An ai-based user intent recognition route planning method and system

Abstract

The application relates to the technical field of artificial intelligence, in particular to an AI-based user intention recognition route planning method and system, which comprises the following steps: receiving route demand information and extracting a trip constraint set; constructing a constraint item queue based on the trip constraint set and traversing, when there is a missing constraint value, generating corresponding supplementary inquiry text and sending the text to a user terminal; receiving supplementary information, extracting a corresponding constraint value and filling the constraint value into the constraint item queue; when a route generation condition is met, executing candidate point recall; generating multiple candidate routes based on the recall result; and in response to an editing operation of the user terminal, updating a route direction of the corresponding candidate route and a corresponding time axis in real time and outputting. The application can effectively solve the problem of inconsistency in time and space dimensions caused by the interruption of the interactive process and the static solidification of the route in the prior art, and has the technical effects of improving the interactive efficiency of route planning and dynamic adaptability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an AI-based user intent recognition route planning method and system. Background Technology

[0002] With the integration of artificial intelligence and geographic information systems, route planning methods based on natural language interaction have been widely applied. Existing planning systems typically parse user intent in a single interaction and then call upon a geographic database to generate a route.

[0003] However, existing systems suffer from the following technical shortcomings when processing real-world unstructured interactive data: First, they struggle to ensure the integrity of input parameters when parsing multimodal or fuzzy route requirements. Specifically, when initial user input lacks constraints, existing systems lack dynamic validation and pre-completion mechanisms based on data structures. They either directly trigger geographic data source queries using incomplete data, resulting in a large amount of irrelevant spatial data, or they simply crash, failing to form an effective data flow loop. Second, the route data output by existing systems is typically statically bound. Specifically, after route planning is complete, if local nodes change, such as when a user initiates an edit operation on a candidate route, the system usually needs to re-enter all nodes into the path planning algorithm for global recalculation. This approach not only lacks dynamic response capabilities to multidimensional constraints but also struggles to locally and adaptively update the physical orientation and timestamp associations of subsequent routes, making it difficult to maintain spatiotemporal consistency in the generated routes. Summary of the Invention

[0004] In view of the aforementioned problems, this application is proposed to provide an AI-based user intent recognition route planning method and system that overcomes or at least partially solves the aforementioned problems, comprising: The above-mentioned objective of this application is achieved through the following technical solution: An AI-based user intent recognition route planning method includes the following steps: Receive route requirement information input from the user and extract a set of travel constraints based on the route requirement information; A constraint item queue is constructed based on the travel constraint set, and the constraint item queue is traversed according to a preset order. When a constraint value is missing during the traversal, a corresponding supplementary query text is generated and sent to the user. Receive supplementary information from the user based on the supplementary query text input, extract the corresponding constraint values ​​based on the supplementary information and populate them into the constraint item queue; When it is determined that the travel constraint set meets the route generation conditions, the corresponding geographic data source is called to perform candidate point recall based on the travel constraint set. Based on the candidate point recall results, multiple candidate routes are generated after filtering and sorting by the travel constraint set. In response to the user's editing operation on any candidate route, the system updates the route direction and timeline of the corresponding candidate route in real time, and outputs the updated candidate route as the planned route.

[0005] The second objective of this invention is achieved through the following technical solution: An AI-based user intent recognition route planning system includes: The constraint extraction module is used to receive route requirement information input by the user and extract a set of travel constraints based on the route requirement information. The queue construction module is used to build a constraint item queue based on the travel constraint set, and traverse the constraint item queue according to a preset order. When a constraint value is missing during the traversal, a corresponding supplementary query text is generated and sent to the user terminal. The constraint supplementation module is used to receive supplementary information from the user based on the supplementary query text input, extract the corresponding constraint values ​​based on the supplementary information, and populate them into the constraint item queue. The recall execution module is used to recall candidate points by calling the corresponding geographic data source based on the travel constraint set when it is determined that the travel constraint set meets the route generation conditions. The route generation module is used to generate multiple candidate routes by filtering and sorting the candidate point recall results in conjunction with the travel constraint set. The route update module is used to respond to the user's editing operation on any candidate route, update the route direction and corresponding timeline of the corresponding candidate route in real time, and output the updated candidate route as the planned route.

[0006] This application also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described AI-based user intent recognition route planning method.

[0007] This application also relates to a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described AI-based user intent recognition route planning method.

[0008] This application has the following advantages: In the embodiments of this application, an AI-based user intent recognition route planning method and system are provided. By dynamically constructing a constraint item queue, the system can intelligently complete missing constraints and support real-time updates of the route direction of the corresponding candidate route and the corresponding timeline. This can effectively solve the problem of inconsistent spatiotemporal dimensions caused by interruption of the interaction process and static fixation of the route in the prior art, and has the technical effect of improving the efficiency of route planning interaction and dynamic adaptability. Attached Figure Description

[0009] Figure 1 This is a flowchart of an embodiment of an AI-based user intent recognition route planning method according to this application; Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0010] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0011] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0012] All terms used in this application, including technical or scientific terms, have the same meaning as understood by one of ordinary skill in the art to which this application pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries, such as those in common dictionaries, should be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and not as having an idealized or highly formalized meaning, unless expressly defined herein.

[0013] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment shall be considered part of the specification.

[0014] First, some nouns or terms that appear in the description of the embodiments of this application shall be interpreted as follows: Route request information refers to raw data about a user's travel plans that they express to the system through various means, such as text, voice, and links. Route request information may include specific locations and times, or it may include vague descriptions or links.

[0015] Trip constraint set: refers to a series of restrictions extracted from route demand information to guide route planning. These may include spatial constraints (such as origin, destination, and waypoints), duration constraints (such as total duration and stay duration), scenario constraints (such as activity type and point of interest preference), and mode of travel constraints (such as walking, driving, and public transportation).

[0016] Constraint queue: refers to the sequence of constraints in the travel constraint set arranged according to a preset priority or processing order. Checking and processing constraints based on the order of this constraint queue can ensure that all necessary constraint values ​​are obtained.

[0017] Supplementary query text: This refers to text that is automatically generated and sent to the user when certain key constraint values ​​are missing or unclear during the processing of user-input route requirement information. It is used to guide the user to provide further information.

[0018] Geographic data sources refer to databases or service interfaces that store various geographic-related data such as geographic location information, road network data, point-of-interest data, and traffic condition data. Typically, route planning systems obtain the underlying data needed to generate routes by calling geographic data sources.

[0019] Candidate point recall refers to the process of retrieving and obtaining a series of qualified geographical locations, such as points of interest and landmarks, from geographic data sources based on preliminary travel constraints, such as spatial range and scene type, as potential nodes for route construction.

[0020] Candidate routes: These are multiple alternative route options generated by a path planning algorithm based on travel constraints and recalled candidate points. These candidate routes typically differ in terms of path, time, and waypoints, and are available for users to choose from.

[0021] Editing operations refer to the actions of users modifying the candidate routes generated by the system. For example, users may want to adjust waypoints, change the route direction, or modify the dwell time at a certain node.

[0022] Route orientation: refers to the specific path of a route in geographical space, that is, the sequence of roads or trajectories connecting various points along the way.

[0023] Timeline: refers to the time information corresponding to the route, including the estimated departure time, arrival time at each point along the way, stop time, and final arrival time at the destination.

[0024] In one embodiment, such as Figure 1 As shown, this application discloses an AI-based user intent recognition route planning method, which specifically includes the following steps: S10: Receive route requirement information input by the user and extract a set of travel constraints based on the route requirement information; In one embodiment of this application, the system can provide a structured form interface for users to input information such as origin, destination, desired duration, and mode of transportation. The system then parses the user's structured input into a set of travel constraints. For example, if a user inputs "Location A" in the "Origin" field and "3 hours" in the "Duration" field, the system can directly extract the spatial and duration constraints. Alternatively, the system can receive unstructured text input by the user, such as "I want to go from Location A to Location B, which will take about half a day." In this case, the system can use rule-based text matching technology to identify keywords and phrases in the text and map them to predefined constraint types and values. For example, "Location A" and "Location B" are identified as spatial constraints, and "half a day" is identified as a duration constraint.

[0025] S20: Construct a constraint item queue based on the travel constraint set, and traverse the constraint item queue according to a preset order. When a constraint value is missing during the traversal, generate a corresponding supplementary query text and send it to the user terminal. In one embodiment of this application, the system can pre-define a fixed order for processing constraints, such as processing spatial constraints first, then duration constraints, and finally scene constraints and travel mode constraints. During the traversal, the system checks whether each constraint item contains a valid constraint value. If a constraint value is found to be empty, the system can select a corresponding query template from a pre-defined text template library based on the type of the missing constraint item, fill in the existing constraint information, and generate supplementary query text. For example, if the "travel mode" constraint is missing, the system may generate the query "Which travel mode would you like to use?".

[0026] S30: Receive supplementary information from the user terminal based on the supplementary query text input, extract the corresponding constraint values ​​based on the supplementary information and fill them into the constraint item queue; In one embodiment of this application, as an implementation, when a user receives a supplementary query text and enters supplementary information, the system receives this supplementary information. For example, the user replies "I want to walk." At this time, the system can use keyword matching or regular expressions to identify the keyword "walking" from the user's reply and fill it into the constraint queue as the value of the "mode of travel" constraint.

[0027] S40: When it is determined that the travel constraint set meets the route generation conditions, the corresponding geographic data source is called to perform candidate point recall based on the travel constraint set; In one embodiment of this application, the system can set a basic set of route generation conditions. For example, if both spatial and time constraints are met, the route generation conditions are considered satisfied. Once the conditions are met, the system will send a query request to a general geographic information database based on existing travel constraints, such as geographic location information in spatial constraints. It will then search within a radius centered on the location corresponding to this geographic location information, recalling all points of interest within this radius as candidate points.

[0028] S50: Based on the candidate point recall results, multiple candidate routes are generated after filtering and sorting by the travel constraint set; In an embodiment of this application, as one implementation, after recalling candidate points, the system can perform preliminary screening of these candidate points, for example, removing points of interest in commercial areas that are completely inconsistent with the user-specified scenario. At this time, the system can sort the candidate points according to their distance from the user-specified location. After obtaining the set of candidate points after screening and sorting, the system can use a path planning algorithm, for example, first calculating the underlying distance matrix between each node using Dijkstra's algorithm or A* algorithm, and then combining it with a traveling salesman algorithm or dynamic programming algorithm, using the user-specified starting point and ending point as a basis, to sequentially combine these candidate points as waypoints to generate multiple different route schemes.

[0029] Furthermore, the specific implementation logic of the above algorithm is as follows: First, the starting point, ending point, and all candidate waypoints are taken as the target node set. The Dijkstra algorithm or A* algorithm is called to iterate through the underlying road network data, calculating the optimal travel cost between each target node, such as the shortest distance or time, thereby constructing a multi-dimensional underlying distance matrix. Then, this underlying distance matrix is ​​used as input parameters into the Traveling Salesman Problem model or dynamic programming algorithm to perform global sequence combination and optimization calculations on the visiting order of multiple candidate waypoints. Finally, the top N non-repeating sequence combinations whose total travel cost meets the preset requirements are output, forming multiple different route schemes, with preset requirements such as minimum cost.

[0030] S60: In response to the user's editing operation on any candidate route, it updates the route direction and corresponding timeline of the corresponding candidate route in real time, and outputs the updated candidate route as the planned route.

[0031] In one embodiment of this application, when a user edits a generated candidate route—for example, by manually adding or deleting a waypoint—the system recognizes this editing operation and can locally cascade updates the route path and timeline from the edit point to the destination based on a pre-built spatiotemporal dependent directed acyclic graph. For instance, if the user adds a new waypoint in the middle of the route, the system only needs to locally recalculate the path and time from the previous waypoint to the new waypoint, and from the new waypoint to the next waypoint, and then cascade the resulting time offset along the directed acyclic graph to subsequent nodes to update their timestamps, without requiring a global route reconstruction.

[0032] Specifically, a spatiotemporally dependent directed acyclic graph (DAG) is a data structure that maps the physical characteristics of a route to the underlying logic of a computer. It uses geographical locations along the route as graph nodes, such as start points, waypoints, and destinations, and the actual travel trajectories between nodes as directed edges. Each graph node is bound to an estimated arrival timestamp and estimated dwell time, and each directed edge is assigned weight parameters including spatial distance and estimated travel time. Based on this DAG structure, the timeline of the route increases unidirectionally with the advancement of spatial nodes, allowing modifications to local temporal or spatial variables to propagate directly backward along the topology without needing to re-invoke the underlying pathfinding engine.

[0033] For example, as a specific implementation, suppose user A plans a trip and enters route requirements on the user's device: "I want to go to a place near location A for about half a day, and it should be suitable for family activities." First, the system receives route requirement information input by user A and extracts a set of travel constraints based on this information. Through text parsing, the system identifies "location A" as the central point of the spatial constraint and "half a day" as the duration constraint. However, the system finds that constraints such as "mode of travel" and specific "scenario type" are missing.

[0034] Furthermore, the system constructs a constraint queue based on the extracted set of travel constraints. This constraint queue may prioritize missing constraints such as "mode of travel" and "scenario type" according to a preset priority. Based on this, the system traverses the constraint queue, detects missing "mode of travel" constraint values, and then generates supplementary query text based on the type of missing constraint, such as: "What mode of travel would you like to use? For example, walking, driving, or public transportation?" and sends it to the user's terminal.

[0035] After receiving the supplementary query text, User A inputs the additional information: "We want to walk, mainly to the park and the children's playground." The system receives this supplementary information and extracts the corresponding constraint values. At this point, the system identifies "walking" as the travel mode constraint value, and "park" and "children's playground" as scenario constraint values, and fills them into the constraint item queue. Based on this, the travel constraint set now includes spatial constraints, duration constraints, travel mode constraints, and scenario constraints.

[0036] Based on this, the system determines that the travel constraint set meets the route generation conditions, such as having spatial, duration, and travel mode constraints. According to this travel constraint set, the system calls the corresponding geographic data source to perform candidate point recall. Using "Location A" as the center point, and combining the duration of "half a day" and the travel mode of "walking," a suitable recall radius is calculated. Simultaneously, based on the scenario constraints of "park" and "children's playground," the system selects the corresponding geographic data source for querying, such as a POI database containing information on parks and children's facilities. The system sends a recall request to this geographic data source to obtain geographical locations within the recall radius that conform to the park and children's playground types, along with their attribute information, forming a candidate point set.

[0037] After obtaining the candidate point recall results, the system filters and sorts the candidate points based on the travel constraint set. For example, the system will prioritize filtering parks and children's playgrounds with higher ratings and better user reviews, and then rate and sort the filtered candidate points in descending order, selecting several as route nodes. At this point, based on these route nodes, the system executes a preset path planning algorithm to generate multiple candidate routes that meet the requirements of "walking" travel mode and "half-day" duration. For example, one route may include "Park 1 near location A → Children's playground 2 → Park 3 near location A", and another route may include "Children's playground 1 near location A → Park 2".

[0038] Finally, User A views the generated candidate routes through the user client. If User A finds that one of the candidate routes is missing a desired destination, "Location B," they can edit the route, for example, by adding "Location B" as a waypoint. In response to this edit, the system updates the route and timeline of the corresponding candidate route in real time. Specifically, the system recalculates the walking path and estimated time from the previous waypoint to "Location B," and from "Location B" to the next waypoint or destination, and adjusts the estimated arrival time of the entire route accordingly. The updated route and timeline are then output as the planned route to User A's user client.

[0039] Based on the above example, existing technologies, when processing user-input route requirement information, often directly trigger geographic data source queries when constraint values ​​are missing. This can easily lead to the return of a large amount of irrelevant spatial data or even direct error interruption. However, the solution in this embodiment constructs a dynamic verification and pre-completion mechanism through the steps of "building a constraint item queue based on the travel constraint set, traversing the constraint item queue according to a preset order, and generating corresponding supplementary query text and sending it to the user when a constraint value is missing during the traversal" and "receiving supplementary information input by the user based on the supplementary query text, extracting the corresponding constraint value based on the supplementary information and filling it into the constraint item queue." In the example above, when user A initially only provided "location A" and "half a day," the system did not immediately perform a recall, but instead initiated a query for the missing "travel mode" and "scenario type." This effectively reduces invalid recalls or system interruptions caused by incomplete information, ensuring an effective closed loop in data flow.

[0040] Furthermore, existing systems typically output statically bound route data. After route planning is complete, if local nodes change, the system usually needs to re-enter all nodes into the path planning algorithm for global recalculation. This lacks dynamic response capability to multi-dimensional constraints and makes it difficult to locally and adaptively update the physical orientation and timestamp association of subsequent routes. The solution in this embodiment achieves local adaptive route updates by "responding to the user's editing operation on any candidate route, updating the corresponding candidate route's orientation and timeline in real time, and outputting the updated candidate route as the planned route." In the example above, when user A adds "location B" to the planned route, the system does not need to globally recalculate the entire route. Instead, it can locally update the affected orientation and timeline, thereby improving the system's dynamic response capability and user experience, ensuring route consistency in the spatiotemporal dimensions, and reducing unnecessary computational resource consumption.

[0041] The following will further explain an AI-based user intent recognition route planning method in this exemplary embodiment.

[0042] In one embodiment, step S10, "receiving route demand information input by the user terminal and extracting a set of travel constraints based on the route demand information," specifically includes: Receive route requirement information input from the user terminal and determine the input type of the route requirement information, wherein the input type includes link information and text information; In embodiments of this application, this step aims to identify the specific form in which the user provides route request information in order to adopt targeted processing strategies. Link information typically refers to web page addresses shared by the user that contain travel guides, attraction descriptions, etc., while text information refers to natural language descriptions directly entered by the user. By distinguishing between these two common input types, the system can flexibly adapt to different user input habits and improve the universality of information reception.

[0043] When the input type is link information, the web page content corresponding to the link information is obtained, and text data and image data are extracted based on the web page content. The text data and image data are then merged to form the text to be parsed. In the embodiments of this application, when the user provides a webpage link, the system first needs to obtain the full content of the webpage pointed to by the link through a network request, such as using the HTTP / HTTPS protocol. Subsequently, the system parses the structure of the webpage, such as HTML or XML, and extracts the visible text information. Simultaneously, for images contained in the webpage, the system can use optical character recognition technology to identify the text information in the images, or use image recognition technology to analyze the image content to obtain relevant semantic information. Finally, all the text and image data extracted from the webpage are integrated to form the parsed text available for subsequent analysis, ensuring that all potential constraints are included in the processing scope. When integrating text and image data, the image data is usually in processed text form.

[0044] Entity information is extracted from the text to be parsed using a pre-trained entity recognition model. The entity information includes place name entities, time entities, and type entities. In the embodiments of this application, the entity recognition model is an artificial intelligence model whose function is to identify and extract named entities with specific meanings from unstructured text. The entity recognition model is typically trained on a large corpus and can accurately identify information such as place names, times, and types in the text. Through the entity recognition model, the system can transform the raw information input by the user into structured entity data, laying the foundation for subsequent constraint construction.

[0045] Specifically, the pre-trained entity recognition model described in this embodiment is a deep neural network architecture based on a pre-trained language model and a conditional random field (CRF). Regarding input and structural features, the model first performs word segmentation mapping on the text to be parsed through a BERT encoding layer, extracting a character-level semantic vector matrix containing global contextual features. Subsequently, this semantic vector matrix is ​​input to a CRF layer, which learns the transition probabilities between adjacent labels, thereby eliminating illegal label sequences, such as the case in the BIO labeling strategy where the internal entity label "I" appears before the entity start label "B". Regarding output and training, the entity recognition model ultimately outputs the entity category label sequence corresponding to each character, which is then merged to obtain place name entities, time entities, and type entities. In the pre-training and fine-tuning stages, the entity recognition model is trained under supervision using a dataset containing a large amount of route planning corpus and geographic POI description text. The negative log-likelihood between the true labels and the predicted sequences is used as the loss function, and the network weights are updated through backpropagation, enabling the entity recognition model to accurately capture geospatial and temporal entities in complex multimodal text.

[0046] The extracted place name entities are matched with the geographic information database to obtain standard place names and their corresponding geographic coordinates. The standard place names and their corresponding geographic coordinates, time entities, and type entities are used as spatial constraints, duration constraints, and scenario constraints, respectively, to obtain a set of travel constraints. In the embodiments of this application, to ensure the accuracy and standardization of place name information, the system compares and matches the place name entities extracted by the entity recognition model with a pre-built geographic information database. This geographic information database typically contains detailed information such as place names, aliases, geographic coordinates, and point-of-interest classifications for global or specific regions. Upon successful matching, standardized place names and their precise geographic coordinates are obtained. Subsequently, the obtained standardized place names and geographic coordinates are defined as spatial constraints, temporal entities as duration constraints, and type entities as scene constraints, collectively forming a set of travel constraints to provide explicit input for subsequent route planning.

[0047] When the input type is text information, a pre-trained natural language understanding model is used to extract multiple constraint values ​​based on route demand information to obtain a set of travel constraints.

[0048] In the embodiments of this application, when a user directly inputs route requirements in natural language text form, the system can invoke a pre-trained natural language understanding model. This model possesses powerful semantic analysis capabilities, enabling it to directly identify the user's intent from the user's free text and extract multiple constraint values ​​related to route planning. For example, for the text "I want to go to Shanghai Disneyland this weekend, preferably a family activity," the natural language understanding model can directly identify "Shanghai Disneyland" as a spatial constraint, "weekend" as a duration constraint, and "family activity" as a scenario constraint.

[0049] For example, as a specific implementation, suppose a user wants to plan a trip and is provided with two different route requirements. One specific example is as follows: Scenario 1: A user inputs a webpage link for a travel guide. Upon receiving this link, the system determines that the input type is a link. Further, the system accesses the webpage, retrieves its HTML content, and extracts text data such as "Beijing Forbidden City," "Great Wall," "three days," and "cultural experience." If the webpage contains images with text, the system can also use optical character recognition (OCR) technology to identify text such as "Beijing" and "Forbidden City" within the images and merge them to form the text to be parsed.

[0050] Based on this, the pre-trained entity recognition model will identify "Beijing Forbidden City" and "Great Wall" as place name entities, "three days" as a time entity, and "cultural experience" as a type entity from the text to be parsed. At this point, the system will match "Beijing Forbidden City" and "Great Wall" with the geographic information database to obtain their standard location names and precise geographic coordinates.

[0051] Ultimately, this information is transformed into spatial constraints, duration constraints, and scenario constraints, which together constitute a set of travel constraints.

[0052] Scenario 2: The user directly enters text information into the input box, such as "I want to go to Shanghai Disneyland this weekend, preferably a family activity." Upon receiving this text information, the system determines that the input type is text. In this case, the pre-trained natural language understanding model directly performs semantic analysis on this text information, identifying the user's intention as travel planning. It then extracts "Shanghai Disneyland" as a spatial constraint, "weekend" as a duration constraint, and "family activity" as a scenario constraint, collectively forming a set of travel constraints.

[0053] Through the above technical solution, this application can effectively solve the problem of information extraction difficulties caused by the diverse input formats of user route demand information. Specifically, by distinguishing between two input types—link information and text information—and employing webpage content parsing combined with entity recognition models, or natural language understanding models, respectively, the system can accurately extract standardized place name entities, time entities, and type entities from different forms of unstructured input, and transform them into structured spatial constraints, duration constraints, and scene constraints, forming a complete set of travel constraints. Based on this, the solution of this embodiment can effectively improve the flexibility and convenience of user input, allowing users to choose the most convenient input method according to their own habits, and ensuring that subsequent route planning steps can obtain high-quality structured input data, thereby improving the accuracy of route planning and user satisfaction.

[0054] In one embodiment, the natural language understanding model includes an encoding layer, a classification layer, and an imputation layer. The step "when the input type is text information, extracting multiple constraint values ​​based on route demand information through a pre-trained natural language understanding model to obtain a travel constraint set" specifically includes: The route requirement information is converted into a semantic vector through the encoding layer, and the intent type of the route requirement information is identified through the classification layer. In the embodiments of this application, the encoding layer is a core component of the natural language understanding model. Its function is to transform the raw route requirement information into a computer-processable numerical representation, namely a semantic vector. This transformation can capture the vocabulary, grammar, and contextual relationships in the text, laying the foundation for subsequent intent recognition and information extraction. The encoding layer can be implemented using models based on the Transformer architecture, such as BERT, RoBERTa, and GPT, which effectively capture long-distance dependencies through self-attention mechanisms; or models based on recurrent neural networks and their variants, such as LSTM and GRU, which excel at processing sequential data. The classification layer follows the encoding layer. Its function is to identify the overall intent type of the user's route requirement information based on the semantic vector output by the encoding layer. The intent type is an abstract summary of the user's core needs, such as "travel planning," "commuting route," and "restaurant recommendation." The classification layer can be implemented using one or more fully connected layers combined with a Softmax activation function to map the semantic vector onto a predefined intent category probability distribution, thereby determining the most likely intent type; or using a multi-label classifier to handle situations where user needs may contain multiple intents.

[0055] The constraint values ​​corresponding to the intent type are extracted from the semantic vector through the padding layer; In the embodiments of this application, the function of the padding layer is to extract specific constraint values ​​related to the intent type identified by the classification layer from the semantic vector generated by the encoding layer. For example, if the intent type is "travel planning," the padding layer can attempt to extract constraint values ​​such as destination, duration, and preference type. The implementation of the padding layer can include: using sequence labeling-based models, such as CRF or Bi-LSTM-CRF, to label each word in the text as a specific entity type, such as location, time, or activity; or using pointer networks or Span Extraction methods to directly locate and extract the text fragments corresponding to the constraint values ​​from the semantic vector.

[0056] The extracted constraint values ​​are validated for validity, and after the validity validation is passed, the intent type and the extracted constraint values ​​are combined to obtain the travel constraint set.

[0057] In the embodiments of this application, the purpose of validity verification is to check whether the extracted constraint values ​​conform to preset rules, formats, or logic, so as to avoid data anomalies caused by model misidentification or user input errors. The implementation of validity verification may include: rule-based verification, such as checking whether the time format and location name exist in a known geographic database; range-based verification, such as whether the duration is within a reasonable range; or verification of the accuracy of entity information by comparing with a knowledge graph. Combining the intent type and the extracted constraint values ​​to obtain a travel constraint set is a step that structurally integrates the validated intent type and the specific constraint values ​​extracted from the text to form a complete travel constraint set. This combination ensures that each constraint value is associated with its corresponding intent type, and that the entire set has logical self-consistency.

[0058] Furthermore, the natural language understanding model described in this embodiment adopts a joint learning multi-task network architecture to achieve simultaneous processing of intent recognition and slot filling. In terms of network layer architecture, the encoding layer includes a multi-layer self-attention mechanism module to map unstructured route demand information into high-dimensional hidden layer semantic vectors; the classification layer is a multi-layer perceptron with an external Softmax activation function, which receives global feature representations from the semantic vectors and outputs the intent type probability distribution of the route demand information; the filling layer uses a pointer network or sequence labeling network, based on the sequence features output by the encoding layer, to accurately segment the specific constraint values ​​corresponding to the intent type by predicting the start and end position indices. Regarding the model training mechanism, a multi-task joint loss function can be constructed. Specifically, the cross-entropy loss of the classification layer and the sequence loss of the filling layer are linearly weighted according to preset weights. During training, a supervision set constructed based on real-world travel business dialogues can be used. By jointly optimizing this weighted loss function, the natural language understanding model can fully utilize intent types as prior knowledge for semantic disambiguation when extracting multiple constraint values, improving its understanding accuracy under complex interactions.

[0059] For example, as one specific implementation, suppose a user enters route requirements on the user's device: "I want to go to Shanghai Disneyland for two days this weekend, preferably a family-friendly one, and with convenient transportation." First, the encoding layer receives the text and uses a pre-trained Transformer model to convert it into a semantic vector containing rich semantic information. This semantic vector captures keywords such as "weekend", "Shanghai Disneyland", "two days", "family theme", "convenient transportation" and their interrelationships.

[0060] Building upon this, the classification layer uses a multilayer perceptron classifier to identify the intent type of this route demand information as "family travel planning" based on this semantic vector. Subsequently, the fill layer extracts the corresponding constraint values ​​from the semantic vector based on this intent type. For example, using a sequence labeling model, it identifies "Shanghai Disneyland" as a spatial constraint, "two days" as a duration constraint, and "family theme" as a scene constraint. Then, the system validates the extracted constraint values. For example, it checks whether "Shanghai Disneyland" is a valid geographical entity and whether "two days" is a reasonable time length.

[0061] Assuming all the above constraints are validated, the system will then combine the intent type "family travel planning" with the validated spatial, duration, and scenario constraints to form a complete set of travel constraints.

[0062] The above technical solution refines the natural language understanding model into a multi-layered architecture including encoding, classification, and padding layers, and introduces a validity verification mechanism, effectively addressing the diversity and complexity of user input text information. This layered processing approach based on the natural language understanding model enables it to accurately identify user intent and then selectively extract constraint values ​​related to that intent, avoiding the challenge of extracting all information from complex text at once. Simultaneously, validating the extracted constraint values ​​allows for timely detection and correction of model recognition errors or non-standard user input, thereby improving the accuracy and reliability of extracting travel constraint sets from text route demand information. This not only enhances the efficiency of user intent recognition but also provides more accurate and high-quality input for subsequent route generation, ultimately optimizing the overall user experience of route planning.

[0063] In one embodiment, step S20, "constructing a constraint item queue based on the travel constraint set, traversing the constraint item queue according to a preset order, and generating corresponding supplementary query text and sending it to the user terminal when a constraint value is missing during the traversal," specifically includes: Based on the preset constraint priority configuration, the constraints in the travel constraint set are arranged in order to construct a constraint queue; In the embodiments of this application, the preset constraint priority configuration refers to a set of rules or weights pre-defined by the system to determine the importance or query order of each constraint in the travel constraint set during processing. The constraint priority configuration can be based on expert experience or historical data statistical analysis, forming a fixed priority list. For example, the origin and destination usually have the highest priority, followed by time, and then the number of people or preferences. Furthermore, the constraint priority configuration can also be dynamically adjusted. For example, the priority of each constraint can be adjusted in real time according to user profiles, current context information, or specific scenario requirements to adapt to different users and situations. Arranging the constraints in the travel constraint set in order to construct a constraint queue aims to ensure that the system can obtain or confirm constraint values ​​according to a predetermined importance or logical order, thereby improving interaction efficiency and user experience. Specifically, each constraint in the travel constraint set and its priority can be stored as key-value pairs, and then sorted using sorting algorithms such as bubble sort or quick sort according to priority to form an ordered queue. Alternatively, a linked list or array structure can be used to dynamically maintain an ordered queue of constraints by inserting them into the correct position in the queue according to their priority when inserting them.

[0064] Traverse the constraint item queue according to the preset order and check whether the constraint value corresponding to each constraint item exists. In the embodiments of this application, traversing the constraint item queue according to a preset order and checking whether the constraint value corresponding to each constraint item exists is to check whether each constraint item has obtained a valid constraint value, so as to identify the information that needs to be supplemented by the user. Specifically, this check can be implemented through an iterator or a loop structure, starting from the head of the queue and sequentially accessing each constraint item, and calling a check function to determine whether the constraint value field of the current constraint item is empty, whether it conforms to a preset data format, or a valid range. Another approach is to perform a preliminary check on all constraint items at once after the constraint item queue is built, and mark the existing and missing constraint values, and only focus on the missing items during subsequent traversals.

[0065] When a constraint value for any constraint term is detected to be missing, supplementary query text is generated based on the type of the current constraint term and the already acquired constraint value, using a pre-trained text generation model. In the embodiments of this application, the purpose of this step is to generate natural and clear query statements in a targeted manner, guiding users to provide missing key information and avoiding vague or repetitive questions. The text generation model can be a pre-trained model based on the Transformer architecture. By inputting the type of the currently missing constraint and other relevant constraint values ​​already acquired as context, the text generation model can generate query statements that conform to the context. Alternatively, it can be based on template filling, with multiple preset query templates. The appropriate template is selected according to the type of missing constraint, and the variables in the template are filled with the acquired constraint values. For example, if the duration is missing and the location is known, it can generate "How long do you plan to stay at [location]?" The generated supplementary query text is sent to the user's client.

[0066] In the embodiments of this application, sending the generated supplementary query text to the user terminal is to present the system-generated query information to the user to obtain necessary supplementary input. Specifically, the text can be sent to the chat window or dialog box of the user interface through an API interface for the user to view and input a reply. Alternatively, the text can be converted into speech using speech synthesis technology and played to the user through a voice assistant or smart speaker.

[0067] Furthermore, to ensure the supplementary query text is contextually natural and clearly targeted, the pre-trained text generation model described in this embodiment adopts an encoder-decoder-based sequence-to-sequence architecture, such as the BART or T5 model. Regarding input construction and generation logic, when the system detects missing constraint values, it does not directly input random strings. Instead, it employs templated prompting technology, using "acquired constraint values" as the context and "the type of missing constraint" as the target guidance point, concatenating them to construct a structured input sequence. Subsequently, the decoder of the text generation model uses an autoregressive approach, combined with a beam search algorithm, to globally optimize the word probabilities generated at each time step, dynamically generating fluent and logically sound supplementary query text. In terms of model training and alignment, this text generation model can not only undergo unsupervised pre-training using massive amounts of general dialogue data in the early stages, but also undergo supervised fine-tuning using standard "customer service-user" travel planning dialogue data during the domain adaptation phase. Its training objectives not only include the standard generative cross-entropy loss, but also introduce a rule-based penalty mechanism. For example, if the generated query text does not contain the lead word of the missing constraint term, an additional loss penalty is applied, thereby ensuring that the generated query text has extremely high relevance and business usability.

[0068] For example, as a specific implementation, suppose a user inputs their route request information via voice: "I want to go to Beijing for about three days." After initial parsing, the system extracts spatial constraints and duration constraints from the travel constraint set, but finds that key information such as scenario constraints, mode of travel constraints, intensity constraints, and group constraints are missing.

[0069] At this point, the system will configure the constraints according to the preset priority settings. For example, if the priority setting prioritizes scenario constraints over travel mode constraints, the system will arrange the constraints in the travel constraint set in order to construct a constraint queue. The order might be: spatial constraints (filled), duration constraints (filled), scenario constraints (missing), travel mode constraints (missing), intensity constraints (missing), and group constraints (missing). The system will then traverse the constraint queue according to the preset order, first checking the spatial and duration constraints and finding that their constraint values ​​already exist. When it reaches the scenario constraint, it will detect that its constraint value is missing. Based on this, the system can generate supplementary query text using a pre-trained text generation model, based on the type of the current constraint and the already acquired constraint value. For example, the text generation model might generate: "You're going to Beijing for three days. What types of attractions do you mainly want to visit? For example, historical sites, natural scenery, or shopping and food?" The system will then send the generated supplementary query text to the user, waiting for the user to input additional information.

[0070] Finally, after receiving the supplementary information input by the user, the system can extract the corresponding constraint values ​​and fill them into the scene constraint items, and then continue to traverse the queue or perform subsequent processing.

[0071] Through the above technical solution, this application can arrange the constraints in the travel constraint set in an orderly manner according to the preset constraint priority configuration, thereby constructing a constraint queue. When traversing the constraint queue and detecting a missing constraint value, the system can generate highly targeted supplementary query text based on the type of the current constraint and the acquired constraint value, using a pre-trained text generation model, and send it to the user terminal. Based on this, the solution of this embodiment can improve the efficiency and accuracy of user interaction, avoid redundant or ambiguous questions, and enable users to provide the required travel constraint information more quickly and accurately. Thus, the system can obtain the complete travel constraint set more efficiently, optimize the user experience, and reduce the cognitive burden and operational costs for users in the information supplementation stage.

[0072] Furthermore, in another embodiment, before the step of "arranging the constraints in the travel constraint set in order according to the preset constraint priority configuration to construct a constraint queue" may further include: Extract entity attributes with acquired constraint values, and calculate the implicit association probability between entity attributes and each missing constraint item based on a pre-defined common sense knowledge graph; When the implicit association probability of a missing constraint is greater than a preset confidence threshold, the predicted constraint value of the missing constraint is generated based on the characteristics of the corresponding association nodes in the common sense knowledge graph, and the query for the missing constraint is skipped. When the implicit association probability is less than or equal to the preset confidence threshold, the information entropy gain of the missing constraint is calculated, and the information entropy gain is used as a weighting factor to dynamically adjust the priority configuration of the constraint.

[0073] Extracting entity attributes from acquired constraint values ​​refers to identifying and extracting descriptive features or properties from user-provided, populated constraint values. Entity attributes can be semantic, categorical, or relational. For example, natural language processing techniques, such as named entity recognition or keyword extraction, can be used to identify specific entities and their attributes from textual constraint values. Alternatively, predefined ontology or classification systems can be used to map acquired constraint values ​​to their corresponding attribute labels. A pre-defined common-sense knowledge graph is a structured knowledge base that represents entities, concepts, and their relationships in graph form. It pre-stores a large amount of common-sense knowledge, which can be used for reasoning and associating different information. The common-sense knowledge graph can be a graph database containing a large number of entities (such as locations, activities, groups, interests), attributes (such as suitable groups, activity types, seasons), and their relationships (such as "suitable," "contains," "located in"), or a hybrid knowledge representation combining semantic networks and ontology. Calculating the implicit association probability between entity attributes and each missing constraint aims to quantify the potential association strength between acquired entity attributes and currently missing constraints, in order to determine whether the missing information can be inferred from known information. Specifically, this can be achieved by performing pathfinding or graph embedding algorithms on a knowledge graph, calculating the semantic distance or similarity between entity attribute nodes and missing constraint node nodes, and converting this into probability values. Alternatively, statistical learning methods can be used to train a classifier or regression model on a large amount of historical user data and knowledge graph data, taking the acquired entity attributes as input and outputting the predicted probability of the missing constraints. A preset confidence threshold is a pre-defined value used to determine whether the implicit association probability is high enough to trust the automatically generated predicted constraint values ​​without asking the user. The confidence threshold can be determined based on experience or experimental testing, or dynamically adjusted according to the application scenario or user preferences. Generating the predicted constraint value of the missing constraint directly based on the characteristics of associated nodes in the knowledge graph means using node information in the knowledge graph that is strongly associated with the attributes of the acquired entity to directly infer and generate the value of the missing constraint. Specifically, this can be achieved by querying attribute values ​​in the knowledge graph that are directly associated with the attributes of the acquired entity and match the type of the missing constraint, or by using a rule-based inference engine that combines facts and predefined rules in the knowledge graph to extract or calculate the predicted constraint value from associated nodes. Skipping the query for the missing constraint means that after successfully predicting the missing constraint value, no further query text is sent to the user regarding the missing constraint. Instead, the predicted value is directly added to the constraint queue, and the constraint is marked as completed.Calculating the information entropy gain of missing constraints aims to measure the importance or information content of a missing constraint in determining its priority. Generally, a larger information entropy gain indicates a greater impact of the missing constraint on subsequent decisions in the current state. This can be calculated by analyzing the missing constraints in historical data and their impact on the final route planning result, or by constructing a decision tree model, treating missing constraints as potential decision nodes, and calculating which constraint, given the current information, can minimize uncertainty. Dynamically adjusting constraint priority configuration using information entropy gain as a weighting factor aims to adjust the query priority of missing constraints in the constraint queue based on their contribution to the information gain, prioritizing queries on constraints with the greatest impact on user intent recognition or route planning. Specifically, the calculated information entropy gain value can be directly used or converted into a priority score and combined with a preset priority configuration to reorder the constraint queue, or a weighted summation model can be used to combine the information entropy gain with a preset static priority to generate a dynamic priority.

[0074] The above solution effectively addresses the problem of inefficient interactions caused by missing information in traditional methods. Specifically, by utilizing entity attributes and common-sense knowledge graphs with already acquired constraint values, it's possible to predict and fill in missing constraint values, reducing the number of times users need to manually input information and thus improving user experience. Furthermore, for constraints that cannot be automatically predicted, the query priority is dynamically adjusted by calculating information entropy gain, ensuring that the most critical information with the greatest impact on route planning is obtained first. This makes the entire information completion process more efficient and accurate, ultimately improving the accuracy of route planning and user satisfaction.

[0075] In one embodiment, the travel constraint set includes spatial constraints, duration constraints, scenario constraints, and travel mode constraints. Step S40, "when it is determined that the travel constraint set meets the route generation conditions, the corresponding geographic data source is invoked to perform candidate point recall based on the travel constraint set," specifically includes: Determine whether all mandatory constraints in the travel constraint set have been filled with constraint values. If all mandatory constraints have been filled with constraint values, determine that the travel constraint set meets the route generation conditions. In the embodiments of this application, the travel constraint set refers to structured data formed after parsing user route demand information, used to limit various conditions of route planning. The travel constraint set includes spatial constraints, duration constraints, scenario constraints, and travel mode constraints. Spatial constraints limit the geographical location of the route's starting point, ending point, transit points, or activity area. This can be represented by specific geographical coordinates, district names, point-of-interest names, or precise spatial retrieval boundaries composed of multiple discrete or continuous polygons. Duration constraints limit the time required for the entire trip or a segment of the trip, such as total duration, daily duration, or a specific time period. Scenario constraints limit the types of activities or points of interest involved in the route, such as predefined categories like food, shopping, culture, and natural scenery, or user-defined keywords. Travel mode constraints limit the mode of transportation selected by the user on the route, such as walking, cycling, public transportation, driving, or a combination of multiple modes. Before recalling candidate points, the system checks whether all mandatory constraints in the travel constraint set have been populated with constraint values. This check ensures that sufficient and necessary route planning information is obtained before computationally intensive candidate point recall, avoiding recall failures or inaccurate results due to missing key information. For example, the system can pre-define a list of mandatory constraints, such as origin, destination, and total duration, and iterate through this list after each update of the travel constraint set, checking whether each mandatory constraint contains a valid constraint value. If all mandatory constraints are populated, the condition is met. Alternatively, a rule-based decision engine can be used, pre-defining a series of logical rules, such as "if there is a spatial constraint but no duration constraint, the condition is not met." The decision condition is met when all rules are passed.

[0076] Extract the geographic coordinates corresponding to the spatial constraints from the travel constraint set, set the recall radius with the geographic coordinates as the center point, and calculate the value of the recall radius based on the duration constraint and travel mode constraint. In the embodiments of this application, this step aims to clarify the geographical scope of candidate point recall, ensuring that the recalled points of interest match the user's spatial, temporal, and travel mode needs, thereby improving the accuracy and efficiency of the recall. For example, the system can extract the geographical coordinates of the explicitly specified starting point or center point in the spatial constraints as the recall center point. The recall radius can be estimated based on the user's expected total duration and the average speed of the selected travel mode; for example, radius = duration * average speed * coverage coefficient. If the spatial constraint is a region rather than a single point, the geometric center of this region can be calculated as the recall center point. Furthermore, the recall radius can be calculated by considering the user's expected activity intensity and travel mode, through table lookup or a preset model.

[0077] The candidate point type is determined based on the scenario constraints, and the corresponding geographic data source is selected based on the candidate point type; In embodiments of this application, this step aims to filter relevant types of points of interest based on user preferences and select the most suitable geographic data source to provide this point of interest data, thereby improving the relevance and data quality of the recall results. For example, the system can maintain a mapping table between scene constraints and candidate point types. When a scene constraint is received, the system queries this mapping table to determine one or more candidate point types, and then selects a pre-configured geographic data source based on the candidate point types. Alternatively, a machine learning-based classification model can be used. The scene constraints are input into the model, the model outputs the corresponding candidate point types, and a geographic data source that can provide high-quality relevant data is selected based on the output candidate point types.

[0078] Send a recall request to the selected geographic data source, the recall request including the geographic coordinates of the center point, the recall radius, and the type of point of interest; In embodiments of this application, this step aims to encapsulate the processed and determined recall parameters into a standard format request and send it to an external geographic data source to obtain geographic location point data that meets the user's needs. For example, the system can construct an HTTP / HTTPS request containing the centroid geographic coordinates, recall radius, and point of interest type in JSON or XML format. Alternatively, it can call the SDK or API interface provided by the geographic data source, passing in the corresponding parameter object.

[0079] Receive a set of candidate points returned by a geographic data source. The set of candidate points includes geographic locations and their attribute information that are within the numerical range of the recall radius and conform to the candidate point type.

[0080] In the embodiments of this application, this step aims to acquire and process the raw data returned by the geographic data source, transforming it into a candidate point set usable within the system, providing foundational data for subsequent filtering and sorting. For example, the system can receive response data in JSON or XML format from the geographic data source, parse the response data, extract attribute information such as the name, latitude and longitude, address, rating, business hours, and image URL of each geographic point, and encapsulate it into a unified candidate point data structure. Alternatively, the received data can be preliminarily validated, for example, checking whether the returned geographic points are indeed within the recall radius and whether they match the requested point of interest type, filtering out points that do not meet the conditions to ensure the quality of the candidate point set.

[0081] For example, as a specific implementation, suppose a user inputs route requirements on their device: "I want to start from People's Square in a certain city, spend 3 hours cycling, and go to a place with historical and cultural significance." After receiving this route requirement, the system will parse it using a pre-trained natural language understanding model and extract the set of travel constraints.

[0082] In this example, the travel constraint set will include: spatial constraint: the starting point is the People's Square in a certain city, and its corresponding geographical coordinates; duration constraint: the total duration is 3 hours; scene constraint: the type is "historical and cultural"; and travel mode constraint: the type is "cycling". Before recalling candidate points, the system will check whether these constraints have been filled. Assuming that "starting point", "total duration", "scene type" and "travel mode" are all defined as mandatory constraints, and in this example, these constraints have all obtained valid values, the system can determine that the travel constraint set meets the route generation conditions.

[0083] Subsequently, the system extracts the geographical coordinates of a city's People's Square from the travel constraint set as the recall center point. Simultaneously, it calculates the recall radius by combining the "3-hour" duration constraint and the "cycling" travel mode constraint. For example, the system can query a preset speed mapping table to find that the average cycling speed is 15 km / h. Therefore, the recall radius can be estimated as: 3 hours * 15 km / h * 0.8 = 36 km, where 0.8 is a coverage coefficient used to adjust the actual reachable range.

[0084] Based on this, and according to the contextual constraint of "historical culture," candidate point types are determined as "historical sites," "museums," and "cultural districts," among others. Based on these candidate point types, the system selects the corresponding geographic data source; for example, it might choose a third-party API interface that specifically provides historical and cultural POI information. Subsequently, the system sends a recall request to this selected geographic data source. The recall request includes the geographic coordinates of a city's People's Square, a recall radius of 36 kilometers, and the point of interest type such as "historical sites, museums, and cultural districts."

[0085] Finally, the system receives a set of candidate points returned by the geographic data source. This set may include geographical locations such as a city museum or a historical building complex, and each location includes its name, latitude and longitude, brief description, and opening hours. Furthermore, these candidate points are all located within a 36-kilometer radius of People's Square, providing a rich and relevant selection for subsequent route planning.

[0086] Through the above technical solutions, this embodiment effectively addresses the technical challenges of determining the recall timing and efficiently defining the recall scope and type in user intent recognition route planning. Specifically, by clearly defining the composition of the travel constraint set and introducing a filling mechanism for mandatory constraints, invalid recall operations can be avoided when key information is incomplete, thereby saving computational resources and improving system efficiency. Furthermore, this embodiment dynamically calculates the recall radius by combining spatial constraints, duration constraints, and travel mode constraints, ensuring the recall scope adapts to the user's actual travel capacity and time budget, avoiding issues of the scope being too large or too small due to fixed-radius recall. Simultaneously, selecting the corresponding geographic data source and candidate point type based on scenario constraints ensures the recall results are highly relevant to user interests, improving recall accuracy. Based on this, the recall strategy enables the system to quickly and accurately obtain the candidate points that best meet user needs from massive amounts of geographic data, providing a solid data foundation for subsequent route generation. Compared to existing basic solutions, this embodiment achieves a high degree of intelligence and customization in the recall stage, improving route planning accuracy and user satisfaction.

[0087] In one embodiment, the step of "extracting the geographic coordinates corresponding to the spatial constraints from the travel constraint set, setting the recall radius with the geographic coordinates as the center point, and calculating the value of the recall radius based on the duration constraint and the travel mode constraint" specifically includes: Extract the origin geographic coordinates corresponding to the spatial constraints from the travel constraint set, and use the origin geographic coordinates as the recall center point; In the embodiments of this application, the starting point geographic coordinates refer to the geographic coordinates of the starting location, whether explicit or implicit, in the user's route requirement information, such as latitude and longitude. These starting point geographic coordinates can be obtained from the spatial constraints parsed from the user's input route requirement information and serve as the center point for subsequent recall operations. The recall center point refers to the geometric center that determines the recall range; in this embodiment, all potential candidate points will be used as a reference for distance calculation.

[0088] Extract the duration value corresponding to the duration constraint and the travel mode type corresponding to the travel mode constraint from the travel constraint set; In the embodiments of this application, the duration value refers to the total trip duration expected or allowed by the user, which can be expressed in time units such as minutes or hours, and can usually be extracted from the duration constraint item in the trip constraint set. The travel mode type refers to the mode of transportation selected by the user, such as walking, cycling, driving, public transportation, etc., and can usually be extracted from the travel mode constraint item in the trip constraint set.

[0089] Based on the mode of travel type, query the preset speed mapping table to obtain the average speed value corresponding to the mode of travel type; In the embodiments of this application, the preset speed mapping table is a data structure that stores different travel mode types and their corresponding average speed values. For example, it can be a database table, a configuration file, or a hash table in memory. The average speed value is obtained by querying the preset speed mapping table according to the travel mode type, and it can represent the travel speed under ideal or average conditions.

[0090] The recall radius is calculated based on the duration, average velocity, and preset coverage coefficient.

[0091] In the embodiments of this application, the preset coverage coefficient is a value between 0 and 1, used to correct the theoretically calculated maximum reachable distance to more accurately reflect the actual reachable range. This coverage coefficient can be an empirical value or a fixed value preset based on historical data or specific regional characteristics. Calculating the recall radius involves multiplying the duration value, average velocity value, and the preset coverage coefficient, or performing a more complex function calculation, to obtain the final radius value used for candidate point recall.

[0092] For example, as a specific implementation, suppose the user's route requirement information includes "starting from a city center square, cycling for 2 hours", and the travel constraint set has extracted the geographical coordinates of the city center square as a spatial constraint, 2 hours as a duration constraint, and cycling as a mode of travel constraint.

[0093] Based on this, the system uses the geographical coordinates of the city center square as the recall center point, extracts the duration value of 2 hours and the travel mode type as "cycling" from the travel constraint set. Then, the system queries a preset speed mapping table; for example, the speed mapping table records an average speed of 15 km / h for "cycling". Simultaneously, it retrieves or queries a preset coverage coefficient, for example, 0.8. Therefore, the recall radius can be calculated as: 15 km / h * 2 hours * 0.8 = 24 km.

[0094] Ultimately, the system will recall candidate locations within a 24-kilometer radius of the city center square.

[0095] Through the above technical solution, the scheme in this embodiment, when calculating the recall radius, not only considers travel time and the average speed of the travel mode, but also further introduces a preset coverage coefficient to correct the theoretical reachable distance, making the calculation of the recall radius closer to the actual reachable range, thereby improving the accuracy and efficiency of candidate point recall. Based on this, the recalled candidate point set will be more relevant, reducing the processing of invalid data and providing higher-quality input for subsequent route screening and sorting, thus improving the overall route planning quality and user experience.

[0096] In some of the above implementations, the recall radius is calculated based on duration, average speed, and a preset coverage coefficient. However, in practical applications, simply using a preset fixed coverage coefficient may not accurately reflect the actual traffic efficiency under different geographical areas and travel modes. For example, in areas with dense road networks or complex terrain, the actual distance covered by the same duration and average speed may be significantly smaller than in areas with sparse road networks or flat terrain. This results in an inaccurate setting of the recall radius, which in turn affects the accuracy of candidate point recall and the quality of route planning.

[0097] In one embodiment, the step of determining the preset coverage coefficient may further include: extracting the spatial grid identifier corresponding to the geographic coordinates of the starting point; Query the pre-constructed geographic feature matrix to obtain the road network node density and terrain undulation corresponding to the spatial grid identifier; By inputting the road network node density and terrain undulation into a preset nonlinear mapping function, the actual coverage coefficient of the starting point's geographical coordinates under the current travel mode type is dynamically calculated. The actual coverage coefficient is negatively correlated with the road network node density and the terrain undulation.

[0098] Spatial grid identification is a method of dividing and encoding geographic space to uniquely identify a geographic region. Its function is to discretize continuous geographic coordinates, facilitating regional data querying and management. For example, Geohash encoding can be used to divide the Earth's surface into a series of grids, each corresponding to a unique string identifier; alternatively, a quadtree structure can be used to recursively divide space into four quadrants to generate grid identifiers at different levels. The geographic feature matrix is ​​a pre-built data structure used to store geographic attribute information for different spatial grid regions. It is generated through offline analysis of large amounts of geographic data, such as map data, satellite imagery, elevation data, and traffic flow data. Its function is to provide rich geographic context information for each spatial grid. For example, the geographic feature matrix can store the road network node density, terrain relief, point of interest distribution density, and region type for each grid. Road network node density refers to the density of road intersections or road connections within a specific geographic region, reflecting the complexity and connectivity of the road network within the region. Generally, a higher road network node density indicates more road choices, but traffic efficiency may decrease due to congestion or frequent turns. Topographic relief refers to the degree of elevation change within a specific geographic area, reflecting the region's topographic features, such as mountains, hills, or plains. Greater topographic relief typically indicates a steeper road gradient, significantly impacting travel speed and energy consumption. A nonlinear mapping function is a mathematical model used to convert multiple input geographic features into a single output value, with a nonlinear relationship between the input and output. Its purpose is to capture the complex impact of different geographic features on actual traffic efficiency. For example, multinomial, exponential, logarithmic, or neural network-based models can be used to construct this nonlinear mapping function to more accurately simulate the complex relationships in the real world. The actual coverage coefficient is a dynamically adjusted parameter used to correct the theoretical recall radius calculated based on average speed and duration. Its function is to adjust the recall radius according to the actual geographic features of the starting point's geographic coordinates, making it more consistent with the actual reachable range. Specifically, the actual coverage coefficient is negatively correlated with road network node density (the denser the road network, the shorter the actual reachable distance and the smaller the coverage coefficient); it is also negatively correlated with topographic relief (the more rugged the terrain, the shorter the actual reachable distance and the smaller the coverage coefficient).

[0099] The above method overcomes the problem of inaccurate recall radius caused by the use of a fixed coverage coefficient in traditional methods. Specifically, by dynamically extracting the spatial grid identifier corresponding to the starting point's geographic coordinates and querying the geographic feature matrix to obtain the road network node density and terrain undulation, a nonlinear mapping function is used to calculate the actual coverage coefficient, ensuring that the recall radius calculation fully incorporates the actual geographic environment of the starting point. This adaptive adjustment mechanism ensures that the recall radius setting is more accurate in different scenarios, such as dense road networks, complex terrain, sparse road networks, and flat terrain. Therefore, when performing candidate point recall, the system can recall more relevant and accessible geographic locations, effectively reducing the recall of invalid points that do not conform to reality and avoiding the omission of potential high-quality points, thereby improving the efficiency of candidate point recall, reducing the computational burden of subsequent filtering and sorting, and improving the accuracy of route planning and user experience.

[0100] In one embodiment, the travel constraint set includes intensity constraints and group constraints. Step S50, "generating multiple candidate routes by filtering and sorting based on the candidate point recall results and the travel constraint set," specifically includes: The candidate point recall results are filtered based on the strength constraints and group constraints in the travel constraint set to obtain the filtered point set; In the embodiments of this application, intensity constraints refer to the user's restrictive requirements on the physical exertion, difficulty level, or challenge of activities or experiences along the route. For example, a user may prefer a route that is "easy walk," "moderate-intensity hike," or "high-difficulty climb." Intensity constraints can be quantified as indicators such as physical exertion index, slope variation range, and altitude variation. This can be achieved by the user directly selecting a preset intensity level or indirectly inferred through analysis of the user's historical behavioral data and health data. Group constraints refer to the user's restrictive requirements on the characteristics of fellow travelers or the needs of group activities. For example, a user may prefer a route suitable for "families with children," "senior groups," "professional photography enthusiasts," or "team building activities." Group constraints can include the number of companions, age range, interests, and special needs. This can be achieved by the user inputting information about fellow travelers or by analyzing user profiles, social relationships, and other data. The candidate point recall results are filtered based on intensity constraints and group constraints in the travel constraint set. This aims to eliminate points that do not meet the user's specific requirements for activity intensity and accompanying group, thereby narrowing the scope of subsequent processing and improving route matching accuracy. For example, if the intensity constraint is "easy," candidate points containing challenging climbing areas can be filtered out; if the group constraint is "family-friendly," candidate points unsuitable for children can be filtered out. The filtering process can be implemented based on a pre-set rule base, machine learning model, or expert system. The filtered point set is the set of geographical points that still meet the user's basic requirements after the initial filtering based on intensity and group constraints.

[0101] Each candidate point in the filter point set is scored, and the filter point set is sorted in descending order according to the scoring results. A preset number of candidate points are selected as route nodes. In the embodiments of this application, each candidate point in the selected point set is scored to quantitatively evaluate the selected candidate points and reflect their matching degree with the user's other travel constraints and potential preferences. The scoring can comprehensively consider multiple dimensions such as the candidate point's popularity, user reviews, unique attributes, harmony with the surrounding environment, and accessibility. The scoring model can be implemented using weighted summation, machine learning regression models, etc. The selected point set is sorted in descending order based on the scoring results, ranking the candidate points from best to second best according to their scores, to prioritize geographical locations that better meet the user's expectations. A preset number of candidate points are selected as route nodes, aiming to select the highest-scoring points from the sorted selected point set as key nodes for route construction. These candidate points, serving as route nodes, will be used as input to the route planning algorithm to ensure that the generated route includes the core locations that are most interesting to the user or best meet their needs. Furthermore, the preset number of candidate points can be dynamically adjusted according to system configuration, user needs, or route complexity.

[0102] Based on the route nodes, a preset path planning algorithm is executed to generate multiple candidate routes.

[0103] In the embodiments of this application, the preset path planning algorithm can consider various factors, such as shortest distance, shortest time, fewest transfers, and minimum energy consumption. For example, a variant of the Traveling Salesman Problem can be used to find the optimal order connecting all nodes, or the A* search algorithm can be used to find the optimal path considering a specific cost function. Generating multiple candidate routes aims to provide more than one route option, increasing the user's choice. By adjusting the parameters of the path planning algorithm, considering different node combinations, or optimization objectives, multiple candidate routes with advantages in different aspects can be generated for the user to choose from.

[0104] For example, as a specific implementation, suppose a user inputs via voice, "I want to go on a moderate-intensity outdoor activity with my family, including a 5-year-old child, preferably with a viewpoint and a picnic area." Upon receiving this route request, the system extracts a set of travel constraints, including: spatial constraints (outdoor activity); duration constraints (undefined, may require further inquiry); scene constraints (viewpoint and picnic area); and the intensity and group constraints introduced in this embodiment, namely moderate intensity and a 5-year-old child, respectively. In the initial candidate point recall phase, the system may recall a large number of geographical locations that meet the basic conditions of "outdoor activity," "viewpoint," and "picnic area."

[0105] Based on this, the system will filter these recall results according to intensity constraints and group constraints. For example, all candidate sites containing challenging rock climbing areas or unsuitable for children's play facilities will be eliminated. At the same time, park trails that are too flat and lack challenge may also be excluded to meet the "moderate intensity" requirement. After filtering, a set of selected sites will be obtained, which may include scenic mountain parks, nature reserves with children's play facilities, etc.

[0106] The system then scores each candidate point in the selected point set. This scoring process may take into account factors such as the view from the observation deck, the completeness of the picnic area's facilities, the abundance of children's play facilities, and user reviews. For example, a mountain park that simultaneously boasts a spacious observation deck, a well-equipped picnic area, and a safe children's play area might receive a higher score. Based on the scoring results, the candidate points are sorted in descending order, and, for example, the top 5 highest-scoring points are selected as route nodes.

[0107] Finally, the system will execute a path planning algorithm based on the five selected route nodes to generate multiple candidate routes connecting these route nodes. For example, a loop route starting from the observation deck, passing through the picnic area, and finally reaching the children's play area, or a linear route starting from the picnic area and passing through the observation deck. These generated candidate routes will meet the user's needs for moderate-intensity and parent-child activities.

[0108] Through the above technical solution, the solution in this embodiment can understand and respond to users' personalized travel needs. After initially recalling candidate points, strength constraints and group constraints are introduced to perform a secondary screening of the recall results. This effectively eliminates geographical locations that do not match the user's underlying intentions, allowing the subsequent route planning process to focus on more relevant goals. By scoring and ranking the set of screened points and selecting high-scoring points as route nodes, it can be ensured that the generated route includes locations that are most interesting to the user and best meet their specific needs. This improves the personalization and matching degree of the generated candidate routes, avoids recommending routes that are unsuitable for the user's physical condition or travel group, thereby significantly improving user satisfaction with the planned route and reducing the need for users to manually adjust and edit the route.

[0109] In some of the above implementations, a preset path planning algorithm is proposed to generate multiple candidate routes based on route nodes. However, in actual route planning, users may have specific requirements for route strength. Traditional path planning algorithms may only focus on distance or time optimization, making it difficult to effectively integrate and satisfy such complex strength constraints. This results in deviations between the generated routes and the user's actual needs, affecting the user experience.

[0110] In one embodiment, the step of generating multiple candidate routes by executing a preset path planning algorithm based on route nodes may further include: obtaining elevation data between route nodes and actual connecting paths; By combining elevation data, the three-dimensional distance of the actual connected paths, and the corresponding travel mode types, the expected energy consumption value corresponding to the directed edges between each route node is calculated. The route nodes and expected energy consumption values ​​are constructed into a weighted directed graph, and the strength constraints in the travel constraint set are used as the maximum energy threshold boundary of the entire graph. A restricted depth-first search algorithm is executed on a weighted directed graph to prune search branches whose cumulative expected energy consumption exceeds the maximum energy threshold boundary, generating multiple candidate routes that satisfy the strength constraints.

[0111] Specifically, elevation data refers to the height information of geographical locations, usually expressed as altitude or relative height. Its function is to provide vertical dimension information for route planning, enabling algorithms to assess route gradient changes and elevation gain. Elevation data can be obtained by querying a Digital Elevation Model (DEM) database using a Geographic Information System (GIS) service interface, or by extracting elevation information from processed satellite remote sensing or aerial photogrammetry data. Actual connected paths refer to the real, passable roads or tracks between two route nodes in geographic space. Their function is to ensure the physical feasibility of route planning and provide a basis for calculating path length and complexity. Actual connected paths can be obtained by querying a road network database to find the shortest or optimal actual road connection between two nodes, or by identifying and extracting the user's actual traveled paths using historical trajectory data or crowdsourced data. Expected energy consumption refers to the energy or physical effort expected to be consumed in moving from one route node to another under a specific travel mode. Its function is to quantify abstract strength constraints, enabling route planning algorithms to optimize based on energy consumption. The expected energy consumption value can be calculated using different energy consumption models for different travel modes. This can be achieved by substituting elevation data, 3D distance, and travel mode type into the corresponding model, or by using machine learning models trained on a large amount of user travel data, including routes, elevations, travel modes, and user feedback on intensity, to predict energy consumption for different route segments. In graph theory, a directed edge represents a connection from one node to another, possessing directionality. Here, it represents unidirectional travel from one route node to another. A weighted directed graph is a graph structure where nodes represent route nodes, directed edges represent connections between route nodes, and each directed edge carries a weight, which is the calculated expected energy consumption value. Its purpose is to transform the route planning problem into a graph search problem, facilitating efficient algorithm processing. This weighted directed graph structure can be represented using data structures such as adjacency matrices or adjacency lists, where the elements of the matrix or the items of the list store the expected energy consumption value, or by utilizing existing graph databases or graph processing frameworks to construct and manage the graph structure. The intensity constraint is a user's limitation on the overall difficulty or physical exertion of the route. Here, it is quantified as a maximum energy threshold, serving as a hard constraint for route planning. Its purpose is to ensure that the generated route does not exceed the user's capacity. The intensity constraint can be set by the user by directly inputting a numerical value, such as "total energy consumption not exceeding 500 kcal," or by selecting an intensity level, such as "easy," "moderate," or "challenging." The system then converts this into a specific energy threshold based on a preset mapping relationship.The constrained depth-first search (DFS) algorithm is a graph traversal algorithm that starts from the initial node and explores the branches of the graph as deep as possible until the target node is reached or further exploration is impossible. The "constrained" aspect refers to the consideration of energy threshold boundaries during the search process. Its purpose is to explore all possible paths while satisfying strength constraints. Specifically, this DFS algorithm adds a variable for accumulated energy consumption to the standard DFS algorithm, updating this variable each time an edge is traversed. When the accumulated energy consumption exceeds the maximum energy threshold boundary, the search of the current branch is immediately stopped, and the algorithm backtracks to the previous node. Pruning refers to the process in which, when a branch is found to be unlikely to lead to a valid solution, further exploration of that branch is immediately stopped. Its purpose is to improve search efficiency and avoid unnecessary computation. Multiple candidate routes refer to all alternative paths that meet the user's strength constraints after algorithm filtering. Their purpose is to provide users with diverse choices, improve planning flexibility, and increase user satisfaction.

[0112] By employing the above method, users' abstract requirements for route intensity can be transformed into a quantifiable energy consumption model, which serves as a hard constraint for route planning. When generating multiple candidate routes, the system fully considers the impact of elevation changes, actual road conditions, and travel modes on energy consumption. Through a constrained depth-first search algorithm, path branches that do not meet the intensity constraints are efficiently pruned, enabling the system to generate routes that truly meet users' personalized intensity preferences. This addresses the problem of traditional planning methods that only optimize based on distance or time while ignoring users' physical endurance, thus improving the accuracy of route planning and user satisfaction. Users can obtain personalized route options that match both their destination requirements and their physical condition or willingness to take on challenges.

[0113] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0114] In one embodiment, an AI-based user intent recognition route planning system is provided, which corresponds one-to-one with the AI-based user intent recognition route planning method described in the above embodiments. The AI-based user intent recognition route planning system includes: The constraint extraction module is used to receive route requirement information input by the user and extract a set of travel constraints based on the route requirement information. The queue construction module is used to build a constraint item queue based on the travel constraint set, and traverse the constraint item queue according to a preset order. When a constraint value is missing during the traversal, a corresponding supplementary query text is generated and sent to the user terminal. The constraint supplementation module is used to receive supplementary information from the user based on the supplementary query text input, extract the corresponding constraint values ​​based on the supplementary information, and populate them into the constraint item queue. The recall execution module is used to recall candidate points by calling the corresponding geographic data source based on the travel constraint set when it is determined that the travel constraint set meets the route generation conditions. The route generation module is used to generate multiple candidate routes by filtering and sorting the candidate point recall results in conjunction with the travel constraint set. The route update module is used to respond to the user's editing operation on any candidate route, update the route direction and corresponding timeline of the corresponding candidate route in real time, and output the updated candidate route as the planned route.

[0115] For specific limitations regarding an AI-based user intent recognition route planning system, please refer to the limitations of an AI-based user intent recognition route planning method described above, which will not be repeated here. The modules in the aforementioned AI-based user intent recognition route planning system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0116] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 2 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database is used for data storage, data processing, and data analysis. The network interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements an AI-based user intent recognition route planning method.

[0117] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements an AI-based user intent recognition route planning method.

[0118] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements an AI-based user intent recognition route planning method.

[0119] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0120] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0121] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A user intent recognition route planning method based on AI, characterized in that, Including the following steps: Receive route requirement information input from the user and extract a set of travel constraints based on the route requirement information; A constraint item queue is constructed based on the travel constraint set, and the constraint item queue is traversed according to a preset order. When a constraint value is missing during the traversal, a corresponding supplementary query text is generated and sent to the user. Receive supplementary information from the user based on the supplementary query text input, extract the corresponding constraint values ​​based on the supplementary information and populate them into the constraint item queue; When it is determined that the travel constraint set meets the route generation conditions, the corresponding geographic data source is called to perform candidate point recall based on the travel constraint set. Based on the candidate point recall results, multiple candidate routes are generated after filtering and sorting by the travel constraint set. In response to the user's editing operation on any candidate route, the system updates the route direction and timeline of the corresponding candidate route in real time, and outputs the updated candidate route as the planned route.

2. The AI-based user intent recognition route planning method according to claim 1, characterized in that, The step of receiving route requirement information input by the user and extracting a set of travel constraints based on the route requirement information specifically includes: Receive route requirement information input from the user terminal and determine the input type of the route requirement information, wherein the input type includes link information and text information; When the input type is link information, the web page content corresponding to the link information is obtained, and text data and image data are extracted based on the web page content. The text data and image data are then merged to form the text to be parsed. Entity information is extracted from the text to be parsed using a pre-trained entity recognition model. The entity information includes place name entities, time entities, and type entities. The extracted place name entities are matched with the geographic information database to obtain standard place names and their corresponding geographic coordinates. The standard place names and their corresponding geographic coordinates, time entities, and type entities are used as spatial constraints, duration constraints, and scenario constraints, respectively, to obtain a set of travel constraints. When the input type is text information, a pre-trained natural language understanding model is used to extract multiple constraint values ​​based on route demand information to obtain a set of travel constraints.

3. The AI-based user intent recognition route planning method according to claim 2, characterized in that, The natural language understanding model includes an encoding layer, a classification layer, and an imputation layer. The step of extracting multiple constraint values ​​based on route demand information using a pre-trained natural language understanding model to obtain a travel constraint set when the input type is text information specifically includes: The route requirement information is converted into a semantic vector through the encoding layer, and the intent type of the route requirement information is identified through the classification layer. The constraint values ​​corresponding to the intent type are extracted from the semantic vector through the padding layer; The extracted constraint values ​​are validated for validity, and after the validity validation is passed, the intent type and the extracted constraint values ​​are combined to obtain the travel constraint set.

4. The AI-based user intent recognition route planning method according to claim 1, characterized in that, The step of constructing a constraint item queue based on a travel constraint set, traversing the constraint item queue according to a preset order, and generating a corresponding supplementary query text to be sent to the user when a constraint value is missing during the traversal, specifically includes: Based on the preset constraint priority configuration, the constraints in the travel constraint set are arranged in order to construct a constraint queue; Traverse the constraint item queue according to the preset order and check whether the constraint value corresponding to each constraint item exists. When a constraint value for any constraint term is detected to be missing, supplementary query text is generated based on the type of the current constraint term and the already acquired constraint value, using a pre-trained text generation model. The generated supplementary query text is sent to the user's client.

5. The AI-based user intent recognition route planning method according to claim 1, characterized in that, The travel constraint set includes spatial constraints, duration constraints, scenario constraints, and travel mode constraints. When the travel constraint set is determined to meet the route generation conditions, the step of calling the corresponding geographic data source to perform candidate point recall based on the travel constraint set specifically includes: Determine whether all mandatory constraints in the travel constraint set have been filled with constraint values. If all mandatory constraints have been filled with constraint values, determine that the travel constraint set meets the route generation conditions. Extract the geographic coordinates corresponding to the spatial constraints from the travel constraint set, set the recall radius with the geographic coordinates as the center point, and calculate the value of the recall radius based on the duration constraint and travel mode constraint. The candidate point type is determined based on the scenario constraints, and the corresponding geographic data source is selected based on the candidate point type; Send a recall request to the selected geographic data source, the recall request including the geographic coordinates of the center point, the recall radius, and the type of point of interest; Receive a set of candidate points returned by a geographic data source. The set of candidate points includes geographic locations and their attribute information that are within the numerical range of the recall radius and conform to the candidate point type.

6. The AI-based user intent recognition route planning method according to claim 5, characterized in that, The steps of extracting the geographic coordinates corresponding to the spatial constraints from the travel constraint set, setting the recall radius with the geographic coordinates as the center point, and calculating the value of the recall radius based on the duration constraint and the travel mode constraint specifically include: Extract the origin geographic coordinates corresponding to the spatial constraints from the travel constraint set, and use the origin geographic coordinates as the recall center point; Extract the duration value corresponding to the duration constraint and the travel mode type corresponding to the travel mode constraint from the travel constraint set; Based on the mode of travel type, query the preset speed mapping table to obtain the average speed value corresponding to the mode of travel type; The recall radius is calculated based on the duration, average velocity, and preset coverage coefficient.

7. The AI-based user intent recognition route planning method according to claim 1, characterized in that, The travel constraint set includes intensity constraints and group constraints. The step of generating multiple candidate routes by filtering and sorting based on the candidate point recall results and the travel constraint set specifically includes: The candidate point recall results are filtered based on the strength constraints and group constraints in the travel constraint set to obtain the filtered point set; Each candidate point in the filter point set is scored, and the filter point set is sorted in descending order according to the scoring results. A preset number of candidate points are selected as route nodes. Based on the route nodes, a preset path planning algorithm is executed to generate multiple candidate routes.

8. An AI-based user intent recognition route planning system, characterized in that, include: The constraint extraction module is used to receive route requirement information input by the user and extract a set of travel constraints based on the route requirement information. The queue construction module is used to build a constraint item queue based on the travel constraint set, and traverse the constraint item queue according to a preset order. When a constraint value is missing during the traversal, a corresponding supplementary query text is generated and sent to the user terminal. The constraint supplementation module is used to receive supplementary information from the user based on the supplementary query text input, extract the corresponding constraint values ​​based on the supplementary information, and populate them into the constraint item queue. The recall execution module is used to recall candidate points by calling the corresponding geographic data source based on the travel constraint set when it is determined that the travel constraint set meets the route generation conditions. The route generation module is used to generate multiple candidate routes by filtering and sorting the candidate point recall results in conjunction with the travel constraint set. The route update module is used to respond to the user's editing operation on any candidate route, update the route direction and corresponding timeline of the corresponding candidate route in real time, and output the updated candidate route as the planned route.

9. A computer electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the steps of an AI-based user intent recognition route planning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of an AI-based user intent recognition route planning method as described in any one of claims 1 to 7.

Images