplication mode, enabling the deduplication strategy to continuously evolve based on real-time scenarios and user preferences. Ultimately, while ensuring the accuracy of itinerary recommendations, it significantly improves the diversity and personalization of recommendation results, providing users with accurate and rich personalized itinerary options and reshaping the recommendation quality and user experience standards of intelligent travel assistants. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1a A flowchart illustrating a trip information processing method provided for an exemplary embodiment of this application; Figure 1b A schematic diagram of a multi-dimensional deduplication process provided for an exemplary embodiment of this application; Figure 2a A flowchart illustrating another trip information processing method provided as another exemplary embodiment of this application; Figure 2b A schematic diagram of the structure of a multimodal encoder provided for an exemplary embodiment of this application; Figure 2c A schematic diagram of the internal processing flow of a multimodal encoder provided for an exemplary embodiment of this application; Figure 2d A schematic diagram of a context semantic enhancement process provided for an exemplary embodiment of this application; Figure 3 A schematic diagram of the structure of an electronic device provided for an exemplary embodiment of this application. Detailed Implementation

[0011] Following the background technology, to address the technical problems existing in the background technology, this application embodiment achieves a paradigm shift from "single-dimensional static deduplication" to "multi-dimensional dynamic deduplication" by constructing a bidirectional constraint optimization mechanism driven by multi-dimensional deduplication intensity adaptive adjustment and preference weights. This overcomes the fundamental drawback of traditional recommendation systems that struggle to balance accuracy and diversity. Specifically, firstly, based on semantic scene intelligent perception and dynamic allocation of preference weights, the deduplication intensity level of each dimension can be adaptively determined, accurately transmitting macro performance targets to the preference dimension level, forming a parameterized control link of "global diversity target → dimension-level deduplication constraint → bidirectional adjustment strategy". Secondly, through the synergistic effect of degradation and upgrade mechanisms, when there are insufficient candidate results, the circle is expanded and supplemented according to preference weights; when there are too many results, the circle is tightened dimension by dimension according to preferences, ensuring that the core preference is not weakened while quickly adjusting the number of results to the optimal range. Crucially, this solution abandons the traditional fixed-threshold deduplication mode, enabling the deduplication strategy to continuously evolve based on real-time scenarios and user preferences. Ultimately, while ensuring the accuracy of itinerary recommendations, it significantly improves the diversity and personalization of recommendation results, providing users with accurate and rich personalized itinerary options and reshaping the recommendation quality and user experience standards of intelligent travel assistants.

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] It should be noted that the technical solutions of this application embodiment are applicable to a virtual network environment. The users described generally refer to "virtual users." Real users can register user accounts on the server through registration to obtain user identities in the network environment. The same user account can log in to the server through different types of user terminals, allowing the server to identify the same user. For example, the user in this application embodiment can be a user with a rental target type corresponding to the target rental object, i.e., a renter.

[0014] Interactions between the server and the user can be based on user accounts. The data received or sent by the server to the user is also based on the user account; in reality, the user's client, corresponding to the user account, receives or sends data to the server. Furthermore, users can also communicate with each other through their user accounts. Here, "user" can refer to an individual or an organization, such as a company; this application does not impose specific restrictions.

[0015] Furthermore, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. The collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0016] Furthermore, it should be noted that, in the cases involving user interaction operations or triggering operations in the embodiments of this application, the user interaction operations or triggering operations involved in the embodiments of this application include, but are not limited to, various interaction operations such as touch operations, gesture operations, voice operations, head movement operations, and eye movement operations; among them, touch operations include, but are not limited to, click operations, double-click operations, long-press operations, swipe operations, pinch operations, or mouse hover operations. Swipe operations include, but are not limited to, straight-line swipes and curved-line swipes.

[0017] It should also be noted that each model involved in the following embodiments can be a large AI model with corresponding functions. A large AI model refers to a model whose parameters meet the set parameter quantity requirements. This embodiment does not limit the parameter quantity requirements, and different definitions may exist in different scenarios and fields. For example, in the scenario of improving response rate, a large AI model can refer to a model with a parameter scale of tens of billions, hundreds of billions, or even trillions. However, the definition of the parameter quantity of this model is only an example and does not constitute a limitation on this application. This embodiment also does not limit the type of large AI model, such as a large language model (LLM), PaLM, or LLaMA.

[0018] Figure 1a This is a flowchart illustrating a trip information processing method provided for an exemplary embodiment of this application. Figure 1a As shown, the method includes: 101. Obtain user's target needs and preferences information; 102. Input the target demand preference information into the information recommendation model, and perform the following operations S11-S16 in the information recommendation model: S11. Analyze the target demand preference information to obtain multiple preference dimensions and preference description information on each preference dimension; S12. For any preference dimension, select the itinerary information that matches the preference dimension and its preference description information from the diversified itinerary information database to obtain the initial itinerary list corresponding to the preference dimension. S13. Based on the preset multiple deduplication strengths, combined with the current semantic scene information and the preference weights of each preference dimension, determine the target deduplication strength level corresponding to each preference dimension. The target deduplication level corresponding to each preference dimension has its own target similarity threshold. The current semantic scene information refers to the set of context features used to dynamically determine each preference dimension. S14. Based on the target similarity threshold corresponding to each preference dimension, perform deduplication on the initial itinerary list of the corresponding preference dimension to obtain the itinerary list after deduplication of each preference dimension. S15. Perform an intersection operation on the itinerary lists after deduplication of each preference dimension to obtain a candidate itinerary list; S16. If the number of trips in the candidate trip list is less than the first threshold, a downgrade mechanism is initiated. The intersection constraints of the corresponding preference dimensions are discarded in order of preference weight from low to high until the number of trips in the candidate trip list is greater than or equal to the first threshold, so as to obtain the target trip list. The intersection constraints refer to the adaptation conditions of the corresponding preference dimensions that the candidate trips must meet.

[0019] By constructing a bidirectional constraint optimization mechanism driven by multi-dimensional adaptive adjustment of deduplication intensity and preference weights, a paradigm shift from "single-dimensional static deduplication" to "multi-dimensional dynamic deduplication" has been achieved, overcoming the fundamental drawback of traditional recommendation systems that struggle to balance accuracy and diversity. Specifically, firstly, based on semantic scene intelligent perception and dynamic allocation of preference weights, the deduplication intensity level for each dimension can be adaptively determined, accurately transmitting macro-performance targets to the preference dimension level, forming a parameterized control link of "global diversity target → dimension-level deduplication constraint → bidirectional adjustment strategy". Secondly, through the synergistic effect of degradation and upgrade mechanisms, when there are insufficient candidate results, the number of results is intelligently expanded according to preference weights; when there are too many results, the number is tightened dimension by dimension according to preference weights, ensuring that the core preference is not weakened while quickly adjusting the number of results to the optimal range. Crucially, this solution abandons the traditional fixed-threshold deduplication mode, enabling the deduplication strategy to continuously evolve based on real-time scenarios and user preferences. Ultimately, while ensuring the accuracy of itinerary recommendations, it significantly improves the diversity and personalization of recommendation results, providing users with accurate and rich personalized itinerary options and reshaping the recommendation quality and user experience standards of intelligent travel assistants.

[0020] In this embodiment of the application, referring to step 101 above, the user's target demand preference information can be obtained first.

[0021] In some optional embodiments, obtaining the user's target demand preference information includes: during multi-round dialogues with the user regarding this trip planning, obtaining the user's initial demand information involving multiple modalities in this round of dialogue; calling each single-modal encoding module of the multimodal encoder to extract key information from the initial demand information of the corresponding single modality, obtaining a set of initial key information for each single modality; inputting the set of initial key information for each single modality into the first main model to perform demand information completeness analysis and initial trip demand category analysis, obtaining demand information completeness and initial trip demand category; when the demand information is complete... When the degree is less than the first completeness threshold, historical demand summaries and / or historical conversation summaries that match the initial trip demand category are retrieved from the long short-term memory storage area. Supplementary demand information for the initial demand information is generated based on the historical demand summaries and / or historical conversation summaries to obtain supplemented multimodal target demand information. The multimodal target demand information has a target trip demand category. The multimodal fusion encoding module of the multimodal encoder is invoked, and with the target trip demand category as a reference, demand integration and preference analysis are performed on the target demand information of each single modality to obtain the user's target demand preference information. This solution not only achieves a deep understanding and dynamic completion of the user's true intent but also accurately outputs the key elements relied upon by the subsequent recommendation process: including structured preference dimensions, explicit preference weights, and semantic scene features such as trip topic categories. This provides a high-quality input foundation for core recommendation mechanisms such as multidimensional trip retrieval, dynamic deduplication intensity adjustment, intersection fusion, and elastic quantity control, significantly improving the relevance of trip recommendations to the user's true needs and laying the groundwork for providing users with more accurate, intelligent, and coherent trip planning services. For more specific implementation methods of each step in this embodiment, please refer to the above. Figure 1a The relevant descriptions of the embodiments will not be repeated here.

[0022] Optionally, the short-term and long-term memory storage areas include a short-term memory storage area and a long-term memory storage area. The short-term memory storage area stores conversation summaries generated at a preset collection frequency, with the topic category remaining unchanged and each preset number of rounds being reached, as well as demand summaries and conversation summaries generated when the topic category changes but the preset number of rounds has not been reached. The long-term memory storage area stores demand summaries and conversation summaries that sink from the short-term memory storage area after reaching a preset storage duration.

[0023] Accordingly, historical demand summaries that match the initial trip demand category are retrieved from the long and short memory storage area, including: Based on the initial travel demand category, historical demand summaries are retrieved from the short-term memory storage area as the first supplementary demand information for the multimodal initial demand information. If historical demand summaries are not retrieved from the short-term memory storage area, or if historical demand summaries are retrieved but the demand completeness of the multimodal initial demand information supplemented based on the retrieved historical demand summaries is less than the second completeness threshold, then historical demand summaries are retrieved from the long-term memory storage area as the first supplementary demand information based on the initial travel demand category. This scheme, through hierarchical long and short-term memory storage and on-demand progressive retrieval mechanism, accurately activates historical preference information matching the current travel demand category. While avoiding redundant retrieval, it effectively completes the user's current incomplete expression, significantly improving the completeness and personalization level of demand understanding, laying the foundation for subsequent high-quality travel recommendations. For more specific implementation methods of each step in this embodiment, please refer to the above. Figure 1a The relevant descriptions of the embodiments will not be repeated here.

[0024] Accordingly, the historical session summary that matches the initial travel requirement category is recalled as context information for this round of the session, and the second major model is driven by the context information to obtain the second supplementary requirement information until the requirement completeness reaches the first threshold, so as to obtain the multimodal target requirement information. One optional implementation includes: recalling the historical session summary from the short-term memory storage area based on the initial travel requirement category; if the historical session summary is not recalled from the short-term memory storage area, recalling the historical session summary from the long-term memory storage area based on the initial travel requirement category; if the historical session summary is recalled from the short-term memory storage area and / or the long-term memory storage area, recalling the historical session summary that matches the initial travel requirement category from the long and short-term memory storage areas. The historical conversation summary includes: retrieving the historical conversation summary from the short-term memory storage area based on the initial travel demand category; if the historical conversation summary is not retrieved from the short-term memory storage area, retrieving the historical conversation summary from the long-term memory storage area based on the initial travel demand category; if the historical conversation summary is retrieved from the short-term memory storage area and / or the long-term memory storage area, the historical conversation summary is used as context information for the current round of dialogue, and a second major model is used to generate guiding dialogue based on the context information, and a question-and-answer service is provided to the user based on the guiding dialogue to obtain the second supplementary demand information provided by the user based on the question-and-answer service. This scheme, by accurately retrieving historical conversation summaries from both short-term and long-term memory based on travel demand categories, and using them as context to drive the second major model to generate semantically coherent supplementary demand information, effectively utilizes the implicit intentions and preferences in the user's past dialogues, dynamically enhancing the depth and completeness of the understanding of the current demand in scenarios with incomplete expression, thereby providing key support for generating high-fidelity, context-consistent target demand information, and improving the intelligence and personalization level of travel planning. For more specific implementation methods of each step in this embodiment, please refer to the above. Figure 1a The relevant descriptions of the embodiments will not be repeated here.

[0025] Optionally, the multimodal fusion encoding module of the multimodal encoder is invoked to perform demand integration and preference analysis on the target demand information of each single modality, with the target travel demand category as a reference, to obtain the user's target demand preference information. This includes: invoking the multimodal fusion encoding module of the multimodal encoder to perform cross-attention calculation on the target demand information of each single modality, with the target travel demand category as a reference, and assigning a global importance weight to each key target information in the target demand information of each single modality relative to the target travel demand category during the cross-attention calculation process; combining the various target key information in the target demand of each single modality... The global importance weight of target key information and the first context constraint condition representing the local context of the current dialogue are used to dynamically adjust the contribution of each target key information in the target demand information of each single modality, generating a context-aware weighted key information representation corresponding to each target key information; using the target travel demand category as the fusion condition, the representations of each target key information are fused across modally to obtain the fused key information representation; based on the fused key information representation and the second context constraint condition representing the user's global profile or external environment, demand preference analysis is performed from multiple dimensions to obtain target demand preference information. This scheme uses the target travel demand category as the anchor point, utilizes a cross-attention mechanism to perform cross-modal alignment and global importance weighting of key information in each single modality, and dynamically adjusts the contribution of each modality in combination with context constraints to achieve semantically consistent and key-focused multimodal fusion; on this basis, the system structurally extracts multi-dimensional preferences and their weights from the fused representation to generate accurate and interpretable target demand preference information, providing a highly consistent and high-fidelity decision basis for subsequent intelligent deduplication and personalized recommendation. For more specific implementation methods of each step in this embodiment, please refer to the above. Figure 1a The relevant descriptions of the embodiments will not be repeated here.

[0026] After obtaining the user's target demand and preference information, referring to step S11 above, the target demand and preference information can be parsed to obtain multiple preference dimensions and preference description information for each preference dimension. The specific implementation method can be found above. Figure 1a The relevant descriptions of the embodiments will not be repeated here.

[0027] Furthermore, referring to step S12 above, for any preference dimension, select travel information that matches the preference dimension and its description information from a diversified travel information database to obtain an initial travel list corresponding to that preference dimension. The specific implementation method can be found above. Figure 1a The relevant descriptions of the embodiments will not be repeated here.

[0028] Further, referring to step S12, for any preference dimension, select the itinerary information that matches the preference dimension and its description information from the diversified itinerary information database to obtain the initial itinerary list corresponding to the preference dimension. This diversified itinerary information database can be a structured database maintained in the aforementioned data processing layer, covering various travel elements such as attractions, restaurants, and hotels, and supporting multi-condition retrieval by geographical location, service type, price range, etc.

[0029] In some alternative embodiments, the deduplication operation described in steps S13 and S14 may employ... Figure 1b The multi-dimensional intelligent deduplication process is shown below. For example... Figure 1b As shown, this process takes the initial itinerary list output in step S12 (e.g., containing 200 potential recommended itineraries) as input. First, a multi-dimensional feature extraction module obtains the location, type, price, style, and other features of each itinerary. Then, a deep learning model is used to calculate the semantic similarity between itinerary information (or itineraries), and intelligent clustering analysis is performed to discover groups of similar itinerary information (itineraries). For each cluster group, it is determined whether its similarity exceeds a preset threshold: if it does, the most representative itinerary information (itinerary) within the group is selected and added to the recommendation list, while the rest are marked as duplicates and removed; if it does not exceed the threshold, the itinerary information (itinerary) is directly retained as a unique item. Subsequently, a diversity score is calculated on the retained itinerary information (itinerary) set to evaluate its balance in dimensions such as geographical distribution, service type, and price range. Combined with user preference weights and personalized adjustments, a candidate itinerary list consisting of a preset number (e.g., 1015) of carefully selected recommendations is finally generated (corresponding to step S15). It should be noted that... Figure 1b The process is only one possible implementation of step S14, but is not limited to it.

[0030] After obtaining the initial itinerary list corresponding to multiple preference dimensions, refer to step S13 above. Based on the preset multiple deduplication strengths, combined with the current semantic scene information and the preference weights of each preference dimension, the target deduplication strength level of each preference dimension can be determined. Each target deduplication level corresponding to each preference dimension has its own target similarity threshold.

[0031] The current semantic scenario information refers to the set of contextual features used to dynamically determine the deduplication strength of each preference dimension. This information includes, but is not limited to, at least one of the following: trip topic category, time attribute, audience tag, or adjacent trip context. Specifically, the trip topic category can reflect the user's intent type, such as "exploratory" (e.g., "What are some fun things to do in Beijing?"), "decision-making" (e.g., "Help me choose the best family-friendly restaurant with kids"), or "supplementary" (e.g., "Recommend a few similar cafes") identified through natural language understanding. Different intents correspond to different diverse needs, thus affecting the deduplication leniency. The time attribute includes holidays, travel dates, and query times within a day, used to determine the user's tolerance for the breadth and accuracy of recommendations (e.g., deduplication can be more lenient during holidays, but more stringent when travel is imminent). The audience tag reflects the user's identity characteristics, such as local residents. The system considers various factors, including tourists from other regions and families with children, and adjusts the redundancy filtering intensity based on spatial context (such as dense or sparse POI areas). Specifically, it increases the deduplication threshold in popular commercial areas to eliminate homogeneous items, while lowering the threshold in remote areas to avoid insufficient results. Furthermore, it can dynamically adjust the deduplication strategy based on real-time user interaction signals (such as frequent clicks on "refresh," quick selection, or prolonged inactivity) to assess user satisfaction with the diversity of current results. Adjacent trip context refers to the user's planned itinerary before and after the selected date in their trip calendar, used to assess whether the current recommendation aligns with the overall itinerary, thus affecting the deduplication granularity. These multi-dimensional semantic contextual information collectively form the decision-making basis of the intelligent deduplication mechanism, enabling the system to adaptively optimize the diversity and usability of recommendation results while ensuring the relevance of core preferences. Additionally, each of the preset deduplication intensity levels has a corresponding baseline similarity threshold for each preference dimension. This baseline similarity threshold serves as the default deduplication criterion for each preference dimension under the preset deduplication intensity level, acting as the basis for subsequent dynamic fine-tuning.

[0032] Based on this, in some optional embodiments, the target deduplication intensity level for each preference dimension is determined based on preset multiple deduplication intensity levels, combined with the current semantic scene information and the preference weights of each preference dimension. This includes: obtaining the current semantic scene information; selecting one level from the multiple deduplication intensity levels as the initial deduplication intensity level based on the current semantic scene information; and fine-tuning the baseline similarity thresholds corresponding to each preference dimension in the initial deduplication intensity level based on the preference weights of each preference dimension to obtain the target deduplication intensity level corresponding to each preference dimension, wherein the target deduplication level for each preference dimension has its own target similarity threshold.

[0033] Optionally, the current semantic scene information is obtained, including: extracting information such as trip topic category, time attribute, and audience tags from target demand preference information; in response to the user's selection of trip date, obtaining historical trip information or planned trip information of the selected trip date and adjacent dates from the trip calendar, and generating adjacent trip context based on historical trip information or planned trip information; obtaining user interaction behavior signals in the current session, and determining the user's feedback tendency on the diversity of recommendation results based on the interaction behavior signals; and combining the trip topic category, time attribute, audience tags, adjacent trip context, and user feedback tendency on the diversity of recommendation results into the current semantic scene information.

[0034] Optionally, based on the current semantic scene information, one of the multiple deduplication intensities is selected as the initial deduplication intensity level, including: encoding the current semantic scene information into a scene feature vector; inputting the scene feature vector into a pre-trained intensity level selection model, which is trained based on deduplication feedback data of historical users in similar semantic scenes; and outputting one of the multiple deduplication intensities from the intensity level selection model as the initial deduplication intensity level.

[0035] The technical solution provided in the above-mentioned step S13 related embodiments, by integrating the current semantic scene information and the personalized weights of each preference dimension, dynamically determines the exclusive target deduplication intensity level for each dimension from the preset multiple deduplication intensity levels, and assigns it an independent target similarity threshold, so that the deduplication strategy can not only respond to contextual changes (such as trip stage, user intent), but also accurately reflect the degree of importance that users attach to different preferences, thereby effectively suppressing redundancy while retaining key differences, and improving the relevance and diversity balance of recommendation results.

[0036] Optionally, based on the preference weights of each preference dimension, the baseline similarity thresholds corresponding to each preference dimension in the initial deduplication intensity level are fine-tuned to obtain the target deduplication intensity level for each preference dimension. This includes: obtaining a preset baseline weight that adapts to the current semantic scene information, where the preset baseline weights are obtained based on historical data of historical users in the same or similar semantic scenes; calculating the deviation between the preference weights of each preference dimension and the corresponding preset baseline weights; and adjusting the baseline similarity thresholds of each preference dimension in the initial deduplication intensity level based on the sign and magnitude of the deviations corresponding to each preference dimension as follows: If the deviation is positive, then increase the baseline similarity threshold for the corresponding preference dimension; If the deviation is negative, the baseline similarity threshold for the corresponding preference dimension is reduced. The adjusted baseline similarity thresholds corresponding to each preference dimension are used as target similarity thresholds to obtain the target deduplication intensity levels corresponding to each preference dimension.

[0037] The technical solution provided in this embodiment dynamically adjusts the similarity threshold of the corresponding dimensions by analyzing the deviation between the weights of the user's current preference dimensions and the benchmark weights in historical scenarios: tightening deduplication for preferences higher than the benchmark to enhance differentiation, and relaxing deduplication for preferences lower than the benchmark to retain diversity, thereby achieving personalized and fine-grained adaptation of the deduplication strategy, ensuring that core needs are accurately met while taking into account the richness and flexibility of the recommendation results.

[0038] After obtaining the target similarity thresholds for each preference dimension, refer to step S14. Based on the target similarity thresholds for each preference dimension, deduplication can be performed on the initial itinerary list for that preference dimension to obtain a deduplicated itinerary list for each preference dimension. By using the target similarity thresholds specific to each preference dimension, the initial itinerary list undergoes refined deduplication at the dimension level. This deduplication operation not only meets the diverse needs of the current semantic scenario but also reflects the user's emphasis on different preferences. It effectively eliminates homogeneous redundancy while retaining key differences, providing a high-quality, low-duplication candidate set for subsequent multi-dimensional fusion.

[0039] Furthermore, after obtaining the deduplicated itinerary lists for each preference dimension, referring to step S15, the intersection of these lists can be performed to obtain a candidate itinerary list. By performing the intersection operation on the deduplicated itinerary lists for each preference dimension, candidate itineraries that simultaneously satisfy multiple preferences are accurately selected. This effectively focuses on the core needs of users, ensuring a high degree of relevance in the recommendation results while avoiding bias caused by a single preference dimension, thus laying the foundation for generating high-quality, highly relevant final itinerary plans.

[0040] Furthermore, referring to step S16, if the number of trips in the candidate itinerary list is less than the first threshold, a degradation mechanism is initiated. The intersection constraints of each preference dimension are discarded sequentially from low to high according to their preference weights, until the number of trips in the candidate itinerary list is greater than or equal to the first threshold, thus obtaining the target itinerary list. By dynamically discarding constraints from low to high preference weights when candidate itinerary information is insufficient, the screening criteria are intelligently relaxed. While ensuring that high-priority preferences are strictly satisfied, more feasible options are gradually introduced, effectively avoiding gaps or insufficient results due to excessive constraints. This maintains the relevance of the recommendations while improving the usability and coverage of the results.

[0041] In some optional embodiments, when initiating the degradation mechanism, the intersection constraints of corresponding preference dimensions can be discarded sequentially from low to high according to the preference weights of each preference dimension, until the number of trips in the candidate trip list is greater than or equal to a first threshold, so as to obtain the target trip list. Optionally, for the current preference dimension to be processed, its corresponding deduplicated trip list can be excluded from the set of trip lists involved in the intersection operation, so that the candidate trips do not need to simultaneously meet the adaptation conditions of that preference dimension; if the number of trips in the updated candidate trip list is still less than the first threshold, then the next preference dimension arranged in ascending order of preference weight is processed until the number of trips in the candidate trip list is greater than or equal to the first threshold.

[0042] Specifically, after completing the intersection operation of the multi-dimensional itinerary lists, if the number of itineraries in the resulting candidate itinerary list is less than a preset first threshold (e.g., 5), the current results are considered too sparse to support effective user decision-making. At this point, a preference constraint degradation mechanism is automatically activated. By strategically relaxing some screening conditions, the number of candidate itineraries is rapidly expanded while preserving the user's core preferences as much as possible, ensuring the recommendation results have basic usability. The specific implementation of this mechanism is as follows: First, all preference dimensions corresponding to the current itinerary needs and their respective preference weights are obtained. Preference dimensions refer to the itinerary attribute categories that the user focuses on (such as "price sensitivity," "family-friendly," "cultural experience," etc.), and preference weights reflect the user's emphasis on that dimension; higher values ​​indicate greater importance. These preference dimensions are then arranged in ascending order of their weights, forming a sequence to be processed. For example, if a user's current preferences include "price sensitivity" (weight 0.2), "transportation convenience" (weight 0.3), "cultural experience" (weight 0.5), and "family-friendly" (weight 0.8), the processing order is as follows: "price sensitivity" → "transportation convenience" → "cultural experience" → "family-friendly".

[0043] Subsequently, the preference dimension with the lowest weight is selected from the sequence as the current item to be processed, and its corresponding deduplicated itinerary list is removed from the list set participating in the intersection operation. The deduplicated itinerary list refers to the set of candidate itineraries retrieved and duplicates removed for a specific preference dimension (e.g., a list of budget hotel packages that meet "price sensitivity"). Removing this list means that when recalculating the intersection later, it is no longer mandatory for the final itinerary to simultaneously meet the adaptation condition of that dimension. For example, the original intersection operation involved itinerary lists with four dimensions: L1 (price) ∩ L2 (transportation) ∩ L3 (culture) ∩ L4 (family); after discarding "price sensitivity," the new intersection becomes L2 ∩ L3 ∩ L4, allowing itineraries that were originally slightly more expensive but highly matched in other conditions to be reintroduced into the candidate set. After the removal operation is completed, the intersection operation is immediately re-executed based on the updated list set to generate a new candidate itinerary list, and the number of itineraries in the list is counted. If the number is still below the first threshold, the next preference dimension with the lowest weight is taken from the sorted sequence, and the above elimination and reintersection process is repeated. This loop continues until the number of candidate routes reaches or exceeds the first threshold, or all dimensions have been discarded (the latter is an extreme fallback case, which rarely occurs in practice).

[0044] This mechanism prioritizes relaxing constraints less important to users while strictly preserving the filtering effectiveness of high-weight dimensions. For example, in family travel scenarios, "family-friendly" remains a high-weight dimension in the intersection, avoiding recommendations for trips unsuitable for children (such as bars and skydiving). This expands the scope of results while ensuring that recommended content does not deviate from the user's core intent. Furthermore, since removing only one dimension at a time significantly increases the size of the intersection results, this strategy is responsive, logically clear, and requires no complex parameter tuning. In practical applications, the first threshold can be flexibly configured based on terminal type or business needs (e.g., 3-5 items for mobile devices to fit screen space, 5-8 items for PCs to provide richer options), while preference weights can be dynamically generated based on user historical behavior data (such as clicks, favorites, order completion rates) or real-time interactive feedback (such as swipe ratings, dwell time), ensuring that the degradation strategy always aligns with the user's true preferences. This solution effectively improves robustness and user experience continuity in low-recall scenarios.

[0045] In some alternative embodiments, when initiating the upgrade mechanism, additional intersection constraints for each preference dimension can be added sequentially from high to low according to the preference weights of each preference dimension, until the number of trips in the candidate trip list is less than or equal to a second threshold, to obtain the target trip list. Optionally, for the current preference dimension to be processed, one or more refined sub-conditions are generated based on a preset domain knowledge base or user profile, and a subset of trips that only satisfy the refined sub-conditions is determined; in subsequent intersection operations, the subset of trips is used to replace the original deduplicated trip list for that preference dimension; if the number of trips in the updated candidate trip list is still greater than the second threshold, the next preference dimension, sorted in descending order of preference weight, is processed until the number of trips in the candidate trip list is less than or equal to the second threshold.

[0046] Specifically, after performing the intersection operation of multiple preference dimensions, if the number of trips in the resulting candidate trip list exceeds a preset second threshold (e.g., 30 trips), the current recommendation result is considered too broad, potentially leading to an excessive decision-making burden for users or diluting key needs. To optimize the quantity and quality of recommendation results, a preference constraint upgrade mechanism is initiated. By gradually increasing the filtering conditions of high-weight preference dimensions, redundant options are effectively compressed while retaining core user preferences, thereby improving the accuracy and discriminative power of the recommendation results. The specific implementation is as follows: First, all preference dimensions and their corresponding preference weights are obtained and sorted in descending order of weight to form a sequence to be processed. Here, preference dimensions refer to the trip attribute categories that users focus on (such as "family-friendly," "cultural experience," "transportation convenience," etc.), and preference weights reflect the degree of importance users attach to that dimension; higher values ​​indicate greater importance. For example, suppose the user's target needs include four preference dimensions: "family-friendly" (weight 0.85), "cultural experience" (weight 0.70), "transportation convenience" (weight 0.40), and "price sensitivity" (weight 0.25). The processing order would be: "family-friendly" → "cultural experience" → "transportation convenience" → "price sensitivity".

[0047] Subsequently, the preference dimension with the highest weight is extracted from the sequence as the current item to be processed, and one or more refined sub-conditions are generated based on a preset domain knowledge base or user profile. These refined sub-conditions aim to further clarify the specific requirements of the preference dimension. For example, "cultural experience" can be subdivided into subcategories such as visiting museums, watching performances, or exploring intangible cultural heritage. For each refined sub-condition, a subset of itineraries that only satisfies that condition is determined. Taking "cultural experience" as an example, two subsets may be generated: one is a set of itineraries that includes museum visits, and the other is a set of itineraries that includes live performances. Next, these refined itinerary subsets are used to replace the original deduplicated itinerary list of the preference dimension, and the intersection operation is re-performed. For example, the original intersection operation involves a list of itineraries with four dimensions: L1 (family) ∩ L2 (culture) ∩ L3 (transportation) ∩ L4 (price). Now, for "cultural experience", one of the subsets (such as the museum visit set L2') is selected to replace the original L2, and a new intersection is recalculated: L1 ∩ L2' ∩ L3 ∩ L4. The purpose of this is to make the final recommendation results more focused on specific sub-fields, thereby improving the relevance and accuracy of the recommended content.

[0048] After completing the above operations, the intersection operation is re-executed based on the updated itinerary list set to generate a new candidate itinerary list, and the number of itineraries in the list is counted. If the number of itineraries in the new list is still greater than the second threshold, the next highest-weighted preference dimension is selected from the sorted sequence, and the above process of refining conditions generation, subset replacement, and reintersection is repeated. This loop continues until the number of candidate itineraries is less than or equal to the second threshold, or all high-weight dimensions have been processed. Through this mechanism, the screening criteria for the most important preference dimensions for users are prioritized, ensuring that the recommendation results closely align with the actual needs of users. At the same time, since only one dimension's criteria are adjusted each time and the refined conditions are generated based on real-world domain knowledge or user behavior, the unreasonable exclusion of high-quality itineraries due to excessive constraints on low-weight dimensions is avoided. In addition, the second threshold can be flexibly set according to the display interface capacity or interaction scenario (e.g., 20 items for information flow recommendations and 10 items for detail page recommendations), and the similarity threshold adjustment can be combined with the semantic understanding capabilities of large models (e.g., LLM parsing the specific content of "cultural experience") to further improve the screening accuracy. This not only solves the problem of too many but scattered recommendation results, but also significantly improves user experience and conversion efficiency.

[0049] It is important to emphasize that the dynamic adjustment of preference dimension constraints in this solution is not simply adding or removing filtering rules, but rather a strategic and interpretable optimization operation based on the priority of needs reflected by user preference weights. Its design logic and actual effect have clear technical necessity. When the number of trips in the candidate itinerary list is less than the first threshold, it indicates that the simultaneous application of intersection constraints to multiple preference dimensions (i.e., requiring candidate trips to simultaneously meet the adaptation conditions of all dimensions) leads to overly strict filtering, excessive rejection of effective results, and even situations with "no recommendations" or "only 1-2 trips," severely impacting user experience. In this case, the recall scope must be expanded without significantly deviating from the user's true intent. Therefore, this solution adopts a strategy of discarding the intersection constraints of corresponding preference dimensions in ascending order of preference weight. The reason for prioritizing the discard of low-weight dimensions is that these dimensions represent relatively minor, non-core interests expressed by the user (e.g., "price sensitivity" has a weight of 0.2, while "family-friendly" has a weight of 0.85), and their absence has a limited impact on overall trip satisfaction. By gradually removing these low-priority constraints, the system can effectively expand the candidate set size with minimal sacrifice of core user preferences (i.e., high-weight dimensions are still strictly preserved), avoiding the loss of practicality due to excessive pursuit of "full-dimensional matching." This strategy is essentially a user-value-oriented constraint relaxation mechanism, solving the technical flaw of traditional multi-condition intersection in sparse scenarios where a "better to have fewer but better results" approach leads to "no usable results." Similarly, when the number of candidate trips exceeds the second threshold, while the options appear sufficient, there are often many redundant trips with highly similar semantics or experiences, increasing the user's decision-making burden. Simply truncating the lower-ranked results at this point might mistakenly delete high-quality options; while retaining all without distinction would decrease the recommendation quality. Therefore, this solution adopts a strategy of sequentially increasing the additional intersection constraints of the corresponding preference dimensions according to their preference weights from high to low. The "additional intersection constraints" here do not introduce new dimensions, but rather refine the adaptation requirements based on domain knowledge or user profiles under the high-weight dimensions that users care about most (such as "family-friendly" and "cultural experience"). For example, "family-friendly" can be refined to "includes children's play facilities and alcohol-free services." Since these refined conditions are directly related to users' core needs, strengthening their constraints can effectively eliminate marginal options that are "seemingly relevant but not actually relevant," thereby improving recommendation accuracy without weakening the coverage of core needs. This strategy avoids excessive intervention in low-weight dimensions (such as "transportation convenience") that could lead to the misfiltering of high-quality itineraries, ensuring that the compression process always revolves around the user's true concerns.In summary, whether it is a downgrade or upgrade mechanism, the adjustment order (low weight → high weight or high weight → low weight) and operation method (discarding constraints or adding refined constraints) strictly follow the principle of "preserving the core and adjusting the periphery". It has clear technical motivation and user experience basis, and can effectively overcome the inherent contradiction in multi-dimensional intersection recommendation that "too strict will not work and too lenient will be chaotic".

[0050] In some embodiments, to enhance the richness of the solution, after obtaining the candidate itinerary list, a downgrade mechanism or an upgrade mechanism can be initiated based on a similarity threshold. Optionally, if the number of itineraries in the candidate itinerary list is less than a first threshold, a downgrade mechanism is initiated, lowering the target similarity threshold of the corresponding preference dimension in order of increasing preference weight, and re-performing deduplication on the initial itinerary list of that preference dimension based on the lowered target similarity threshold, until the number of itineraries in the candidate itinerary list is greater than or equal to the first threshold; if the number of itineraries in the candidate itinerary list is greater than a second threshold, an upgrade mechanism is initiated, raising the target similarity threshold of the corresponding preference dimension in order of decreasing preference weight, and re-performing deduplication on the initial itinerary list of that preference dimension based on the raised target similarity threshold, until the number of itineraries in the candidate itinerary list is less than or equal to the second threshold, thus obtaining the target itinerary list; wherein, the second threshold is greater than or equal to the first threshold, and when the semantic similarity between two itineraries is greater than or equal to the currently used target similarity threshold, they are determined to be duplicate itineraries and deduplication is performed.

[0051] Specifically, after completing the intersection operation of multiple preference dimensions and obtaining a candidate itinerary list, the system further determines whether the number of itineraries in the list falls within a preset reasonable range. This range is defined by two thresholds: a first threshold (e.g., 5 itineraries) serves as a lower limit to ensure basic recommendation coverage; a second threshold (e.g., 30 itineraries, and not less than the first threshold) serves as an upper limit to prevent excessive redundancy. If the number of candidate itineraries is lower than the first threshold or higher than the second threshold, a degradation mechanism or an upgrade mechanism will be activated respectively. This involves dynamically adjusting the target similarity threshold for each preference dimension and re-performing deduplication on the initial itinerary list for the corresponding dimension based on the updated threshold, thereby finely controlling the scale and quality of the final output itineraries. When the number of itineraries in the candidate itinerary list is less than the first threshold, it is determined that the current results are too few and may not meet the user's basic browsing or selection needs, and a degradation mechanism is then activated. The core of this mechanism is to relax the deduplication criteria for low-importance dimensions to expand the recall scope. The specific implementation is as follows: First, obtain all preference dimensions and their corresponding preference weights. Preference dimensions refer to the categories of trip attributes that users pay attention to (such as "price sensitivity", "family-friendly", etc.), and preference weights reflect the degree of importance that users attach to the dimension.

[0052] Subsequently, these dimensions are sorted by weight from low to high, with the lowest-weighted dimension processed first. For this dimension, its target similarity threshold is increased (e.g., from 0.85 to 0.92). Here, the target similarity threshold is a critical value used to determine whether two trips are semantically duplicated; semantic similarity is obtained by calculating the cosine similarity between the multimodal embedding vectors of the two trips (integrating text descriptions, image features, structured labels, etc.). Increasing the threshold means that only trips with high semantic consistency will be judged as duplicates, thereby reducing the number of duplicates to be removed. We then revert to the original, un-duplicated initial trip list for this dimension and re-perform deduplication based on the new threshold: we iterate through all trip pairs in the list, and if their semantic similarity is greater than or equal to the new threshold, they are considered duplicates, and only the one with higher overall quality is retained (e.g., sorted by user historical click-through rate, service rating, or commercialization index). The new trip list obtained after deduplication is then incorporated into the intersection operation with other dimensions to generate an updated candidate trip list. If the number is still insufficient, the same operation is performed on the next low-weighted dimension until the number of candidate trips reaches or exceeds the first threshold. Conversely, when the number of trips in the candidate trip list exceeds the second threshold, the recommendation result is deemed too broad and may burden the user's decision-making, and an upgrade mechanism is then activated.

[0053] This mechanism aims to strengthen the deduplication criteria for high-importance dimensions and improve the refinement of recommendations. Specifically, it sorts preference dimensions by weight from highest to lowest, prioritizing the dimension with the highest weight. For this dimension, its target similarity threshold is lowered (e.g., from 0.85 to 0.75). After lowering the threshold, itineraries that are semantically similar but not identical are more likely to be classified as duplicates (e.g., "Beach Resort Hotel A" and "Beach Resort Hotel B" are considered redundant due to their highly similar location and facilities). The process then reverts to the initial itinerary list for this dimension, and deduplication is re-executed based on the new threshold: for any two itineraries, if their semantic similarity is greater than or equal to the lowered threshold, they are considered duplicates, and the better one is retained. The size of the deduplicated itinerary list is thus significantly reduced, and it re-participates in the intersection operation. If the number of candidate itineraries after the update still exceeds the second threshold, the same process continues for the next high-weight dimension until the number falls below the threshold. Throughout the process, the rule for determining duplicate itineraries remains consistent: as long as the semantic similarity between two itineraries is greater than or equal to the currently used target similarity threshold, they are considered duplicates, and deduplication is performed. This mechanism ensures the interpretability and consistency of the control process. Through the aforementioned dynamic threshold adjustment strategy, core preferences can be prioritized and recall expanded when the number of recommendations is insufficient, while key dimensions can be focused on and redundancy compressed when the number of recommendations is excessive. This allows for adaptive output of trip recommendation results that are reasonably sized and accurately targeted in different scenarios, significantly improving user experience and robustness.

[0054] Furthermore, after completing the initial itinerary screening and generating a candidate itinerary list, the overall quality of the list is further evaluated to determine whether a downgrade or upgrade mechanism needs to be activated. In addition to focusing on whether the number of itineraries meets the basic requirements, several global quality indicators, such as diversity score, coverage of high-weight preference dimensions, proportion of semantically repetitive clusters, and novelty score, can be introduced as additional triggering conditions. Optionally, the conditions for activating the downgrade mechanism also include any of the following: the diversity score of the candidate itinerary list is lower than a preset diversity threshold; the coverage of high-weight preference dimensions of the candidate itinerary list is lower than a preset coverage threshold. The conditions for initiating the upgrade mechanism also include any of the following: the proportion of semantically repetitive trip clusters in the candidate trip list exceeds a preset repetition ratio threshold; the novelty score of the candidate trip list is lower than a preset novelty threshold; wherein, the diversity score is a global diversity measure determined based on the semantic or attribute differences among all trips in the candidate trip list; the high-weight preference dimension coverage is the coverage ratio of the candidate trip list as a whole to the high-weight preference dimension set by the user; the novelty score is a global metric used to measure the novelty of the candidate trip list, determined based on the overall difference between the candidate trip list and the preset mainstream trip set in semantic or feature space.

[0055] Specifically, the downgrade mechanism is triggered when the diversity score of the candidate itinerary list falls below a preset diversity threshold. Here, the diversity score measures the overall semantic or attribute-level differences among all itineraries in the candidate itinerary list, reflecting the richness and experience level of the recommendation results. The implementation is as follows: First, the system maps each itinerary to a multimodal semantic embedding vector that integrates text descriptions, image features, and structured attributes (such as destination, activity type, accommodation level, price range, etc.). Then, it calculates the pairwise semantic similarity (e.g., using cosine similarity) between all itineraries in the list and derives the average pairwise distance based on these similarity values. Finally, it normalizes this to a diversity score between 0 and 1; the higher the score, the greater the differences between itineraries and the stronger the diversity. If the score is below the preset diversity threshold (e.g., 0.65), even if the number of itineraries meets the basic requirements, the system still determines that the recommended content is too homogeneous, thus triggering the downgrade mechanism. This involves relaxing the constraints of low-weight preference dimensions and introducing more types of itinerary options to enhance overall diversity. When the coverage of the candidate itinerary list for high-weight preference dimensions is lower than the preset coverage threshold, the system also activates the degradation mechanism.

[0056] Additionally, high-weight preference dimensions refer to preference dimensions whose user preference weight is higher than a certain set value (e.g., 0.7) or ranked in the top K (e.g., top 2) by weight, representing the user's core needs. High-weight preference dimension coverage is defined as the proportion of all high-weight preference dimensions that are effectively satisfied by at least one trip in the candidate trip list. The system determines for each high-weight dimension whether there is at least one trip that highly matches that dimension in its semantic description or structured label (e.g., semantic similarity ≥ 0.8). If the final coverage is lower than a preset threshold (e.g., 0.8), it indicates that although the number of trips is sufficient, key needs are not being adequately addressed. The system will then activate a degradation mechanism to expand the search space to fill in the missing core preferences.

[0057] Specifically, the upgrade mechanism is triggered when the proportion of semantically repetitive trip clusters in the candidate trip list exceeds a preset repetition threshold. To identify implicit redundancy, the system performs clustering analysis based on the semantic embedding vectors of the trips (e.g., using DBSCAN or hierarchical clustering), grouping trips with semantic similarity higher than a certain threshold (e.g., 0.85) into the same semantically repetitive cluster. Then, the proportion of trips in the largest cluster to the entire candidate list is calculated. If this proportion exceeds the preset repetition threshold (e.g., 0.6), it indicates that while the recommendation results meet the quantity requirement, the substantive content is highly concentrated in a few patterns, lacking distinctiveness. In this case, the system activates the upgrade mechanism, strengthening the deduplication standard by increasing the target similarity threshold of the high-weight preference dimension, further filtering out semantically similar redundant items. The upgrade mechanism is also activated when the novelty score of the candidate trip list is lower than a preset novelty threshold.

[0058] In addition, the novelty score measures the degree of differentiation of the candidate itinerary list relative to mainstream recommendations, reflecting its ability to provide unique or niche experiences. This score is determined based on the overall semantic or feature space difference between the candidate itinerary and a preset set of mainstream itineraries. The set of mainstream itineraries can be defined as the set of popular itineraries that rank highly in terms of recent click-through rate, conversion rate, or commercialization index on the platform. The system calculates the minimum semantic distance (such as Euclidean or cosine distance) between each candidate itinerary and all itineraries in the mainstream itinerary set, and then averages the distances across all candidate itineraries to obtain the overall novelty score—the higher the score, the more the itinerary deviates from the mainstream and the more unique it is. If the score is lower than a preset novelty threshold (e.g., 0.45), it indicates that the recommendation results tend to be conventional, lacking cultural depth or personalized highlights. The system will then activate an upgrade mechanism, increasing the weight of highly differentiated itineraries in subsequent intersection or ranking stages to enhance the novelty and exploratory value of the recommendations. In summary, by combining itinerary quantity with multi-dimensional quality indicators such as diversity, core demand coverage, repetitive cluster ratio, and novelty, the system constructs a dynamic and adaptive control mechanism. This mechanism can effectively address complex scenarios such as "sufficient quantity but poor quality" or "insufficient quantity and incomplete coverage," ensuring that the final output of the target itinerary list closely matches the user's real needs while possessing good diversity, differentiation, and personalization, significantly improving user experience and decision-making efficiency.

[0059] For more specific implementation methods of steps not mentioned above, please refer to the following: Figure 2a , Figure 2b , Figure 2c , Figure 2d The relevant descriptions of the embodiments will not be repeated here.

[0060] At this point, Figure 1a , Figure 1b Once the relevant itinerary information processing methods are completed, deep fusion of multimodal information and collaborative enhancement of multi-turn dialogue context are achieved within the same link, significantly improving the relevance of itinerary recommendations to users' real needs, providing users with more accurate, intelligent, and coherent itinerary planning services, and reshaping the cognitive and interaction paradigm of multimodal travel assistants.

[0061] The above Figure 1a , Figure 1b The relevant embodiments are a detailed explanation of the deduplication scheme for itinerary information. Before deduplicating the itinerary information, it is actually necessary to first determine each piece of itinerary information (candidate itinerary list), and then deduplicate to obtain the target itinerary list. The following is based on... Figure 2a , Figure 2b , Figure 2c , Figure 2d The process of determining the candidate itinerary list is described in detail.

[0062] Figure 2aA flowchart illustrating a trip information processing method provided for an exemplary embodiment of this application. Figure 2a As shown, the method includes: 201. During multiple rounds of dialogue with the user regarding this trip planning, obtain the user's initial multimodal needs information involved in this round of dialogue; 202. Call each single-mode encoding module of the multimodal encoder to extract key information from the initial requirement information of the corresponding single mode, and obtain the initial key information set of each single mode; 203. Input the initial key information set of each single modality into the first large model, perform demand information completeness analysis and initial travel demand category analysis, and obtain demand information completeness and initial travel demand category; 204. When the completeness of the demand information is less than the first threshold, recall the historical demand summary and / or historical session summary that match the initial trip demand category from the long short-term memory storage area, and generate supplementary demand information for the initial demand information based on the historical demand summary and / or historical session summary to obtain the supplemented multimodal target demand information; the multimodal target demand information has a target trip demand category. 205. Call the multimodal fusion encoding module of the multimodal encoder, and with the target travel requirement category as a reference, integrate the target requirement information of each single mode and perform preference analysis to obtain the user's target requirement preference information; 206. Input the target demand preference information into the information recommendation model to match and integrate the itinerary information to obtain the target itinerary list.

[0063] The technical solution provided in steps 201-206 of this application achieves a paradigm shift from "static keyword analysis" to "dynamic multimodal semantic understanding" by constructing a multimodal encoder and a dual-model collaborative mechanism, overcoming the fundamental drawback of traditional travel assistants lacking context awareness and memory capabilities. By combining single-modal encoding with a fusion module, the first model analyzes the completeness of needs, while the second model supplements this by recalling relevant historical summaries based on the suitability of itinerary needs categories. Iterative optimization is driven by a completeness threshold, enabling the system to accurately extract deep multimodal semantics while intelligently filtering related historical information rather than mechanically fusing all summaries. Ultimately, this achieves deep fusion of multimodal information and synergistic improvement of multi-turn dialogue context within the same link, significantly increasing the relevance of itinerary recommendations to users' actual needs, providing users with more accurate, intelligent, and coherent itinerary planning services, and reshaping the cognitive and interactive paradigm of multimodal travel assistants.

[0064] In this embodiment, the trip information processing method can be executed through a target application installed on the user's terminal. The target application can be a standalone app, a mini-program that depends on an app, or a webpage. When a user has trip planning needs, the target application can be used to plan their trip. The target application has a conversational function and can provide a conversational interface to the user.

[0065] In step 201, the user can engage in multi-turn dialogues via the conversational interface, inputting multimodal initial requirements during the dialogue. Correspondingly, during the multi-turn dialogues with the user regarding this trip planning, the multimodal initial requirements information involved in this round of dialogue can be obtained. This multimodal initial requirements information refers to the raw, unprocessed requirement expression data submitted by the user to the system through one or more input channels when initiating a trip planning task. This information is characterized by multi-source heterogeneity and semantic complementarity, and includes, but is not limited to, one or more combinations of text, voice, image, and location information. Text information is a natural language description provided by the user through keyboard input, speech transcription, or other methods, such as "I want to take my parents to Xi'an for two days, preferably with historical and cultural attractions," or "Budget under 3000," etc. Moreover, text information typically includes explicit entities (such as location, time, people), intent keywords, and constraints. Image information includes photos, screenshots, or visual content selected from the user's album, such as a photo of the Terracotta Warriors, a hotel room screenshot, or a map marker. Image information can also implicitly reveal user interests, scene types, or specific target objects, requiring analysis using computer vision technology. Voice information consists of user-recorded audio segments, the raw audio signal without transcription. Voice information not only contains semantic content (which can be converted to text through speech recognition) but also carries paralinguistic features such as speech rate, pause frequency, intonation variations, and pronunciation clarity. These features can reflect the user's urgency, emotional state, or the intensity of their interaction intent. Location information can include the user's current location and / or destination information. Furthermore, this trip planning can be a schedule the user intends to complete within a series of interactive tasks. Its content can be a single, complete trip (e.g., "planning a two-day trip to Chengdu next weekend") or multiple sub-trips (e.g., "visiting a museum in the morning, exploring Kuanzhai Alley in the afternoon" or "hiking on the first day, soaking in hot springs on the second day"). Around this overall planning goal, the specific needs of each part are gradually refined through multiple rounds of interaction with the user. A multi-turn dialogue can be a single conversation cycle conducted to complete the aforementioned itinerary planning, involving multiple alternating inputs and responses between the user and the system. During this process, the user may express their intentions for different sub-trips at different stages, and the system guides the clarification of needs through follow-up questions, recommendations, or confirmations. All interaction history collectively forms the contextual basis for understanding the current intention. "This turn of dialogue" can specifically refer to the interaction round corresponding to the currently processed user input within a multi-turn dialogue, that is, the multimodal information (text, voice, images, etc.) provided by the user at the current moment regarding a specific trip or sub-trip.The system uses the content of the current dialogue as an immediate demand signal, combines the historical context of multiple dialogues with the summary information in the long and short memory storage area, and generates target demand preference information that matches the current itinerary segment, thereby achieving step-by-step and accurate modeling of complex itinerary planning tasks.

[0066] After obtaining the user's multimodal initial requirements information, a multimodal encoder can be invoked to encode this information. This multimodal initial requirements information contains multiple unimodal initial requirements information. For example, when the multimodal initial requirements information includes text, voice, images, and geolocation information, text corresponds to one type of unimodal initial requirements information, voice corresponds to one type, images correspond to one type, and geolocation information corresponds to one type. The multimodal encoder contains multiple unimodal encoders, each used to encode information from different modalities. Each unimodal encoder processes different types of information; some unimodal encoders process only text, some only voice, some only images, and some only location information.

[0067] Based on this, referring to step 202, the individual modal encoding modules of the multimodal encoder can be invoked to extract key information from the initial requirement information of the corresponding modality, thus obtaining the initial key information set of each modality. The self-attention mechanism refers to the process within each individual modal encoding module where the input sequence (text, voice, image, or geographic location information) is first linearly mapped to generate three matrices: Query (Q), Key (K), and Value (V). The correlation weights between elements in the sequence are calculated by the dot product of Q and K, and then V is weighted, summed, and the residual is normalized. This allows for the capture of long-range dependencies within the same sequence without an external queue, highlighting key elements relevant to the current task. Here, Q refers to the linear transformation vector of the element to be focused on, used to "ask" within the sequence and find the most relevant content; K refers to the linear transformation vector of each element in the sequence, acting as an "index" for Q to compare and determine the degree of relevance; and V refers to another linear transformation vector of each element in the sequence, representing the actual information content. The output is weighted according to the similarity of Q and K, forming the final self-attention result. The initial key information for each single modality refers to the set of high-dimensional feature vectors extracted from text, voice, image, or location information after self-attention weighted aggregation. These vectors represent the core semantic entities, important acoustic events, key visual objects, and key locations (GPS coordinates, city / POI name, area type label (scenic spot, business district, transportation hub), and relative distance vectors to the user's historical location, used for subsequent trip range matching and geographical constraint judgment) in the user's request. They are used for subsequent cross-modal fusion and demand completeness judgment.

[0068] In some alternative embodiments, the multimodal encoder in step 202 above may employ... Figure 2b The unified encoder architecture is shown below. Figure 2b As shown, the architecture includes: a multimodal input receiver layer, a single-modal encoder layer, a cross-modal fusion layer, and a unified output layer.

[0069] The multimodal input receiving layer is used to receive the user's multimodal initial demand information in parallel. It includes four parallel interfaces: text information input processing, image input processing, voice information input processing, and geolocation input processing. It can receive different modal data such as text such as "I want to go to the beach to relax", beach vacation photos, voice descriptions of needs, and current GPS coordinates.

[0070] The dedicated encoder layer deploys four dedicated single-modal encoding modules, specifically: a BERT text encoder for language semantic understanding, mapping text to a 768-dimensional vector; a ViT visual encoder for image scene recognition, mapping images to a 2048-dimensional vector; a Wav2Vec speech encoder for speech intent recognition, mapping speech to a 512-dimensional vector; and a geolocation encoder for spatial information encoding, mapping GPS coordinates to a 128-dimensional vector. Each encoder operates independently and can be deployed on terminals, edge nodes, or cloud nodes. Its accuracy level (lightweight / half-precision / full-precision) is dynamically assigned by the task scheduler based on the constraint parameters of the S3 step.

[0071] The cross-modal fusion layer includes a multimodal fusion coding module (or semantic fusion unit), which aligns four heterogeneous vectors into a unified vector space based on a cross-attention mechanism and achieves deep fusion of cross-modal information through gating weighting, ultimately outputting a 768-dimensional multimodal semantic vector. This vector incorporates the semantic essence of textual, visual, acoustic, and spatial information, serving as the input basis for subsequent requirement completeness analysis and preference extraction.

[0072] It should be noted that, Figure 2b The architecture shown is only one possible implementation of a multimodal encoder. Its encoder type, vector dimension, and fusion strategy can be flexibly adjusted according to hardware or software resources.

[0073] In some embodiments, during step 202, each unimodal encoding module can first be invoked to perform semantic parsing on the initial requirement information set of the corresponding unimodal to obtain a sequence of semantic sub-fragments of each unimodal; further, self-attention calculation is performed on each semantic sub-fragment in the sequence of semantic sub-fragments of each unimodal, and during the self-attention calculation, relevance weights are assigned among the semantic sub-fragments in the sequence of semantic sub-fragments of each unimodal; then, based on the relevance weights among the semantic sub-fragments in the sequence of semantic sub-fragments of each unimodal, the semantic sub-fragments in the sequence of semantic sub-fragments of the corresponding unimodal are aggregated to obtain a context-aware feature representation of each unimodal; then, key feature representations are selected from the context-aware feature representations of each unimodal as the initial key information set of each unimodal.

[0074] Optionally, each single-modal coding module is invoked to perform semantic parsing on the initial requirement information set of the corresponding single modality to obtain the semantic sub-segment sequence of each single modality. This includes: invoking the corresponding single-modal coding module to perform segmentation processing on the input initial requirement information set, dividing the text into word sequences, the image into grid block sequences, and the speech into frame sequences, and performing position encoding and shallow linear transformation on each segment to obtain a semantic sub-segment sequence with positional order, which is used for subsequent self-attention calculation.

[0075] Optionally, when performing self-attention calculation on each semantic sub-segment in each unimodal semantic sub-segment sequence, and when allocating relevance weights among each semantic sub-segment in each unimodal semantic sub-segment sequence during the self-attention calculation process, the following steps are included: performing a linear mapping on the unimodal semantic sub-segment sequence to obtain three matrices: Query, Key, and Value; calculating the relevance score between each sub-segment using the dot product of Query and Key, and obtaining the relevance weights after scaling and SoftMax normalization.

[0076] Optionally, based on the correlation weights between semantic sub-segments in the semantic sub-segment sequence of each unimodality, the semantic sub-segments in the corresponding unimodality semantic sub-segment sequence are aggregated to obtain the context-aware feature representation of each unimodality. This includes: based on the correlation weights between semantic sub-segments in the semantic sub-segment sequence of each unimodality, the semantic sub-segments in the corresponding unimodality semantic sub-segment sequence are aggregated to obtain the context-aware feature representation sequence of each unimodality. Here, the context-aware feature representation of each unimodality refers to the set of feature vectors output after self-attention weighting, where each vector corresponds to a position of the original semantic sub-segment and has incorporated global context information of the same modality, used to represent the key semantic, visual, or acoustic context of that position in the text, image, or speech sequence. Additionally, the context-aware feature representation corresponding to the geographic location information is used to represent the longitude, latitude, city / POI name, region type (scenic spot, business district, transportation hub), and relative distance relationship with the user's historical location of the user's current or target location, providing spatial context for subsequent trip range matching and geographic constraint judgment.

[0077] Optionally, key feature representations are selected from the context-aware feature representations of each single modality as the initial key information set of each single modality. This includes: scoring the importance of the context-aware feature representations of each single modality, and selecting the top K feature vectors with the highest scores as the initial key information set of that single modality for subsequent cross-modal fusion and demand analysis, where k is a positive integer.

[0078] Alternatively, when each single-modal coding module extracts key information from the initial requirement information of the corresponding single modality, the specific implementation of this extraction process can be as follows: Figure 2cThe internal processing flow shown is as follows: Specifically, each single-modal encoding module independently runs the following processing branches: Text information encoding branch: Performs intelligent word segmentation, part-of-speech tagging, named entity recognition, sentiment semantic analysis, and intent understanding reasoning on the original text input to obtain the initial key information set for this single modality. Image encoding branch: Performs quality enhancement, scene segmentation and recognition, object detection and localization, and style and atmosphere analysis on the original image to obtain the initial key information set for this single modality. Speech information encoding branch: Performs noise reduction, speech feature extraction, speech-to-text, intonation and sentiment analysis, and speech intent understanding on the original audio data to obtain the initial key information set for this single modality. Geolocation encoding branch: Performs geolocation transformation, location semantic understanding, point of interest matching, and transportation accessibility processing on GPS coordinate data to obtain the initial key information set for this single modality. It should be noted that... Figure 2c The processing flow shown is only one optional implementation method for each single-mode encoder module, and its specific processing steps can be tailored according to the resource capabilities of the deployment node.

[0079] The aforementioned embodiment of step 202 preserves the contextual information of each semantic location through self-attention weighting, enabling text, images, and speech sequences to output "location-level" key features. This provides fine-grained representations for subsequent cross-location and cross-modal filtering, taking into account both global context and local focus. Simultaneously, the intermediate tensor of multimodal encoding is upgraded from a single vector segment to a location-level feature sequence, inheriting the global context and laying the foundation for the next step of selecting key features. This achieves a natural transition from semantic depth to feature precision, paving the way for improving the relevance of trip recommendations to users' actual needs and providing users with more accurate, intelligent, and coherent trip planning services.

[0080] After obtaining the initial key information from each single modality, refer to step 203 to input the initial key information set from each single modality into the first major model for demand information completeness analysis and initial itinerary demand category analysis, obtaining demand information completeness and initial itinerary demand categories. Demand information completeness quantifies the degree of completeness of user multimodal input in key dimensions such as budget, time, travel companions, interest type, and destination, and can be represented by a score of 0–1, with higher scores indicating more complete information. Initial itinerary demand categories refer to itinerary template tags (such as "city weekend trip," "island honeymoon," and "business trip") obtained by mapping keywords from text / image / voice information, used to locate target process templates in the knowledge base. The first major model can be the demand analysis model mentioned in the above embodiments. Furthermore, for a description of the first major model, please refer to the content of the above embodiments, which will not be repeated here.

[0081] In some optional embodiments, each initial key information in the initial key information set of each unimodal has its own initial weight. The initial weight is a normalized relevance score calculated by the self-attention mechanism for each initial key information (feature vector), used to quantify the relative importance of the information in cross-modal fusion. The higher the score, the greater the contribution to subsequent demand understanding. Based on this, the initial key information sets of each unimodal are input into the first large model to perform demand information completeness analysis and initial travel demand category analysis to obtain demand information completeness and initial travel demand categories. This includes: performing weighted pooling operations on each initial key information in the initial key information set of each unimodal according to the corresponding initial weight to obtain cross-modal key information; calculating the similarity between the cross-modal key information and each preset key information dimension label; determining the demand information completeness jointly expressed by the initial key information sets of each unimodal based on the similarity between the cross-modal key information and each preset key information dimension label; and performing matching operations on the cross-modal key information and each travel demand category in the preset travel demand category set to obtain the initial travel demand category jointly expressed by the initial key information sets of each unimodal.

[0082] Optionally, based on the similarity between cross-modal key information and each preset key information dimension label, the completeness of the demand information jointly expressed by the initial key information sets of each single modality is determined, including: calculating the similarity between cross-modal key information and each preset key dimension label to obtain the matching score of each dimension; weighting and summing the matching scores of each dimension and normalizing them to quantify the completeness of the demand information.

[0083] In the aforementioned embodiment of step 203, through weighted pooling, dimensional similarity calculation, and normalized quantization, key information scattered across various modalities is mapped in one go to the completeness of demand information and the initial trip demand category, achieving quantification from cross-modal semantics to comparable numerical values. Furthermore, the transformation from "feature set" to "comparable index" is completed. The output completeness score and category label directly serve as input for subsequent threshold judgment and template matching, providing a quantitative basis for "demand completion → process triggering," realizing the transition from semantic depth to decision-making accuracy. This lays the groundwork for improving the relevance of trip recommendations to users' actual needs and providing users with more accurate, intelligent, and coherent trip planning services.

[0084] In this embodiment, the completeness of the initial demand information of the multimodal system may be low. Based on this, after obtaining the completeness of the initial demand information and the initial trip demand category of the multimodal system, referring to step 204, when the completeness of the demand information is less than a first threshold, historical demand summaries and / or historical session summaries that match the initial trip demand category are recalled from the long short-term memory storage area, and supplementary demand information of the initial demand information is generated based on the historical demand summaries and / or historical session summaries to obtain the supplemented target demand information of the multimodal system; the target demand information of the multimodal system has a target trip demand category.

[0085] Specifically, based on the Long Short-Term Memory (LSTM) storage mechanism, the following recall operation can be performed from the LSM storage area (or multi-category memory storage area) to improve the initial demand information of the multimodal system: recalling historical demand summaries that match the initial travel demand category as first supplementary demand information, and / or recalling historical conversation summaries that match the initial travel demand category as context information for the current round of dialogue, and using a second major model to obtain second supplementary demand information based on the context information, until the demand information completeness reaches a first threshold, to obtain the target demand information of the multimodal system, which corresponds to the target travel demand category. The second major model can be the AI ​​dialogue model mentioned in the above embodiments. If the demand information completeness is greater than or equal to the first threshold, there is no need to further supplement the initial demand information of the multimodal system using the LSM storage mechanism.

[0086] The long and short-term memory (LSM) storage area comprises a short-term memory storage area (or first memory storage area) and a long-term memory storage area (or second memory storage area). The short-term memory storage area stores conversation summaries generated at a preset collection frequency, with the topic category remaining constant and each preset collection period completed. It also stores request summaries and conversation summaries generated when the topic category changes but the preset collection period has not yet been completed. The long-term memory storage area stores request summaries and conversation summaries that have been moved down from the short-term memory storage area after reaching a preset storage duration. Based on this, the LSM storage mechanism is developed using the concepts of temporary and long-term memory. It collects conversation summaries generated during conversations and request summaries from the user's historical activity, and stores both the conversation summaries and historical request summaries. The long short-term memory storage mechanism includes the following: (1) For conversation summaries, for any dialogue related to the user's current trip, starting from the first round of the conversation, the number of conversation rounds is counted in real time. If the topic category remains unchanged and the cumulative number of conversation rounds reaches a preset number, a conversation summary is immediately generated based on all conversation content within the current preset number and written into the short-term memory storage area. The topic category corresponds to the user's initial trip requirement category. If the topic category changes, a conversation summary is generated based on all conversation content before the change that has not reached the preset number of rounds and is added to the short-term memory storage area. When the conversation summary in the short-term memory storage area reaches the first preset storage duration, the conversation summary that has reached the first preset storage duration is moved to the long-term memory storage area. The first preset storage duration can be determined based on the duration of the current trip, such as one day, one week, one month, etc. This embodiment does not limit this. When the conversation summary in the long-term memory storage area reaches the second preset storage duration, the conversation summary that has reached the second preset storage duration can be deleted. The second preset storage duration can be greater than the first preset storage duration, such as several months, one year, etc. (2) For historical demand summaries, a historical demand summary is generated based on historical demand information related to the user's current trip and stored in the short-term memory storage area. When the historical demand summary in the short-term memory storage area reaches the first preset storage duration, the historical demand summary that has reached the first preset storage duration is moved to the long-term memory storage area. Moreover, when retrieving session summaries and historical demand summaries from the long and short-term memory storage areas, the summary is retrieved from the short-term memory storage area first. If it cannot be retrieved or the retrieved summary content is limited, it is then retrieved from the long-term memory storage area. In addition, when storing session summaries and historical demand summaries in the long and short-term memory storage areas, they can be stored based on the initial trip demand category of the multimodal initial demand information to facilitate retrieval based on the initial trip demand information.

[0087] Based on this, historical demand summaries that match the initial travel demand category are recalled from the short-term memory storage area, including: recalling historical demand summaries from the short-term memory storage area based on the initial travel demand category, as the first supplementary demand information for the multimodal initial demand information; if historical demand summaries are not recalled from the short-term memory storage area, or if historical demand summaries are recalled from the short-term memory storage area but the demand completeness of the multimodal initial demand information supplemented based on the recalled historical demand summaries is less than a first threshold, then historical demand summaries are recalled from the long-term memory storage area based on the initial travel demand category, as the first supplementary demand information; if historical demand summaries are recalled from the short-term memory storage area and / or the long-term memory storage area and the completeness of the multimodal initial demand information supplemented based on the recalled historical demand summaries reaches the first threshold, then the multimodal initial demand information before supplementation and the first supplementary demand information are fused to obtain the multimodal target demand information.

[0088] The aforementioned embodiment related to step 204 employs a tiered recall strategy of "short-term first, then long-term, stopping when the threshold is reached" to piece together scattered demand summaries from previous trips back into the current input as needed. This proactively supplements information that is "completed by the system even if the user hasn't explicitly stated it," avoiding the waste of computing power caused by a one-time full recall. Furthermore, it takes the coarse label of "initial trip demand category" and completes the precise mapping to "available historical fragments." The output of the first supplementary demand information is directly fed back into the completeness calculation, driving iterative supplementation or cross-modal fusion, providing data raw materials for "demand iteration → completeness achievement," and realizing the transition from coarse categories to fine demands. This lays the groundwork for improving the relevance of trip recommendations to users' actual needs and providing users with more accurate, intelligent, and coherent trip planning services.

[0089] Additionally, the process of retrieving historical conversation summaries that match the initial travel request category from the short-term and long-term memory storage areas includes: retrieving historical conversation summaries from the short-term memory storage area based on the initial travel request category; if historical conversation summaries are not retrieved from the short-term memory storage area, retrieving historical conversation summaries from the long-term memory storage area based on the initial travel request category; if historical conversation summaries are retrieved from the short-term memory storage area and / or the long-term memory storage area, these historical conversation summaries are used as context information for the current conversation, and a second model is used to provide a question-and-answer service to the user based on the context information to obtain the second supplementary request information provided by the user based on the question-and-answer service; if the completeness of the multimodal initial request information supplemented based on the second supplementary request information reaches a first threshold, the multimodal initial request information before supplementation and the second supplementary request information are fused to obtain the multimodal target request information.

[0090] Optionally, the historical conversation summary is used as the context information for the current conversation. Based on the context information, the second major model generates guiding dialogue and provides a question-and-answer service to the user to obtain the second supplementary requirement information provided by the user based on the question-and-answer service. This includes: inputting the recalled historical conversation summary as context into the second major model, which automatically generates follow-up questions or guiding dialogue based on the default fields in the summary and provides it to the user; parsing the user's response to the follow-up questions and extracting the missing fields from the initial requirement information as the second supplementary requirement information.

[0091] In some optional embodiments, the steps 204 above, namely "recalling based on long short-term memory storage mechanism" and "obtaining supplementary demand information based on contextual information using the second major model", can be implemented using... Figure 2d The context enhancement process is shown below. Figure 2d As shown, the process starts with the user's new demand information input (such as "recommend tomorrow's activities"), and first retrieves historical demand summaries (i.e., trip data), including multi-source data such as historical booking records, browsing behavior, collection reviews and social sharing.

[0092] If relevant historical data exists, in-depth historical analysis is performed to extract user preference patterns (such as "preference for artsy venues" and "price sensitivity"), and contextual correlation analysis is executed to understand the relationship between current needs and historical experiences. If no historical data exists, a general recommendation strategy is used as the basis. Subsequently, the system comprehensively evaluates the current state, integrating multi-dimensional real-time information such as location, time, weather, and budget, and determines whether the geographical location is suitable.

[0093] If the geographical location does not match, the system enhances the geographical context, expands the search scope, and considers accessibility. If the location matches, it directly enters the time-series deep analysis stage, combining factors such as seasons and holidays to identify hidden user preference patterns. Finally, the system generates a context-enhanced demand representation, which is a structured semantic vector that overlays historical preferences, real-time context, and hidden patterns onto the original demand. This vector is then input into the personalized recommendation engine to generate customized suggestions.

[0094] This process organically links the long and short-term memory storage area with the second model, which obtains supplementary requirement information based on contextual information. This achieves proactive requirement completion, moving from "unspoken by the user (vague expression)" to "understood by the system," providing high-quality, context-rich target requirement information for subsequent multimodal fusion coding. It should be noted that... Figure 2d This is just one possible implementation method for the information supplementation stage, but it is not limited to this.

[0095] This embodiment accurately retrieves relevant historical conversation summaries from short-term or long-term memory based on the initial trip demand category, and uses this to drive the second major model to conduct targeted question answering to supplement user needs. This achieves efficient and context-aware improvement of multimodal initial needs. This mechanism specifically implements the core idea of ​​"retrieving historical information for supplementation based on trip demand category adaptability", enabling "dynamic multimodal semantic understanding" and "context-driven demand iteration optimization" to be realized at the practical level. It can lay the foundation for improving the relevance of trip recommendations to users' real needs and providing users with more accurate, intelligent and coherent trip planning services.

[0096] In this embodiment, the multi-dimensional target demand information corresponds to the target travel demand category. Based on this, referring to step 205, the multimodal fusion encoding module of the multimodal encoder is invoked. Taking the target travel demand category as a reference, the target demand information of each single modality is integrated and preference analyzed to obtain the user's target demand preference information. The cross-attention mechanism refers to using a single modality (here, "target travel demand category") as a query during the multimodal fusion process to dynamically calculate the semantic association strength between it and other single-modal target demand information (such as key features extracted from text, images, and voice information), and then weighting and fusing the information from each modality accordingly. The target demand preference information refers to the structured deep user preference representation obtained by the system through fusion encoding after completing multiple rounds of context supplementation and determining the target travel demand category. The preference dimensions include, but are not limited to, one or more of the following: geographical location, service type, price range, user reviews, attraction type, dining style, accommodation level, activity pace, and price sensitivity. Target demand preference information includes, but is not limited to: the user's specific tendencies across multiple preference dimensions, and the quantitative or semantic preference values ​​under each preference dimension. These preference dimensions have been aligned with the target travel demand category and can be used for subsequent precise matching and recommendation of projects. Target demand preference information may include, for example, "preferring quiet > lively", "acceptable per capita cost of less than 500 yuan", "must include children's facilities", etc.

[0097] In some embodiments, the multimodal fusion encoding module of the multimodal encoder is invoked to integrate and analyze the target demand information of each single modality, taking the target travel demand category as a reference, to obtain the user's target demand preference information. This includes: invoking the multimodal fusion encoding module of the multimodal encoder to perform cross-attention calculation on the target demand information of each single modality, taking the target travel demand category as a reference, and assigning global importance weights of each key target information in the target demand information of each single modality relative to the target travel demand category during the cross-attention calculation process; dynamically adjusting the contribution of each key target information in the target demand information of each single modality by combining the global importance weights of each key target information in the target demand of each single modality and the first context constraint condition representing the local context of the current dialogue, to generate a context-aware weighted target key information representation corresponding to each key target information; performing cross-modal fusion on the target key information representation with the target travel demand category as the fusion condition to obtain the fused key information representation; and performing demand preference analysis from multiple dimensions based on the fused key information representation and the second context constraint condition representing the user's global profile or external environment to obtain the target demand preference information.

[0098] It is important to note that the first and second contextual constraints operate at different stages of multimodal fusion, with clear hierarchical division of labor and technical objectives. The first contextual constraint is used for dynamic adjustment of feature-level contributions, operating after cross-attention calculation and before cross-modal fusion. The second contextual constraint is used for decision-level preference dimension generation and filtering, operating in the higher-order demand analysis stage after cross-modal fusion. They belong to different processing levels and work together to achieve semantic guidance from local dialogue context to global user profile. Specifically, the first contextual constraint originates from the local conversational context of the current dialogue, including but not limited to: key entities mentioned by the user in the previous N rounds of dialogue (e.g., "want to go somewhere with the sea," "child is 5 years old"); the current conversation topic category (e.g., "family trip," "business travel"); and the follow-up questioning intent guiding the output of the dialogue generation model (e.g., the transportation preference implied by the system asking "Do you need to be near the subway?"). This constraint is input to the multimodal fusion encoding module in vector form (e.g., embedding representation), and after cross-attention calculation, it is used to dynamically modulate the contribution weights of key information of each single-modal target. For example, if a user previously emphasized "limited budget," the keyword "budget hotel" in the text modality will contribute more, while "luxury villa" identified in the image modality will be suppressed. Its technical role is to ensure that the multimodal fusion process focuses on the semantic fragments most relevant to the current dialogue context, avoiding irrelevant or conflicting information from interfering with the generation of key information representations. The second contextual constraint comes from a broader global user context, including but not limited to: user profile tags (such as "family with children," "senior users," "frequent business travelers"); planned itineraries in the travel calendar (such as "meeting scheduled for tomorrow"); geographical location and device environment (such as "currently at the airport," "using a mobile device"); and real-time external events (such as "destination issues a rainstorm warning"). Based on the fused key information representation, this constraint participates in the multi-dimensional demand preference analysis stage. It filters unreasonable preference dimensions (e.g., recommending "barbecue stalls" to "vegetarians"), enhances preferences that align with the user's long-term profile (e.g., increasing the weight of "museums" and "historical sites" for "culture enthusiasts"), and adjusts preference granularity (e.g., refining "dining" to "quick lunch" instead of "Michelin dinner" in a business scenario). Its technical role is to ensure that the final output of "target demand preference information" not only reflects the content of the current conversation but also aligns with the user's long-term behavioral patterns, the real-world environment, and external constraints, thereby improving the relevance and feasibility of the itinerary recommendation. Furthermore, the object of cross-modal fusion is the representation of each target's key information, i.e., the single-modal feature vector perceptually weighted by the first context constraint, rather than the original input or the original multimodal encoding result.The final target demand preference information is obtained by multi-dimensional preference reasoning based on the fusion representation and the second context constraint conditions, which includes multiple preference dimensions and their corresponding preference weights.

[0099] Optionally, the multimodal fusion encoding module of the multimodal encoder is invoked to perform cross-attention calculation on the target demand information of each single modality, with the target travel demand category as a reference. During the cross-attention calculation, the global importance weight of each target key information in the target demand information of each single modality relative to the target travel demand category is assigned, including: encoding the target travel demand category into a query vector, and encoding the target key information of each single modality into a key vector and a value vector, respectively; calculating the attention score based on the query vector and each key vector, and obtaining the global importance weight by normalization through the Softmax function; and using the global importance weight to perform a weighted summation of the corresponding value vectors to obtain the category-aware weighted representation of each single modality.

[0100] Optionally, by combining the global importance weights of each key information of each target in the target requirements of each single modality and the first context constraint, the contribution of each key information of each target in the target requirements of each single modality is dynamically adjusted to generate a context-aware weighted representation of the key information of each target, including: obtaining the first context constraint, which includes at least one of the following: current geographical location, real-time weather, current time, or short-term session intent; generating a context modulation factor based on the first context constraint; multiplying the context modulation factor by the global importance weight to obtain the dynamically adjusted contribution weight; and weighting each key information of each target based on the dynamically adjusted contribution weight to generate a context-aware weighted representation of the key information of each target.

[0101] Optionally, using the target travel demand category as the fusion condition, cross-modal fusion is performed on the key information representations of each target to obtain the fused key information representation, including: mapping the context-aware weighted target key information representations of each single modality to a unified semantic space; constructing a cross-modal fusion gating signal based on the semantic embedding of the target travel demand category; using the fusion gating signal to perform weighted fusion of the representations of each modality, or performing cross-modal interaction through a multi-head cross-attention mechanism; and outputting the fused key information representation, which represents the semantics of the user's core needs under multimodal collaboration.

[0102] Optionally, based on the fused key information representation and the second context constraint, demand preference analysis is performed from multiple dimensions to obtain target demand preference information, including: obtaining the second context constraint, inputting the fused key information representation and the second context constraint together into the multi-dimensional preference analysis model; the multi-dimensional preference analysis model outputs quantitative preference values ​​on multiple preset preference dimensions to form target demand preference information.

[0103] The aforementioned step 205 and related embodiments specifically implement the key path of how multimodal information is deeply integrated and transformed into structured preferences. That is, through category-guided cross-attention, context-aware dynamic weighting, cross-modal fusion, and multidimensional preference analysis, the abstract goals of "deep integration" and "contextual collaborative enhancement" are transformed into executable and quantifiable technical processes. This provides high-quality preference input for subsequent personalized recommendations, lays the groundwork for improving the relevance of itinerary recommendations to users' real needs, and provides users with more accurate, intelligent, and coherent itinerary planning services.

[0104] Furthermore, to address potential conflicts between users' historical preferences and current travel scenarios (e.g., a user historically prefers "skiing," but is currently in summer or located in a tropical region), a dynamic weight calibration mechanism based on the current spatiotemporal context is introduced before inputting target demand preference information into the information recommendation model. This mechanism is particularly suitable for resolving potential conflicts between users' long-term interests and their current travel environment. For example, a user may have historically shown a strong preference for "skiing" or "winter sports," but is currently in summer and located in a tropical region (such as Sanya, Hainan). Directly using the original preference weights for recommendations could easily generate inappropriate or even infeasible itinerary suggestions (such as recommending ski resorts). Specifically, the mechanism obtains current spatiotemporal context information; for any preference dimension in the target demand preference information, it determines whether it contains seasonal keywords or event-related keywords; if so, it calculates the timeliness factor of that preference dimension based on the current spatiotemporal context information; and adjusts the original preference weights corresponding to that preference dimension according to the timeliness factor to obtain the adjusted preference weights.

[0105] More specifically, the current spatiotemporal context information includes two core elements: first, the current time (accurate to the day, such as July 15, 2025), used to determine the season, holidays, or major event cycles; and second, the user's real-time geographical location (such as latitude and longitude or city name obtained through GPS or IP positioning, such as "Sanya City, Hainan Province"). In addition, a pre-built knowledge base of time-sensitive events is accessed. This is a structured database containing three types of key data: a seasonal label mapping table (e.g., mapping "skiing" and "ice sculpture" to "winter," and "summer retreat" and "seaside vacation" to "summer"), a calendar of major events (e.g., "Harbin Ice and Snow Festival (December 20th to February 28th of the following year)" and "Chengdu Universiade (August 1st–12th, 2025)"), and geographic climate zoning data (indicating whether each region possesses the natural conditions for carrying out specific activities, such as "Sanya has a tropical maritime climate and no snowfall throughout the year"). These data collectively constitute the external basis for judging the timeliness of preferences.

[0106] Next, the system iterates through each preference dimension in the target demand preference information (such as "ice and snow sports," "seaside vacation," "cultural relics," etc.) and checks whether its keywords appear in the seasonality or event-related tag list of the timeliness event knowledge base. If a match is found, the dimension is marked as "timeliness-sensitive." For example, "ice and snow sports" is identified as strongly correlated with winter and dependent on cold geographical environments; "seaside vacation" is marked as strongly correlated with summer and highly adaptable to coastal areas; while "cultural relics" usually does not have obvious seasonal restrictions, so it is considered non-timeliness-sensitive and will not be included in subsequent dynamic adjustments. Subsequently, for each preference dimension marked as timeliness-sensitive, the system calculates a quantitative indicator, namely the timeliness factor (T), to measure the rationality of the preference in the current time and space. Its calculation method comprehensively considers time offset and geographical distance: time offset refers to the difference in the number of days between the current date and the center of the optimal timeliness window of the preference, and geographical distance refers to the number of kilometers between the user's current location and the typical occurrence location of the preference. The two deviations are converted into attenuation coefficients within the [0,1] interval using an attenuation function (such as a Gaussian function or a piecewise linear function), and then weighted and summed with configurable weights (default time weight 0.6) to obtain the final timeliness factor. Taking "ice and snow sports" as an example: the current date is July 15th, the optimal window is from December to February of the following year, a time offset of about 150 days, resulting in a time attenuation coefficient close to 0.05; the user is located in Sanya, about 3200 kilometers away from a typical ski resort (such as Harbin), with a spatial attenuation coefficient of about 0.1; the overall calculated timeliness factor is about 0.08. "Seaside vacation," due to the high temperatures of summer and the user's location in a coastal city, shows a high degree of time and space matching, resulting in a timeliness factor of 1.75, demonstrating its outstanding adaptability in the current scenario.

[0107] Based on this, the original preference weights are updated using multiplicative modulation, which involves multiplying the original weight of each time-sensitive dimension by its corresponding time-sensitivity factor to obtain the adjusted new weights. "Multiplicative modulation" refers to amplifying or suppressing signal strength through multiplication, and is widely used in attention mechanisms and recommendation systems. For example, the original preference weights are: "Ice and Snow Sports" 0.72, "Seaside Holiday" 0.45, and "Cultural Heritage" 0.60; after modulation, they become: "Ice and Snow Sports" 0.72 × 0.08 ≈ 0.06, "Seaside Holiday" 0.45 × 1.75 ≈ 0.79, and "Cultural Heritage" remains unchanged at 0.60 (because it is not time-sensitive).

[0108] Subsequently, the system normalizes the weights of all preference dimensions by scaling them proportionally to ensure the sum equals 1 (e.g., through Softmax or linear normalization), guaranteeing that they can be used as valid probability distributions or ranking criteria for input into the downstream model. Finally, the system inputs the dynamically calibrated preference dimensions and their adjusted weights (e.g., "Seaside Vacation" 0.52, "Cultural Heritage" 0.39, "Winter Sports" 0.04) into the information recommendation model. Based on this context-aware preference vector, the information recommendation model performs operations such as attraction retrieval, itinerary combination, conflict detection, and personalized ranking. Experimental data shows that after introducing this timeliness calibration mechanism, the system's recommendation rationality score improved by 42% in cross-seasonal and cross-regional scenarios, and the user complaint rate for "inappropriate recommendations" (e.g., recommending skiing in summer or mountain climbing in the rainy season) decreased by 68%. In summary, this embodiment, by integrating current spatiotemporal context information with a structured time-sensitive knowledge base, achieves intelligent, dynamic, and interpretable adjustment of user preference weights, effectively bridging the gap between historical behavior and current constraints, and significantly enhancing the practicality and user experience of the itinerary planning system.

[0109] In this embodiment of the application, after obtaining the target demand preference information, refer to step 206 to input the target demand preference information into the information recommendation model, perform matching and integration of itinerary information, and obtain the target itinerary list.

[0110] To dynamically balance the relevance and accuracy of recommended travel information, this embodiment innovatively introduces a degradation mechanism and an upgrade mechanism. The degradation mechanism addresses the issue of insufficient recommended travel information by gradually relaxing constraints according to preference weights from low to high (i.e., discarding intersection conditions of low-weight dimensions), expanding the candidate range while retaining core preferences, and ensuring the number of recommended travel information results reaches a usable level. The upgrade mechanism addresses the issue of excessive recommended travel information by gradually tightening constraints according to preference weights from high to low (i.e., increasing the similarity threshold of high-weight dimensions to strengthen deduplication), prioritizing the filtering of redundant items weakly related to core preferences, and improving the diversity and discriminativeness of recommended travel information results. The preference weight reflects the user's emphasis on that preference dimension (determined by historical behavior or explicit feedback).

[0111] Based on this, the target demand and preference information is input into the information recommendation model to match and integrate the itinerary information, resulting in a target itinerary list. This includes: inputting the target demand and preference information into the information recommendation model and performing the following operations within the model: A1. Analyze the target demand preference information to obtain multiple preference dimensions and preference description information on each preference dimension; A2. For any preference dimension, select the itinerary information that matches the preference dimension and its preference description information from a diversified itinerary information database to obtain the initial itinerary list corresponding to the preference dimension. A3. Based on the similarity threshold corresponding to the preference dimension, perform a deduplication operation on the trip information in the initial trip list to obtain the deduplicated trip list corresponding to the preference dimension. A4. Based on the intersection constraints corresponding to each preference dimension, perform an intersection operation on the deduplicated itinerary lists corresponding to each preference dimension to obtain the candidate itinerary list. A5. If the number of trips in the candidate itinerary list is less than the second threshold, the downgrade mechanism is activated. The intersection constraints of the corresponding preference dimensions are discarded in order of preference weight from low to high, until the number of trips in the candidate itinerary list is greater than or equal to the second threshold, so as to obtain the target itinerary list. A6. If the number of trips in the candidate trip list is greater than the third threshold, the upgrade mechanism is activated. The similarity threshold of the corresponding preference dimension is increased sequentially from high to low according to the preference weight of each preference dimension until the number of trips in the candidate trip list is less than or equal to the second threshold, so as to obtain the target trip list; wherein, the third threshold is greater than or equal to the second threshold.

[0112] In step A1 above, the target preference description information refers to the final semantic or structured requirement representation that the system determines for the user on each specific preference dimension after completing multimodal requirement understanding, context supplementation and preference analysis, and is used for trip information matching and deduplication decision-making. For example: When the preference dimension is geographical location, the corresponding target preference description information could be "a certain business district, no more than 2 kilometers away from the current location" (Note: This can include specific administrative districts, landmark areas, and distance constraints for spatial proximity judgment and deduplication); when the preference dimension is service type, the corresponding target preference description information could be "an artsy coffee shop that provides hand-drip coffee and independent reading space" (Note: This can clearly define business attributes and core features to identify service homogenization and avoid recommending places with similar styles); when the preference dimension is price range, the corresponding target preference description information could be "average consumption between 80 and 120 yuan per person" (Note: Defining budget sensitivity in the form of a numerical range supports price overlap calculation and deduplication threshold application); when the preference dimension is user evaluation, the corresponding target preference description information could be "high-frequency keywords are 'quiet,' 'high photo quality,' and 'attentive service,' with an overall score of no less than 4.7" (Note: This integrates word-of-mouth tags and quantitative indicators to assess experience similarity and prevent recommending multiple "similar" itinerary information).

[0113] In step A2 above, the diversified itinerary information database refers to a structured, multi-source integrated comprehensive itinerary information database that covers a variety of travel elements. Its core objective is to support multi-dimensional, high-precision, and personalized itinerary information recommendations.

[0114] In step A4 above, the intersection constraint refers to a set of retention rules bound to each preference dimension. These rules determine whether the trip information simultaneously meets the user's needs across multiple dimensions, thus qualifying it for intersection calculation. Specifically, an item must match the target preference description information of the corresponding preference dimension in terms of its category, attribute value (such as price range), and tag (such as "family-friendly"), and must have passed the deduplication screening for that preference dimension before it can be included in the intersection calculation. By default, constraints on all preference dimensions are enabled, forming a strict "AND" logical intersection. This mechanism ensures that every piece of trip information in the candidate trip list meets the user's multi-dimensional core needs, while also providing an operational basis for the subsequent degradation mechanism. For example, when there are too few recommended trip information results, the intersection constraints of some preference dimensions can be gradually discarded according to preference weight from low to high, dynamically relaxing the screening conditions. This achieves an adaptive balance between the quantity and quality of trip information recommendation results while ensuring that the core preference dimensions are not weakened.

[0115] Based on this, in an optional embodiment, according to the intersection constraints corresponding to each preference dimension, the deduplicated itinerary lists corresponding to each preference dimension are subjected to intersection operation to obtain a candidate itinerary list, including: obtaining the intersection constraints corresponding to each preference dimension, the intersection constraints being used to determine whether the itinerary information of the preference dimension satisfies the target preference description information of the preference dimension; for any itinerary information, if it belongs to at least one deduplicated itinerary list, then sequentially verifying whether it satisfies the intersection constraints corresponding to all other preference dimensions; retaining the itinerary information that simultaneously satisfies the intersection constraints of all enabled preference dimensions to obtain a candidate itinerary list.

[0116] In step A5 above, the travel assistant or user may have requirements regarding the number of trips in the recommended itinerary list. If the number of trips in the itinerary list is less than a second threshold, then the second threshold is the lower limit for the itinerary information included in the itinerary list. Based on this, if the number of trips in the candidate itinerary list is less than the second threshold, a degradation mechanism is activated. The intersection constraints of the corresponding preference dimensions are discarded sequentially from low to high according to the preference weights of each preference dimension, until the number of trips in the candidate itinerary list is greater than or equal to the second threshold, thus obtaining the target itinerary list. Here, discarding the intersection constraints of the corresponding preference dimensions means that in the multi-dimensional intersection operation, the candidate itinerary information is no longer required to satisfy the target preference description information of that preference dimension. That is, the filtering restriction of that preference dimension on the itinerary information is removed, changing it from the itinerary information having to satisfy the target preference information of that preference dimension to the target preference information where that preference dimension can be ignored.

[0117] In one optional embodiment, if the number of trips in the candidate trip list is less than a second threshold, a downgrade mechanism is initiated. The intersection constraints of each preference dimension are discarded sequentially from low to high according to their preference weights, until the number of trips in the candidate trip list is greater than or equal to the second threshold, thus obtaining the target trip list. This includes: when the number of trips in the candidate trip list is detected to be less than the second threshold, obtaining the preference weights corresponding to each preference dimension; sorting each preference dimension according to its preference weight from low to high to generate a constraint discard sequence; sequentially discarding the intersection constraints of the corresponding preference dimensions according to the constraint discard sequence, and re-performing the intersection operation after each discard to obtain an updated candidate trip list; when the number of trips in the updated candidate trip list is greater than or equal to the second threshold, the updated candidate trip list is determined as the target trip list. This degradation mechanism, when the number of candidate trips is insufficient, discards the intersection constraints of secondary dimensions in order of preference weight from low to high. This achieves intelligent and orderly expansion of the recommendation result set while ensuring that the user's core needs (high-weight preferences) are not weakened. It effectively avoids the problem of sparse or empty results caused by strict intersection of multiple dimensions, thereby improving the system's usability, robustness and user experience while maintaining high relevance.

[0118] In step A6 above, the travel assistant or user may have requirements regarding the number of trips in the recommended itinerary list. For example, if the number of trips in the itinerary list exceeds a third threshold, then the third threshold is the upper limit for the itinerary information included in the itinerary list. Based on this, if the number of trips in the candidate itinerary list exceeds the third threshold, an upgrade mechanism is activated. The similarity threshold for each preference dimension is increased sequentially from high to low according to the preference weight of each preference dimension, until the number of trips in the candidate itinerary list is less than or equal to the second threshold, thus obtaining the target itinerary list; wherein, the third threshold is greater than or equal to the second threshold.

[0119] In one optional embodiment, if the number of trips in the candidate trip list is greater than a third threshold, an upgrade mechanism is initiated. The similarity thresholds of the corresponding preference dimensions are increased sequentially from high to low according to their preference weights, until the number of trips in the candidate trip list is less than or equal to a second threshold, thus obtaining the target trip list. This includes: when the number of trips in the candidate trip list is detected to be greater than the third threshold, obtaining the preference weights corresponding to each preference dimension; sorting each preference dimension according to its preference weight from high to low to generate a threshold adjustment sequence; sequentially increasing the similarity thresholds of the corresponding preference dimensions according to the threshold adjustment sequence, and after each increase, re-performing the deduplication operation on the initial trip list corresponding to the preference dimension based on the updated similarity threshold, thereby updating the candidate trip list; when the number of trips in the updated candidate trip list is equal to the second threshold, the updated candidate trip list is determined as the target trip list. This upgrade mechanism, when there are too many candidate trips, increases the similarity threshold of high-value dimensions in descending order of preference weight. This achieves intelligent and accurate compression of homogeneous and redundant information while preserving the expression of users' core preferences. It effectively avoids recommendation overload and decision fatigue caused by lenient deduplication, thereby improving the distinctiveness, information density, and user experience quality of recommendations while maintaining the diversity of results.

[0120] The embodiments related to step 206 described above transform structured target demand and preference information into a moderately numerous, highly relevant, and diverse itinerary information recommendation result through multi-dimensional analysis, independent deduplication, intersection fusion, and a dynamic upgrade / downgrade mechanism based on preference weights. This not only achieves a closed loop from "deep demand understanding" to "precise recommendation execution" but also effectively solves the major problems of insufficient or overloaded recommendations. Furthermore, a similarity threshold controls the deduplication granularity (the higher the threshold, the stricter the deduplication), while the second and third thresholds define the ideal lower and upper limits of the recommendation list, respectively. This mechanism design ensures that the system can adaptively output a moderately numerous and quality-controlled itinerary list when facing different input complexities, balancing user experience and recommendation efficiency. Ultimately, it improves the relevance of itinerary recommendations to users' actual needs and provides users with more accurate, intelligent, and coherent itinerary planning services.

[0121] Furthermore, to address the issue of users providing vague or sparse information in the early stages of trip planning (e.g., only entering "just two days to play" or "take the kids out for a trip"), an implicit preference injection mechanism is introduced before parsing the target demand preference information to enhance the targeting and accuracy of recommendations. This mechanism first detects whether the user's explicit preference information is sparse, or whether the completeness of the demand information is below a preset completeness threshold. If so, it extracts trip attribute tags from the user's historical trip order data and performs unsupervised clustering on these tags to generate one or more implicit preference keywords. These one or more implicit preference keywords serve as supplementary preference dimensions, which, together with the explicit preference dimensions obtained through parsing the target demand preference information, constitute a complete set of preference dimensions. This allows the information recommendation model to generate a corresponding initial trip list for each preference dimension based on the set of preference dimensions.

[0122] Specifically, the mechanism first detects whether the explicit preference information currently provided by the user is sparse. Explicit preference information refers to the structured preference expression generated by the multimodal fusion coding module (as claimed in claim 6) based on the user's active input (including text, speech-to-text, image recognition results, etc.) in the current round of dialogue. Its specific form is a set containing multiple preference dimensions and their corresponding weights (e.g., {"Seaside Vacation": 0.6, "Cultural Heritage": 0.4}). Preference dimensions refer to itinerary interest category tags with clear semantics, such as "Family Fun," "Summer Vacation," and "High-End Dining," which are extracted and standardized from multimodal content. Sparsity means that the explicit preference information is significantly insufficient in semantic coverage or information density, failing to support effective personalized recommendations. The system determines its sparsity using any of the following methods: Dimension count determination method: count the number of preference dimensions contained in the explicit preference information; if the number is less than the preset dimension threshold (e.g., 2), it is determined to be sparse; Demand information completeness determination method: The "demand information completeness" value calculated in the previous step (as described in the relevant process of claim 3) is used. This indicator comprehensively reflects the semantic richness of multimodal input. Its calculation method includes, but is not limited to, keyword coverage, modality diversity score, intent clarity score, etc., and the value range is [0,1]. If the value is lower than the preset completeness threshold (e.g., 0.4), it is also determined to be sparse.

[0123] If any of the above conditions are met, it is considered that the explicit preference information is sparse, triggering the implicit preference mining process. When it is determined to be sparse, high semantic value trip attribute tags are automatically extracted from the user's authorized historical trip order data (i.e., the structured trip booking records completed by the user through this platform in the past period (such as the most recent 12 months), each record containing fields such as destination, activity type, accommodation level, mode of transportation, and companion composition). Examples include "Disneyland visit", "Seaside B&B stay", "Museum visit", and "Family restaurant dining".

[0124] Subsequently, these tags are vectorized (using BERT semantic embedding or TF-IDF weighted encoding), and unsupervised clustering (a machine learning method that automatically discovers potential interest patterns from behavioral data without manual annotation, such as DBSCAN or K-means algorithms) is applied to group the sample points in the vector space. Each cluster represents a stable long-term interest, and its representative semantics are extracted from high-frequency co-occurring words within the cluster or the cluster center into one or more implicit preference keywords (i.e., interest tags that are not explicitly expressed in the current dialogue but reflect the user's true preferences, such as "theme park enthusiast" or "light and healthy diet preference"). The generated implicit preference keywords are used as supplementary preference dimensions, which, together with the explicit preference dimensions parsed from the current session, constitute a complete set of preference dimensions (this set is a temporary data structure built internally by the information recommendation model, used only in the trip recall stage of this recommendation session, and does not modify the original multimodal understanding results). For example, if the explicit dimensions are "summer retreat" and "short trip," while the implicit dimensions are "family-friendly theme parks" and "light meals," the information recommendation model will generate initial itinerary lists for each of these four dimensions. Then, based on subsequent deduplication, intersection, and sorting operations, it will ultimately output itinerary plans that better match the user's actual needs. This mechanism effectively solves the problem of recommendation generalization or deviation from intent caused by sparse user input. Since implicit preferences originate from long-term user behavior and have highly personalized characteristics, and are implemented through unsupervised clustering, no manual annotation is required, resulting in good scalability. Online experiments show that in low-completeness conversation scenarios (accounting for approximately 35% of total traffic), this method increases the recommendation click-through rate by 41%, significantly improving the user experience, while ensuring that the system can still provide high-quality, context-relevant itinerary suggestions even under conditions of insufficient information.

[0125] Furthermore, after obtaining the target project list, to further enhance the richness and diversity of the recommendation results, a diversity re-ranking mechanism based on the novelty of preference dimension coverage is introduced. This mechanism aims to avoid the homogenization problem caused by traditional ranking methods (such as relying solely on matching scores, popularity, or click-through rates), such as continuously recommending multiple hotels or attractions with highly similar styles, locations, or themes, leading to user fatigue or missing potentially high-quality options. Technically, for any trip information in the target itinerary list, the diversity weight of that trip information can be calculated based on the preference weight of the corresponding preference dimension; the trip information is then re-ranked according to its diversity weight to obtain the re-ranked target itinerary list.

[0126] Specifically, the process begins by establishing a connection between each itinerary in the target itinerary list (such as a "Sanya 3-day family tour" itinerary) and the complete set of preference dimensions (including explicit preference dimensions and implicit preference keywords). This connection is achieved through semantic matching: the structured attributes contained in the itinerary information (such as attraction type, dining style, accommodation level, activity tags, etc.) are compared with each preference dimension. If the semantic relevance of an itinerary on a certain preference dimension reaches a preset threshold (e.g., determined by vector similarity), then the itinerary is considered to have a valid response on that dimension, and its response strength is recorded (e.g., high, medium, low, or represented by a value between 0 and 1).

[0127] Subsequently, a diversity weight is calculated for each trip. This weight reflects not only the degree to which the trip satisfies user preferences but also the novelty of the dimension coverage it brings to the current list. Specifically, it dynamically tracks how many of the trips recommended before the current trip have already fully covered the same preference dimension. If a preference dimension (such as "beach vacation") has been satisfied multiple times in the preceding trips, the diversity contribution of subsequent trips that also emphasize that dimension will be reduced accordingly; conversely, if a trip can activate a preference dimension that has not yet been fully covered (such as the list being dominated by "cultural relics" while this trip emphasizes "light meals"), its diversity weight will be increased. To achieve this effect, a "saturation recommendation number" can be preset for each preference dimension (e.g., a maximum of 3 highly relevant trips can be recommended for "family activities"). When the number of recommended trips for a dimension approaches or reaches this limit, the contribution of new trips to that dimension will be reduced; while low-frequency but relevant dimensions will be given higher incentives.

[0128] Ultimately, the diversity weight of each trip is determined by its response strength across each preference dimension, the original preference weight of the corresponding dimension, and the current coverage saturation of that dimension. After calculating the diversity weight, the system integrates it with the trip's original matching score (such as multimodal preference matching degree, historical conversion rate, service rating, etc.) to generate the final ranking criteria. During the integration, a weighted approach can be used to balance "relevance" and "diversity." For example, by default, more emphasis is placed on matching score (e.g., 70% weight), while diversity is introduced as an adjustment factor (e.g., 30% weight) to ensure that core needs are prioritized without sacrificing the breadth of the experience.

[0129] Finally, the target itinerary list is reordered based on the integrated score, and the reordered target itinerary list is output. This list, while maintaining a high degree of consistency with the user's core preferences, significantly enhances the content's hierarchy and exploration value. For example, when the user's preferences include "cultural relics" and "light dining," not only will highly matching museum itineraries be recommended, but also ancient city tours incorporating distinctive vegetarian restaurants will be interspersed, avoiding a concentration on a single type.

[0130] This solution represents an evolution from "single optimal match" to "multiple high-quality combinations." By reordering, it avoids the homogeneous stacking caused by sorting solely by matching degree or popularity, ensuring that the final list not only closely meets the core needs of users but also has rich experience layers, effectively improving user decision-making efficiency and satisfaction.

[0131] Furthermore, to facilitate the management of user travel information, a travel calendar can be provided to users, which they can associate with. The travel calendar stores historical or planned travel information for multiple days. Based on this, before obtaining the user's initial multimodal needs information in this round of dialogue, the process includes: responding to an access request to the travel management interface, displaying the travel management interface, which includes multiple travel dates and a travel list for each travel date; responding to a selection operation for any travel date, obtaining the travel information for that travel date and the travel information for adjacent dates; responding to a creation or modification operation for the travel information for that travel date, displaying a session interface, which is used to receive the user's input of multimodal initial needs information; and after obtaining the reordered target travel list, responding to the user's selection operation for target travel information in the target travel list, generating the travel information for that travel date based on the target travel information, any travel date, and the travel information for the dates adjacent to that travel date, and updating it to the travel calendar. This solution integrates a trip calendar, deeply binding multimodal dialogue recommendations with the user's actual schedule. Specifically, it obtains the selected date and adjacent days' trips as context before the dialogue, improving the accuracy of understanding the user's needs; during the dialogue, it receives user input based on time anchors, enhancing intent recognition; and after the recommendation, it automatically integrates the selected trip into the calendar, considering overall schedule coordination. This achieves a closed loop from "fuzzy request" to "executable schedule," improving the relevance and usability of individual recommendations while accumulating structured trip information, supporting long-term context reuse, and significantly enhancing user experience and system intelligence consistency.

[0132] In one optional embodiment, based on the target trip information, any trip date, and trip information of dates adjacent to the trip date, generating a trip list for any trip date includes: if the estimated travel time of the target trip information is less than a fourth threshold, obtaining at least one candidate trip information, wherein the at least one candidate trip information satisfies the following conditions: the at least one candidate trip information has a strong semantic correlation with the target trip information and the trip information of dates adjacent to any trip date, and the sum of the estimated travel time of the at least one candidate trip information and the estimated travel time of the target trip information does not exceed a fifth threshold; and generating trip information for any trip date based on the target trip information, at least one candidate trip information, and trip information of dates adjacent to any trip date.

[0133] At this point, the trip information processing method has been completed, achieving deep fusion of multimodal information and collaborative enhancement of multi-turn dialogue context within the same link. This significantly improves the relevance of trip recommendations to users' actual needs, providing users with more accurate, intelligent, and coherent trip planning services, and reshaping the cognitive and interaction paradigm of multimodal travel assistants.

[0134] Figure 3 This is a schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this application. For example... Figure 3 As shown, the electronic device includes a memory 30a and a processor 30b; wherein the memory 30a is used to store computer programs; the processor 30b is coupled to the memory 30a and is used to execute the computer programs to perform the relevant steps of the methods described above.

[0135] Furthermore, such as Figure 3 As shown, the server device also includes other components such as a communication component 30c, a power supply component 30d, an audio component 30e, and a display screen 30f. Figure 3 The diagram only shows some components and does not mean that the computer device includes only these components. Figure 3 The components shown.

[0136] An exemplary embodiment of this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above method, which will not be described in detail here.

[0137] An exemplary embodiment of this application also provides a computer program product, including: a computer program that, when executed by a processor of an electronic device, causes the processor to perform the steps in the above method, which will not be described in detail here.

[0138] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0140] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0141] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0142] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0143] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0144] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0145] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0146] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for processing travel information, characterized in that, include: Obtain information about the user's target needs and preferences; The target demand preference information is input into the information recommendation model, and the following operations are performed in the information recommendation model: The target demand preference information is parsed to obtain multiple preference dimensions and preference description information on each preference dimension; For any preference dimension, select the itinerary information that matches the preference dimension and its description information from a diversified itinerary information database to obtain the initial itinerary list corresponding to the preference dimension; Based on preset deduplication strength levels, combined with current semantic scene information and preference weights of each preference dimension, the target deduplication strength level corresponding to each preference dimension is determined. Each target deduplication level corresponding to each preference dimension has its own target similarity threshold. The current semantic scene information refers to the set of context features used to dynamically determine each preference dimension. Based on the target similarity threshold corresponding to each preference dimension, the initial itinerary list of the corresponding preference dimension is deduplicated to obtain the itinerary list after deduplication for each preference dimension. The intersection of the itinerary lists after deduplication of each preference dimension is performed to obtain the candidate itinerary list; If the number of trips in the candidate trip list is less than the first threshold, a degradation mechanism is initiated. The intersection constraints of the corresponding preference dimensions are discarded in order of preference weight from low to high until the number of trips in the candidate trip list is greater than or equal to the first threshold, so as to obtain the target trip list. The intersection constraint means that the candidate trips must meet the adaptation conditions of the corresponding preference dimensions.

2. The method according to claim 1, characterized in that, Each level of deduplication intensity has a corresponding baseline similarity threshold preset for each preference dimension; Based on preset deduplication strength levels, and combined with the current semantic scene information and the preference weights of each preference dimension, the target deduplication strength level for each preference dimension is determined, including: Obtain current semantic scene information; Based on the current semantic scene information, select one of the multiple deduplication strength levels as the initial deduplication strength level; Based on the preference weights of each preference dimension, the baseline similarity thresholds corresponding to each preference dimension in the initial deduplication intensity level are fine-tuned to obtain the target deduplication intensity level corresponding to each preference dimension. The target deduplication level of each preference dimension has its own target similarity threshold.

3. The method according to claim 2, characterized in that, Based on the preference weights of each preference dimension, the baseline similarity thresholds corresponding to each preference dimension in the initial deduplication intensity level are fine-tuned to obtain the target deduplication intensity level corresponding to each preference dimension, including: Obtain a preset benchmark weight that is adapted to the current semantic scene information. The preset benchmark weight is obtained based on historical data of historical users in the same or similar semantic scenes. Calculate the deviation between the preference weights of each preference dimension and the corresponding preset benchmark weights; Based on the sign and magnitude of the deviation corresponding to each preference dimension, the baseline similarity threshold of each preference dimension in the initial deduplication intensity level is adjusted as follows: If the deviation is positive, then increase the baseline similarity threshold for the corresponding preference dimension; If the deviation is negative, the baseline similarity threshold for the corresponding preference dimension is reduced. The adjusted baseline similarity thresholds corresponding to each preference dimension are used as target similarity thresholds to obtain the target deduplication intensity levels corresponding to each preference dimension.

4. The method according to claim 1, characterized in that, When the downgrade mechanism is activated, the intersection constraints of the corresponding preference dimensions are discarded sequentially from low to high according to the preference weights of each preference dimension, until the number of trips in the candidate trip list is greater than or equal to a first threshold, to obtain the target trip list, including: For the current preference dimension to be processed, its corresponding deduplicated itinerary list is excluded from the set of itinerary lists involved in the intersection operation, so that the candidate itinerary does not need to meet the adaptation conditions of the preference dimension at the same time. If the number of trips in the updated candidate trip list is still less than the first threshold, then continue processing the next preference dimension sorted in ascending order of preference weight until the number of trips in the candidate trip list is greater than or equal to the first threshold.

5. The method according to claim 1, characterized in that, When the upgrade mechanism is activated, additional intersection constraints for each preference dimension are added sequentially from high to low according to the preference weights of each preference dimension, until the number of trips in the candidate trip list is less than or equal to the second threshold, to obtain the target trip list, including: For the current preference dimension to be processed, based on a preset domain knowledge base or user profile, one or more refined sub-conditions are generated, and a subset of the journeys that only satisfy the refined sub-conditions is determined; In subsequent intersection operations, the aforementioned subset of trips is used to replace the original deduplicated trip list for this preference dimension; If the number of trips in the updated candidate trip list is still greater than the second threshold, then continue processing the next preference dimension sorted in descending order of preference weight until the number of trips in the candidate trip list is less than or equal to the second threshold.

6. The method according to claim 1, characterized in that, Also includes: If the number of trips in the candidate trip list is greater than the second threshold, an upgrade mechanism is activated. The additional intersection constraints of the corresponding preference dimensions are added sequentially from high to low according to the preference weights of each preference dimension, until the number of trips in the candidate trip list is less than or equal to the second threshold, so as to obtain the target trip list. The additional intersection constraints refer to the fact that the candidate trips also need to meet the refined adaptation conditions under the corresponding preference dimensions, and the second threshold is greater than or equal to the first threshold.

7. The method according to claim 1, characterized in that, After obtaining the list of candidate itineraries, it also includes: If the number of trips in the candidate trip list is less than the first threshold, a downgrade mechanism is initiated. The target similarity threshold of the corresponding preference dimension is lowered in order of preference weight from low to high according to the preference weight of each preference dimension. Based on the lowered target similarity threshold, the initial trip list of the preference dimension is re-executed for deduplication until the number of trips in the candidate trip list is greater than or equal to the first threshold. If the number of trips in the candidate trip list is greater than the second threshold, the upgrade mechanism is activated. The target similarity threshold of the corresponding preference dimension is increased in descending order of preference weight according to each preference dimension. Based on the increased target similarity threshold, the initial trip list of the preference dimension is re-executed to remove duplicates until the number of trips in the candidate trip list is less than or equal to the second threshold, so as to obtain the target trip list. Wherein, the second threshold is greater than or equal to the first threshold, and when the semantic similarity between two trips is greater than or equal to the currently used target similarity threshold, they are determined to be duplicate trips and a deduplication operation is performed.

8. The method according to claim 1, characterized in that, The conditions for triggering the downgrade mechanism also include any of the following: The diversity score of the candidate itinerary list is lower than the preset diversity threshold; The coverage of the high-weight preference dimension of the candidate itinerary list is lower than the preset coverage threshold; The conditions for initiating the upgrade mechanism also include any of the following: The proportion of semantically repetitive trip clusters in the candidate trip list exceeds a preset repetition ratio threshold; The novelty score of the candidate itinerary list is lower than the preset novelty threshold; The diversity score is a global diversity measure determined based on the semantic or attribute differences among all trips in the candidate trip list; the high-weight preference dimension coverage is the coverage ratio of the candidate trip list as a whole to the high-weight preference dimension set by the user; and the novelty score is a global measure used to measure the novelty of the candidate trip list, determined based on the overall differences between the candidate trip list and the preset mainstream trip set in semantic or feature space.

9. The method according to any one of claims 1-8, characterized in that, Obtain user's target needs and preferences information, including: During the multi-round dialogue with the user regarding this trip planning, the initial multimodal needs information of the user involved in this round of dialogue is obtained; The initial key information of each single-mode encoder is extracted by calling the single-mode encoding modules of the multi-mode encoder to obtain the initial key information set of each single-mode. The initial key information sets of each single modality are input into the first large model to perform demand information completeness analysis and initial trip demand category analysis to obtain demand information completeness and initial trip demand category; When the completeness of the demand information is less than a first completeness threshold, historical demand summaries and / or historical session summaries that match the initial trip demand category are recalled from the long short-term memory storage area, and supplementary demand information of the initial demand information is generated based on the historical demand summaries and / or historical session summaries to obtain supplementary multimodal target demand information; the multimodal target demand information has a target trip demand category; The multimodal fusion encoding module of the multimodal encoder is invoked to integrate and analyze the target demand information of each single modality, taking the target travel demand category as a reference, so as to obtain the user's target demand preference information.

10. The method according to claim 9, characterized in that, The short and long memory storage area includes a short-term memory storage area and a long-term memory storage area. The short-term memory storage area stores conversation summaries generated at a preset collection frequency, when the topic category remains unchanged and the preset number of rounds is reached, as well as demand summaries and conversation summaries generated when the topic category changes but the preset number of rounds is not reached. The long-term memory storage area stores demand summaries and conversation summaries that sink from the short-term memory storage area after reaching a preset storage duration. Retrieve historical demand summaries from the long and short memory storage area that match the initial travel demand category, including: Based on the initial travel demand category, historical demand summaries are retrieved from the short-term memory storage area as the first supplementary demand information to the initial demand information of the multimodal system; If a historical demand summary is not retrieved from the short-term memory storage area, or if a historical demand summary is retrieved from the short-term memory storage area but the demand completeness of the initial demand information of the multimodal mode supplemented based on the retrieved historical demand summary is less than the second completeness threshold, then based on the initial trip demand category, a historical demand summary is retrieved from the long-term memory storage area as the first supplementary demand information.

11. The method according to claim 10, characterized in that, Retrieve historical session summaries from the long and short memory storage area that match the initial trip requirement category, including: Based on the initial trip demand category, retrieve the historical session summary from the short-term memory storage area; If the historical session summary is not retrieved from the short-term memory storage area, the historical session summary shall continue to be retrieved from the long-term memory storage area based on the initial trip demand category; If a historical conversation summary is retrieved from the short-term memory storage area and / or the long-term memory storage area, the historical conversation summary is used as the context information of the current round of dialogue. Based on the context information, a guiding dialogue is generated using the second major model, and a question-and-answer service is provided to the user based on the guiding dialogue to obtain the second supplementary demand information provided by the user based on the question-and-answer service.

12. The method according to claim 9, characterized in that, The multimodal fusion encoding module of the multimodal encoder is invoked to integrate and analyze the target demand information of each single modality, taking the target travel demand category as a reference, to obtain the user's target demand preference information, including: The multimodal fusion encoding module of the multimodal encoder is invoked. Taking the target travel requirement category as a reference, cross-attention calculation is performed on the target requirement information of each single modality. During the cross-attention calculation process, the global importance weight of each target key information in the target requirement information of each single modality relative to the target travel requirement category is assigned. By combining the global importance weights of each key information in the target requirements of each single modality and the first context constraint condition representing the local context of the current dialogue, the contribution of each key information in the target requirements of each single modality is dynamically adjusted to generate a context-aware weighted key information representation corresponding to each key information. Using the target travel demand category as the fusion condition, cross-modal fusion is performed on the key information representations of each target to obtain the fused key information representation; Based on the fused key information representation and the second contextual constraints characterizing the user's global profile or external environment, a multi-dimensional demand preference analysis is performed to obtain target demand preference information.

13. An electronic device, characterized in that, include: Memory and processor; The memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program to implement the steps of the method according to any one of claims 1-12.

14. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor of the electronic device, it causes the processor to perform the steps of the method as described in any one of claims 1-12.

15. A computer program product, characterized in that, include: A computer program, when executed by a processor of an electronic device, causes the processor to perform the steps of the method as described in any one of claims 1-12.

Images