Traffic comprehensive query optimization method and system applied to intelligent human-computer interaction

Abstract

The application provides a traffic comprehensive query optimization method and system applied to intelligent human-computer interaction, traffic query requests input by a target user in an interaction scene are acquired, multi-modal feature analysis is performed on the traffic query requests, semantic understanding features and scene association features are generated, dynamic interaction feature fusion processing is performed on the semantic understanding features and the scene association features based on a preset language large model, multi-dimensional intention features are generated, service path optimization is performed on the traffic query requests of the target user according to the multi-dimensional intention features, and a comprehensive response strategy is generated; the comprehensive response strategy is converted into interactive feedback information, device parameters in real-time environment association information are dynamically adapted, a feedback operation is performed to complete the optimized response of the traffic query request. The application can improve the response real-time performance, decision rationality and execution reliability of intelligent traffic query services.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a traffic comprehensive query optimization method and system applied to intelligent human-computer interaction. BACKGROUND

[0002] With the deep integration of intelligent transportation systems and human-computer interaction technology, the current mainstream intelligent traffic query method usually generates traffic path suggestions by analyzing the natural language text input by the user, and relies on the pre-set database to provide fixed mode feedback information, such as path planning based on keyword matching or single-dimensional real-time road condition broadcast. However, such methods have significant limitations: on the one hand, traditional natural language processing technology can only extract explicit user requirements, and cannot dynamically understand real-time environmental parameters (such as device response delay, terminal compatibility) to cause deviations between feedback strategies and user real requirements and physical scenes; on the other hand, existing path optimization algorithms mostly focus on theoretical optimal path calculation, and lack of consideration of the adaptability of heterogeneous terminal device execution capabilities, so that the generated path scheme is easy to fail when executed on the terminal device due to protocol mismatch or resource overrun. SUMMARY

[0003] The present application provides a traffic comprehensive query optimization method and system applied to intelligent human-computer interaction.

[0004] In a first aspect, an embodiment of the present application provides a traffic comprehensive query optimization method applied to intelligent human-computer interaction, comprising:

[0005] Obtaining a traffic query request input by a target user in an interaction scene, the traffic query request comprising a natural language text and real-time environmental association information;

[0006] Performing multi-modal feature analysis on the traffic query request to generate semantic understanding features and scene association features, the scene association features being used to indicate real-time traffic environment states related to the traffic query request;

[0007] Based on a pre-set language large model, performing dynamic interaction feature fusion processing on the semantic understanding features and the scene association features to generate multi-dimensional intention features;

[0008] According to the multi-dimensional intention features, performing service path optimization on the traffic query request of the target user to generate a comprehensive response strategy;

[0009] Converting the comprehensive response strategy into interactive feedback information, and dynamically adapting device parameters in the real-time environmental association information to perform feedback operations to complete the optimized response of the traffic query request.

[0010] In a second aspect, the embodiment of the present application provides a human-computer interaction system, comprising:

[0011] a memory, wherein the memory stores a computer program;

[0012] a processor, configured to load the computer program to implement the traffic comprehensive query optimization method applied to intelligent human-computer interaction.

[0013] The traffic comprehensive query optimization method applied to intelligent human-computer interaction provided by the present application effectively overcomes the scene perception limitation of traditional traffic query systems caused by single modal data processing by integrating multi-modal feature analysis of user natural language text and real-time environment associated information to generate composite data expression with semantic understanding features and scene association features. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0015] Figure 1 is a flowchart of the traffic comprehensive query optimization method applied to intelligent human-computer interaction provided by the embodiment of the present application.

[0016] Figure 2 is a composition schematic diagram of the human-computer interaction system provided by the embodiment of the present application. DETAILED DESCRIPTION

[0017] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0018] Please refer to Figure 1 , Figure 1 The flowchart of the traffic comprehensive query optimization method applied to intelligent human-computer interaction provided by the embodiments of the present application can be executed by a human-computer interaction system. The traffic comprehensive query optimization method applied to intelligent human-computer interaction can include the following steps:

[0019] Step S100: obtaining a traffic query request input by a target user in an interaction scene, the traffic query request including a natural language text and real-time environment associated information.

[0020] In the traffic query scene of intelligent human-computer interaction, the target user is an individual who initiates the traffic query operation, and the interaction scene covers various environments in which users interact with the system, such as in the intelligent vehicle system, the traffic query application on the mobile phone, etc. The traffic query request is the content input by the user to obtain traffic-related information, which is composed of a natural language text and real-time environment associated information. The natural language text is the query content expressed by the user in daily language, such as "how to get from home to work" "where is the subway station nearby" and the like. The real-time environment associated information is data related to the current real-time environment, such as the user's current geographic location, the current time, the real-time traffic flow condition, etc.

[0021] As an implementation manner, in step S100, obtaining the traffic query request input by the target user in the interaction scene, specifically, the following steps S110-S150 can be included:

[0022] Step S110: receiving the original query information submitted by the target user through voice input, text input or touch input.

[0023] In different interaction scenarios, target users can submit raw query information in various ways. Voice input involves users conveying their queries to the system by speaking; for example, in a smart car system, a user can say, "Query the route to the airport." The system then uses speech recognition technology to convert the speech signal into text information. Speech recognition technology can be based on deep learning models, such as recurrent neural networks (RNNs) and their variants, long short-term memory networks (LSTMs) and gated recurrent units (GRUs). These models are trained on large amounts of speech data to learn the mapping relationship between speech features and corresponding text. In practical applications, the system performs preprocessing operations such as sampling and feature extraction on the user's input speech, and then inputs the processed features into the trained speech recognition model to obtain the corresponding text information.

[0024] Text input involves users directly typing text into the input box to express their query, such as manually typing "find nearby bus stops" in a mobile transportation query app. Touch input involves users touching specific areas or icons on the screen to perform queries, such as clicking the "nearby parking lot search" icon on a smart transportation query tablet. The system receives and performs preliminary formatting of these raw query information inputs from different methods to standardize subsequent processing flows.

[0025] Step S120: Perform noise filtering on the original query information to remove irrelevant characters or invalid audio segments.

[0026] The original query information may contain irrelevant characters or invalid audio segments, which can affect subsequent processing. Irrelevant characters may include punctuation marks entered incorrectly by the user, garbled characters, etc., while invalid audio segments may include background noise, the user's cough, etc. The purpose of noise filtering is to remove this interfering information and improve the quality of the query information.

[0027] For raw query information in text form, noise filtering can be achieved using regular expression matching. Regular expressions are tools for describing string patterns; by defining pattern rules, irrelevant characters can be matched and removed. For example, a regular expression pattern can be defined to match all characters that are not letters, numbers, or common punctuation marks, and then the regular expression library functions in the programming language can be used to replace these characters with an empty string from the raw query information.

[0028] For raw query information in audio format, noise filtering can employ audio processing algorithms such as spectral subtraction and Wiener filtering. Spectral subtraction is a frequency-domain-based noise estimation and removal method. It estimates the spectral characteristics of the noise and then subtracts the noise spectrum from the original audio spectrum to obtain the noise-removed audio signal. Wiener filtering, on the other hand, is an optimal linear filtering method that filters the audio signal based on the statistical characteristics of the noise and prior knowledge of the signal to achieve noise removal.

[0029] Step S130: Perform intent pre-identification on the filtered original query information to determine whether it contains traffic-related key semantic units.

[0030] Intent pre-identification involves semantic analysis of the filtered raw query information to determine whether it contains key semantic units related to transportation. Key semantic units refer to words or phrases that can clearly express the intent of a transportation query, such as "route," "bus," "subway," and "parking lot."

[0031] Intent pre-identification can employ a combination of keyword matching and semantic understanding models from natural language processing techniques. Keyword matching involves pre-defining a list of traffic-related keywords and then searching for the presence of these keywords in the filtered original query information. If they are found, it is considered that the query may contain key semantic units related to traffic. The semantic understanding model can use deep learning models, such as convolutional neural networks (CNNs) and bidirectional long short-term memory networks (Bi-LSTMs). These models can learn the semantic features of text and determine whether the input query information is relevant to the traffic query intent by encoding and classifying it. In practical applications, keyword matching is first used for initial screening, and then the filtered query information is input into a trained semantic understanding model for further evaluation.

[0032] Step S140: If the original query information contains key semantic units, mark it as a valid traffic query request and trigger subsequent processing.

[0033] When the filtered original query information is determined through intent pre-identification to contain key semantic units related to traffic, the original query information is marked as a valid traffic query request. The purpose of marking is to distinguish between valid and invalid query information so that the valid query information can be processed in a targeted manner later. Triggering the subsequent processing flow means that the system will begin to execute step S200 and subsequent steps to further analyze and process the valid traffic query request in order to provide an accurate traffic query response.

[0034] Step S150: If the key semantic unit is not included, send guidance information to the target user to re-enter or supplement the query content.

[0035] If, during the intent pre-identification process, the filtered original query information is found to lack key semantic units related to transportation, it indicates that the query may be irrelevant to transportation or poorly expressed. In this case, the system will send guidance information to the target user. This guidance information can be text prompts, such as "Your query seems unrelated to transportation; please enter information related to routes, buses, subways, etc."; or it can be voice prompts, using speech synthesis technology to convert the guidance information into speech and play it to the user. Speech synthesis technology is typically based on neural network models, such as the Tacotron series models, which can learn the mapping relationship between text and speech, converting the input text into a natural and fluent speech signal. By sending guidance information, the system guides the target user to re-enter or supplement the query content, enabling them to accurately express their transportation query intent.

[0036] Step S200: Perform multimodal feature parsing on the traffic query request to generate semantic understanding features and scene association features. The scene association features are used to indicate the real-time traffic environment status related to the traffic query request.

[0037] Multimodal feature parsing refers to the comprehensive analysis of natural language text and real-time environmental information in traffic query requests to extract key features. Semantic understanding features are used to characterize the core semantic intent of traffic query requests, reflecting the user's true query needs. Scene-related features are features related to the real-time traffic environment, indicating the current traffic conditions, such as real-time traffic flow distribution, device response latency, and the spatial distance between the user's location and the target traffic node.

[0038] Multimodal feature parsing requires the integration of natural language processing (NLP) and data analysis techniques. For natural language text, its grammatical structure and semantic meaning need to be analyzed to extract key information. For real-time environmental information, various environmental parameters need to be processed and analyzed to generate traffic environment-related features. The semantic understanding features and scene-related features generated through multimodal feature parsing provide an important foundation for subsequent feature fusion and intent analysis.

[0039] As one implementation method, step S200 involves performing multimodal feature parsing on the traffic query request to generate semantic understanding features and scene association features, which may specifically include the following steps S210~S240:

[0040] Step S210: Perform word segmentation on the natural language text to obtain multiple semantic units, and extract environmental parameters from the real-time environment association information to obtain multiple environmental state parameters.

[0041] Word segmentation divides natural language text into independent semantic units. For example, the text "How to get from home to the company" can be segmented into semantic units such as "from," "home," "to," "company," "how," and "get there." Word segmentation can be performed using various methods, such as rule-based, statistical, and deep learning-based methods. Rule-based methods segment the text according to predefined rules, such as forward maximum matching and backward maximum matching. Statistical methods segment the text by statistically analyzing large amounts of text data to learn the probability of word occurrence and contextual relationships. Deep learning-based methods use neural network models, such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs), to learn the semantic features of the text and automatically perform word segmentation.

[0042] Real-time environmental information includes various traffic environment-related data, and environmental parameter extraction involves extracting useful environmental state parameters from this data. For example, parameters such as traffic flow and speed on different road segments can be extracted from real-time traffic flow data; device response latency parameters can be extracted from device information; and the user's current location and the location of the target traffic node can be extracted from geographic location information. Environmental parameter extraction can be achieved through data mining and data analysis techniques, such as using data filtering and feature selection methods to filter out environmental state parameters relevant to traffic query requests from a large amount of real-time environmental information.

[0043] Step S220: Perform contextual semantic encoding on multiple semantic units to generate semantic understanding features; wherein, the semantic understanding features are used to characterize the core semantic intent of the traffic query request.

[0044] Contextual semantic encoding encodes multiple semantic units obtained from word segmentation, taking into account their contextual relationships to generate features that represent the core semantic intent of a traffic query request. Contextual semantic encoding can use deep learning models, such as the BERT (Bidirectional Encoder Representations from Transformers) model. The BERT model is a pre-trained language model based on the Transformer architecture, which learns the contextual information of the text through a bidirectional attention mechanism. In contextual semantic encoding, multiple semantic units are first input into the BERT model, which encodes each semantic unit, generating a corresponding word vector representation. Then, these word vectors are further processed, such as through pooling operations (average pooling, max pooling, etc.), to obtain a comprehensive semantic understanding feature vector. This feature vector reflects the core semantic intent of the traffic query request, such as whether the user is querying routes, finding bus stops, or checking traffic conditions.

[0045] Step S230: Perform spatiotemporal correlation analysis on multiple environmental state parameters to generate an environmental state coding vector.

[0046] Spatiotemporal correlation analysis (SCR) comprehensively analyzes multiple environmental state parameters, considering their temporal and spatial relationships. In traffic query scenarios, the temporal distribution and spatial topological relationships of environmental state parameters can help understand the real-time traffic environment. For example, traffic flow distribution may vary significantly across different time periods, and traffic conditions at different geographical locations can influence each other. SCR can be implemented by constructing a spatiotemporal coding network. This network consists of a temporal convolutional layer and a spatial graph attention layer. The temporal convolutional layer extracts the temporal dependencies of environmental state parameters by performing convolution operations on these parameters over time, capturing their changing patterns. The spatial graph attention layer extracts the association weights between geographical locations by representing geographical location information as a graph structure and learning the degree of association between different geographical locations through an attention mechanism.

[0047] In spatiotemporal correlation analysis, timestamp data and geographic location data are first extracted from real-time environmental correlation information. Then, the temporal distribution characteristics of environmental state parameters are determined based on the timestamp data, and the spatial topological relationships of environmental state parameters are determined based on the geographic location data. Next, the temporal distribution characteristics and spatial topological relationships are input into a spatiotemporal coding network to generate spatiotemporal correlation features. Finally, the spatiotemporal correlation features are subjected to dimensionality reduction processing to obtain the environmental state coding vector. Dimensionality reduction processing can use methods such as principal component analysis (PCA) to reduce the dimensionality of features while retaining the main information.

[0048] As one implementation method, step S230 involves performing spatiotemporal correlation analysis on multiple environmental state parameters to generate an environmental state coding vector, which may specifically include the following steps S231~S235:

[0049] Step S231: Extract timestamp data and geographic location data from real-time environmental association information.

[0050] Timestamp data records the time when environmental state parameters were collected, accurate to the specific date and time. Geographic location data indicates the location where these parameters were collected, typically expressed in latitude and longitude coordinates. Extracting timestamp and geographic location data from real-time environmental information can be achieved through data parsing and extraction methods. For example, in real-time traffic flow data, each record may contain a timestamp field and a geographic location field; by parsing these fields, the timestamp and geographic location data can be extracted.

[0051] Step S232: Determine the temporal distribution characteristics of environmental state parameters based on timestamp data.

[0052] Temporal distribution characteristics reflect the changing patterns of environmental state parameters over time. Determining the temporal distribution characteristics of environmental state parameters based on timestamp data can be achieved through statistical analysis and time series analysis. For example, the average, maximum, and minimum values ​​of environmental state parameters within different time periods can be calculated to understand the overall situation of the parameters in different time periods. Time series analysis methods, such as the ARIMA (Autoregressive Integrated Moving Average) model and seasonal decomposition methods, can also be used to model and analyze the time series of environmental state parameters and predict future trends. In practical applications, the timestamp data is first sorted chronologically, and then the environmental state parameters are grouped by time period, calculating the statistics for each time period or using time series analysis methods for modeling.

[0053] Step S233: Determine the spatial topological relationship of environmental state parameters based on geographical location data.

[0054] Spatial topology describes the relative positions and connections between different geographic locations. Determining the spatial topology of environmental state parameters based on geographic location data can be achieved using Geographic Information System (GIS) technology and graph theory methods. First, geographic location data is converted into points in geospatial space. Then, a graph structure is constructed based on the distances and connections between these points. In the graph structure, each node represents a geographic location, and edges represent the connections between nodes. Graph theory algorithms, such as shortest path algorithms and centrality analysis algorithms, can be used to analyze the spatial distribution and propagation patterns of environmental state parameters. For example, Dijkstra's algorithm can be used to calculate the shortest path between different geographic locations to understand the propagation of traffic flow across different road segments. In practical applications, GIS software or open-source geographic information processing libraries, such as GeoPandas and NetworkX, can be used to process and analyze geographic location data to determine the spatial topology of environmental state parameters.

[0055] Step S234: Input the temporal distribution features and spatial topological relationships into the spatiotemporal coding network to generate spatiotemporal correlation features.

[0056] Spatiotemporal coding networks are neural network models used to process spatiotemporal data. They can fuse temporal and spatial information to generate spatiotemporal correlation features. Temporal distribution features and spatial topological relationships are input into the spatiotemporal coding network. The input data is first preprocessed to meet the network's input requirements. For example, the temporal distribution features and spatial topological relationships are converted into a suitable tensor format and normalized to improve the network's training performance.

[0057] The temporal convolutional layer of the spatiotemporal coding network performs convolution operations on temporal distribution features to extract temporal series dependencies. The temporal convolutional layer can use a one-dimensional convolutional kernel that slides along the time dimension to extract features from the temporal distribution. The spatial graph attention layer processes spatial topological relationships, learning the association weights between different geographical locations through an attention mechanism. The spatial graph attention layer can represent geographical location information as a graph structure and use a graph attention mechanism to update the features of nodes in the graph. Finally, the outputs of the temporal convolutional layer and the spatial graph attention layer are fused to generate spatiotemporal correlation features.

[0058] Step S235: Perform dimensionality reduction on the spatiotemporal correlation features to obtain the environmental state encoding vector; wherein, the spatiotemporal encoding network includes a temporal convolutional layer and a spatial graph attention layer, the temporal convolutional layer is used to extract time series dependencies, and the spatial graph attention layer is used to extract the association weights between geographical locations.

[0059] Dimensionality reduction aims to reduce the dimensionality of spatiotemporal correlation features while retaining key information, thereby improving the efficiency and accuracy of subsequent processing. Principal Component Analysis (PCA) can be used to reduce the dimensionality of spatiotemporal correlation features. PCA projects high-dimensional data into a low-dimensional space by identifying the principal components. The specific steps are as follows: First, the spatiotemporal correlation features are centered to have a mean of zero. Then, the covariance matrix of the spatiotemporal correlation features is calculated. Next, the eigenvalues ​​and eigenvectors of the covariance matrix are solved, and the k eigenvectors with the largest eigenvalues ​​are selected as principal components. Finally, the spatiotemporal correlation features are projected onto these k principal components to obtain the dimensionality-reduced environment state encoding vector. Dimensionality reduction transforms high-dimensional spatiotemporal correlation features into low-dimensional environment state encoding vectors, facilitating subsequent feature fusion and intent analysis.

[0060] Step S240: Construct scene association features based on the device type parameters in the environment state coding vector and the real-time environment association information.

[0061] Scene association features are used to indicate the real-time traffic environment status related to traffic query requests. They include real-time traffic flow distribution parameters, device response latency parameters, and spatial distance parameters between the user's location and the target traffic node. Constructing scene association features based on the environment state encoding vector and the device type parameter in the real-time environment association information requires comprehensive consideration of both environmental status and device characteristics.

[0062] First, the environmental state encoding vector already reflects the spatiotemporal correlation information of environmental state parameters, and is used as part of the scene association features. Then, combined with the device type parameter from the real-time environmental association information, the impact of different device types on traffic queries is determined. For example, different types of devices may have different response latency characteristics; the corresponding device response latency parameter can be obtained through the device type parameter. The spatial distance parameter between the user's location and the target traffic node can be calculated using the geographic location data from the real-time environmental association information. Finally, the environmental state encoding vector, device response latency parameter, and spatial distance parameter between the user's location and the target traffic node are combined and weighted to construct the scene association features. In practical applications, the weights of each parameter can be adjusted according to different traffic query scenarios and needs to obtain more accurate scene association features.

[0063] Step S300: Based on the preset language big model, perform dynamic interaction feature fusion processing on semantic understanding features and scene association features to generate multi-dimensional intent features.

[0064] The pre-defined language model is a pre-trained natural language processing model with powerful semantic understanding and feature representation capabilities. Dynamic interaction feature fusion processing refers to fusing semantic understanding features and scene-related features based on the dynamic relationship between them, generating more comprehensive and accurate multi-dimensional intent features. These multi-dimensional intent features include explicit user needs, implicit scene adaptation features, and real-time resource constraint features, which can more comprehensively reflect the user's traffic query intent.

[0065] When performing dynamic interaction feature fusion, semantic understanding features and scene-related features are effectively combined, taking into account their interaction in different scenarios. The pre-defined large language model can provide a powerful feature fusion framework, which automatically adjusts the weights of semantic understanding features and scene-related features through the model's learning capabilities to achieve dynamic interaction feature fusion.

[0066] As one implementation method, step S300 involves dynamically fusing semantic understanding features and scene association features based on a preset language model to generate multi-dimensional intent features. This may specifically include the following steps S310-S360:

[0067] Step S310: Map the semantic understanding features to the first feature vector and the scene association features to the second feature vector.

[0068] Mapping is the process of converting semantic understanding features and scene-related features into vector representations. Semantic understanding features are obtained by encoding the contextual semantics of natural language text, and can be a complex feature representation. Mapping them to a first feature vector can be achieved through linear or non-linear transformations. For example, a fully connected layer can be used to linearly map the semantic understanding features, resulting in a first feature vector of fixed dimensions.

[0069] Scene-related features are constructed based on environment state encoding vectors and device type parameters, and they also need to be mapped to a second feature vector. The mapping process is similar to that of semantic understanding features and can be performed using fully connected layers or other neural network layers. Mapping semantic understanding features and scene-related features to vector representations facilitates subsequent feature fusion and computation.

[0070] Step S320: Perform dynamic weight allocation processing on the first feature vector and the second feature vector to determine the semantic weight coefficient and the environmental weight coefficient.

[0071] Dynamic weight allocation is used to dynamically adjust the weights of semantic understanding features and contextual features in feature fusion based on different traffic query scenarios and needs. The semantic weight coefficient represents the importance of semantic understanding features in feature fusion, while the contextual weight coefficient represents the importance of contextual features in feature fusion.

[0072] When performing dynamic weight allocation, historical interaction data and the current query context need to be considered. First, the set of sample semantic feature vectors and the set of sample environmental feature vectors are obtained from the historical interaction dataset. Each sample semantic feature vector is associated with a weight value representing the influence of semantic features in the actual decision-making process, and each sample environmental feature vector is associated with a weight value representing the influence of environmental features in the actual decision-making process. Then, a weight allocation model is trained using this historical data to learn the weight allocation patterns of semantic and environmental features under different circumstances.

[0073] As one implementation method, step S320 involves performing dynamic weight allocation processing on the first feature vector and the second feature vector to determine the semantic weight coefficient and the environment weight coefficient. Specifically, this may include the following steps S321~S3210:

[0074] Step S321: Obtain the set of sample semantic feature vectors and the set of sample environmental feature vectors from the historical interaction dataset. Each sample semantic feature vector is associated with the influence weight value of semantic features in the actual decision-making process, and each sample environmental feature vector is associated with the influence weight value of environmental features in the actual decision-making process.

[0075] The historical interaction dataset is data recorded by the system during past traffic query interactions. It contains a large number of sample semantic feature vectors and sample environment feature vectors. The sample semantic feature vectors are feature vectors obtained by semantically encoding the natural language text in the historical queries, while the sample environment feature vectors are feature vectors obtained by processing the real-time environmental association information in the historical queries.

[0076] Each sample semantic feature vector and sample environment feature vector is associated with the influence weights of semantic and environmental features in the actual decision-making process. These influence weights are determined based on the actual results of historical queries and user feedback, reflecting the importance of semantic and environmental features to the final decision in different traffic query scenarios. For example, in some cases, the user's query intent is mainly determined by semantic information, in which case the influence weight of semantic features is higher; while in other cases, the real-time traffic environment has a greater impact on the query results, in which case the influence weight of environmental features is higher. Obtaining the set of sample semantic feature vectors and sample environment feature vectors from the historical interaction dataset can be achieved through a data storage and management system, such as using a database system to store historical interaction data, and then extracting the required sample feature vectors and influence weights through query statements.

[0077] Step S322: Input the sample semantic feature vector and sample environment feature vector into the initial weight prediction model to generate the corresponding semantic weight prediction sequence and environment weight prediction sequence.

[0078] The initial weight prediction model is a model used to predict semantic weights and context weights. It can use a neural network model, such as a multilayer perceptron (MLP). A multilayer perceptron is a feedforward neural network consisting of an input layer, hidden layers, and an output layer. In the input layer, the sample semantic feature vector and the sample context feature vector are used as input data. The hidden layer can contain multiple neurons that perform feature transformation on the input data using a non-linear activation function (such as the ReLU function). The output layer outputs the semantic weight prediction sequence and the context weight prediction sequence.

[0079] When training the initial weight prediction model, the sample semantic feature vector and the sample environment feature vector need to be normalized to ensure they have the same scale and range. Then, the normalized sample feature vectors are input into the initial weight prediction model, and the semantic weight prediction sequence and environment weight prediction sequence are obtained through forward propagation. These prediction sequences represent the model's weight prediction results for semantic features and environment features under different samples.

[0080] Step S323: Calculate the error loss for each predicted value in the semantic weight prediction sequence and its associated actual influence weight value to generate a semantic weight error set.

[0081] Error loss calculation measures the difference between each predicted value and the actual weight value in the semantic weight prediction sequence. Mean squared error (MSE) can be used as the error loss function, calculated as follows: ,in This is the actual impact weight value. is the predicted value in the semantic weight prediction sequence, and n is the number of samples.

[0082] In practical applications, each predicted value in the semantic weight prediction sequence is iterated and compared with its corresponding actual impact weight value to calculate the error loss. The error loss values ​​of all samples are collected to generate a semantic weight error set. The semantic weight error set reflects the accuracy of the initial weight prediction model in predicting semantic weights; the smaller the error value, the more accurate the model's prediction.

[0083] Step S324: Calculate the error loss of each predicted value in the environmental weight prediction sequence and its associated actual impact weight value to generate an environmental weight error set.

[0084] Similar to the semantic weight error calculation, the error loss calculation for each predicted value in the environmental weight prediction sequence and its associated actual impact weight value can also use the mean squared error (MSE) as the error loss function. Each predicted value in the environmental weight prediction sequence is iterated over, compared with its corresponding actual impact weight value, and the error loss is calculated. The error loss values ​​of all samples are collected to generate an environmental weight error set. The environmental weight error set reflects the accuracy of the initial weight prediction model in predicting environmental weights; the smaller the error value, the more accurate the model's prediction.

[0085] Step S325: The semantic weight error set and the environmental weight error set are weighted and fused to generate a total error loss value. The parameters of the initial weight prediction model are iteratively adjusted according to the total error loss value until the total error loss value converges to a stable range, thus generating a trained weight allocation model.

[0086] Weighted fusion combines the semantic weight error set and the environment weight error set, generating a total error loss value through weighted summation. A weight coefficient can be assigned to each of the semantic and environment weight error sets. Then, their respective error losses are multiplied by their respective weight coefficients and summed to obtain the total error loss value. For example, let the weight coefficient of the semantic weight error set be... The weighting coefficients of the environmental weight error set are Total error loss value ,in These are the error values ​​in the semantic weight error set. It is the number of samples in the semantic weight error set. These are the error values ​​in the environmental weight error set. It is the number of samples in the environmental weight error set.

[0087] The gradient descent algorithm can be used to iteratively adjust the parameters of the initial weight prediction model based on the total error loss value. Gradient descent is an optimization algorithm that calculates the gradient of the total error loss value with respect to the model parameters, and then updates the model parameters in the opposite direction of the gradient, gradually reducing the total error loss value. In each iteration, the gradient of the total error loss value is calculated, and the model parameters are updated according to the learning rate. This process is repeated until the total error loss value converges to a stable region, at which point the trained weight allocation model is generated.

[0088] Step S326: Input the first feature vector and the second feature vector into the trained weight allocation model, and output the initial semantic weight value and the initial environment weight value before normalization.

[0089] The first and second feature vectors are input into the trained weight allocation model. The model processes the first and second feature vectors according to the learned weight allocation rules, outputting the initial semantic weight values ​​and initial environment weight values ​​before normalization. The trained weight allocation model has learned the weight allocation rules of semantic and environment features in historical interaction data and can predict their weights in feature fusion based on the current first and second feature vectors.

[0090] Step S327: Perform dynamic balancing constraint processing on the initial semantic weight value and the initial environment weight value so that the sum of the two satisfies the preset constant relationship, and generate intermediate semantic weight value and intermediate environment weight value.

[0091] Dynamic balancing constraints are used to ensure a balance between semantic weights and environmental weights. Preset constants can be set according to specific traffic query scenarios and requirements; for example, their sum could be set to 1. Dynamic balancing constraints on the initial semantic weights and initial environmental weights can be achieved through normalization. For example, let the initial semantic weight be... The initial environmental weight value is Then the intermediate semantic weight value intermediate environment weight value By using dynamic balance constraints, the sum of intermediate semantic weights and intermediate environmental weights is made to satisfy a preset constant relationship, thus ensuring a balance in the relative importance of semantic features and environmental features in feature fusion.

[0092] Step S328: Perform nonlinear activation transformation on the intermediate semantic weight values ​​to eliminate abrupt fluctuations in the weight values ​​and generate smooth semantic weight coefficients.

[0093] Nonlinear activation transformation can be performed using functions such as the Sigmoid or Tanh. Taking the Sigmoid function as an example, when intermediate semantic weight values ​​are input into the Sigmoid function, the function performs a nonlinear transformation on these values, mapping them to the range of 0 to 1. This nonlinear activation transformation can eliminate abrupt fluctuations in weight values, resulting in smoother weight values. For instance, when intermediate semantic weight values ​​exhibit large fluctuations, the Sigmoid function can compress them into a relatively stable range, generating smooth semantic weight coefficients.

[0094] Step S329: Perform nonlinear activation transformation on the intermediate environment weight values ​​to eliminate abrupt fluctuations in the weight values ​​and generate smooth environment weight coefficients.

[0095] Similar to the processing of intermediate semantic weights, the intermediate environment weights are also processed through non-linear activation transformations, using functions such as the Sigmoid or Tanh. The intermediate environment weights are input into the non-linear activation function, which performs a non-linear transformation, eliminating abrupt fluctuations in the weight values ​​and generating smooth environment weight coefficients. This smoothing process makes the environment weight coefficients more stable, avoiding the impact of drastic changes in weight values ​​on the feature fusion results.

[0096] Step S3210: Perform joint verification processing on the smoothed semantic weight coefficients and smoothed environment weight coefficients to ensure that the interaction contribution of the two is matched with the first feature vector and the second feature vector, and generate the final usable semantic weight coefficients and environment weight coefficients.

[0097] Joint verification is performed to ensure the rationality and effectiveness of the smoothed semantic weight coefficients and smoothed environment weight coefficients in feature fusion. Verification can be performed by calculating the interaction contribution of the smoothed semantic weight coefficients and smoothed environment weight coefficients with the first and second feature vectors. The interaction contribution can be measured by calculating the sum of the products of the feature vectors and the weight coefficients. For example, calculate the sum of the products of the first feature vector and the smoothed semantic weight coefficients, and the sum of the products of the second feature vector and the smoothed environment weight coefficients, and compare their magnitudes and proportions. If the interaction contributions are mismatched, the smoothed semantic weight coefficients and smoothed environment weight coefficients need to be adjusted. Through joint verification, the final usable semantic weight coefficients and environment weight coefficients are generated, ensuring the reasonable contribution of semantic features and environment features in feature fusion.

[0098] Step S330: Adjust the first feature vector according to the semantic weight coefficient to obtain the adjusted semantic feature vector.

[0099] Weighted adjustment involves performing element-wise multiplication of the semantic weight coefficients with the first feature vector. Through weighted adjustment, the importance of each element in the first feature vector is adjusted according to the magnitude of the semantic weight coefficients. If the semantic weight coefficients are large, the elements in the first feature vector have a larger proportion in the adjusted semantic feature vector; conversely, if the semantic weight coefficients are small, the elements in the first feature vector have a smaller proportion in the adjusted semantic feature vector.

[0100] Step S340: Adjust the second feature vector according to the environmental weight coefficient to obtain the adjusted environmental feature vector.

[0101] Similar to the weighted adjustment of the first feature vector, the second feature vector is also weighted according to the environmental weight coefficients. This weighted adjustment adjusts the importance of each element in the second feature vector based on the magnitude of the environmental weight coefficients. If the environmental weight coefficients are large, the elements in the second feature vector have a larger proportion in the adjusted environmental feature vector; conversely, if the environmental weight coefficients are small, the elements in the second feature vector have a smaller proportion in the adjusted environmental feature vector.

[0102] Step S350: Concatenate the adjusted semantic feature vector with the adjusted environmental feature vector to generate the initial fused feature.

[0103] Concatenation involves joining the adjusted semantic feature vector and the adjusted environmental feature vector sequentially to generate a longer vector. Through this concatenation operation, semantic and environmental features are fused to obtain an initial fused feature containing more information. This initial fused feature integrates information from semantic understanding features and scene-related features, providing a foundation for subsequent intent dimension expansion.

[0104] Step S360: Call the intent recognition layer in the language big model to expand the intent dimension of the initial fusion features and generate multi-dimensional intent features; among which, the multi-dimensional intent features include user explicit demand features, implicit scene adaptation features and real-time resource constraint features.

[0105] The intent recognition layer analyzes and processes the input features to identify the intent information within them. Expanding the intent dimension of the initial fused features aims to extract more intent information from them, generating more comprehensive multi-dimensional intent features.

[0106] Multi-dimensional intent features include explicit user demand features, implicit scenario adaptation features, and real-time resource constraint features. Explicit user demand features are demand information directly extracted from the user's query text, such as explicitly stated requests like route queries or bus stop searches. Implicit scenario adaptation features are inferred from the real-time traffic environment and the user's query, suggesting potentially suitable travel options, such as recommending more appropriate modes of transportation in congested traffic conditions. Real-time resource constraint features consider current traffic resource limitations, such as road capacity and public transportation capacity, to constrain and optimize the user's query.

[0107] In this embodiment of the invention, the intent recognition layer can be specifically implemented as including a semantic analysis sublayer, a scene reasoning sublayer, a resource evaluation sublayer, and a feature integration sublayer.

[0108] The main task of the semantic analysis sublayer is to deeply analyze the semantic information in the initial fused features and extract explicit user demand features. It can employ a multi-head attention mechanism based on the Transformer architecture. The Transformer architecture has powerful parallel computing capabilities and long sequence processing capabilities, enabling it to capture semantic relationships at different locations in the text.

[0109] Multi-head attention mechanisms allow models to focus on different parts of an input sequence in parallel across different representation subspaces. Specifically, the initial fused features of the input are first linearly projected onto multiple low-dimensional subspaces, forming multiple "heads." Each head independently computes an attention score, determining the importance of each position by calculating the similarity between the query, key, and value. These attention scores are then weighted and summed to obtain the output of each head. Finally, the outputs of all heads are concatenated and subjected to a linear transformation to obtain the final semantic analysis result.

[0110] For example, when a user queries "How do I get from home to the mall by subway at 10 am tomorrow", the semantic analysis sublayer will use a multi-head attention mechanism to focus on "10 am tomorrow" to determine the travel time, "from home to the mall" to determine the starting point and destination, and "take the subway" to determine the mode of transportation, thereby accurately extracting the user's explicit demand features.

[0111] The scene inference sublayer, based on scene association information from the initial fused features, combines real-time traffic data and predefined scene rules to perform dynamic scene adaptation analysis and generate implicit scene adaptation features. This sublayer can use a graph neural network (GNN) to process the spatial topology information in the traffic network. A graph neural network can represent a traffic network as a graph, where nodes represent traffic nodes (such as bus stops, subway stations, intersections, etc.), and edges represent the connections between nodes (such as roads, routes, etc.). Through a message passing mechanism, nodes can receive and aggregate information from their neighboring nodes, thereby updating their own feature representation.

[0112] Meanwhile, the scenario reasoning sublayer also incorporates time series analysis methods to consider changes in traffic conditions over time. For example, it uses Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs) to process time series data of traffic flow and predict traffic conditions over a future period.

[0113] In practical applications, the scenario inference sublayer infers the dynamic constraints of traffic resources on the user's current scenario based on information such as current traffic flow, weather conditions, and special events. For example, if it is a weekday peak hour and the queried route passes through a busy commercial area, the scenario inference sublayer may infer that the road segment is congested, thereby generating implicit scenario adaptation features that suggest the user choose alternative routes or modes of transportation.

[0114] The resource assessment sublayer is responsible for evaluating traffic resources and determining real-time resource constraints. It combines real-time traffic databases and machine learning models to quantitatively analyze the availability and cost of using traffic resources.

[0115] The real-time traffic database contains information such as road capacity, bus and subway capacity, and the number of available parking spaces. The resource assessment sublayer queries these databases in real time to obtain the latest resource status. The resource assessment sublayer can use Support Vector Regression (SVR) to predict resource consumption and demand. SVR finds an optimal hyperplane that minimizes the error between the predicted and actual values. In resource assessment, it can predict the resource consumption of different travel options, such as time cost, energy consumption, and expenses, based on historical data and current traffic conditions.

[0116] The feature integration sublayer integrates the explicit user demand features extracted by the semantic analysis sublayer, the implicit scene adaptation features generated by the scene reasoning sublayer, and the real-time resource constraint features determined by the resource evaluation sublayer to generate the final multi-dimensional intent features. This sublayer uses residual connections and layer normalization techniques to ensure feature stability and effective information transfer. Residual connections allow the model to skip some layers during training and directly pass the input features to subsequent layers, avoiding the gradient vanishing problem and enabling the model to learn more complex feature representations. Layer normalization normalizes the input of each layer, ensuring that the features have a similar distribution across different dimensions, improving the model's training efficiency and generalization ability.

[0117] During feature integration, the feature integration sublayer performs weighted fusion of different features. The weights are determined based on statistical analysis of historical data and the model's training results. For example, for users with high requirements for travel time, the weight of time-related features in the explicit user demand feature will be relatively high; while for users who are sensitive to cost, the weight of cost-related features in the real-time resource constraint feature will be greater.

[0118] Ultimately, the feature integration sublayer outputs multi-dimensional intent features that include explicit requirements, implicit scenario adaptations, and real-time resource constraints, providing comprehensive and accurate information for subsequent service path optimization.

[0119] As one implementation method, step S360 involves calling the intent recognition layer in the large language model to expand the intent dimension of the initial fused features and generate multi-dimensional intent features. Specifically, this may include the following steps S361~S366:

[0120] Step S361: Decompose the initial fused features into a semantic-dominant feature branch and an environment-dominant feature branch; wherein, the semantic-dominant feature branch consists of features after channel attention filtering of the adjusted semantic feature vector, and the environment-dominant feature branch consists of features after spatial attention filtering of the adjusted environment feature vector.

[0121] Channel attention filtering processes the adjusted semantic feature vector, selecting the feature channels that are more important for semantic understanding through a channel attention mechanism. This mechanism can utilize channel attention modules in convolutional neural networks, such as the Squeeze-and-Excitation (SE) module. The SE module compresses the adjusted semantic feature vector spatially using global average pooling, obtaining a global feature representation for each channel. Then, a fully connected layer and a non-linear activation function (such as the sigmoid function) are used to process the global feature representation, yielding attention weights for each channel. Finally, element-wise multiplication is performed between the attention weights and the adjusted semantic feature vector to select the feature channels most important for semantic understanding, forming the semantically dominant feature branch.

[0122] Spatial attention filtering processes the adjusted environmental feature vector and selects the feature spaces that are more important for understanding the environment through a spatial attention mechanism. This spatial attention mechanism can utilize the spatial attention module in a convolutional neural network, such as the SpatialAttentionModule (SAM). The SAM module processes the adjusted environmental feature vector through convolution operations to obtain a spatial attention map. Then, element-wise multiplication is performed between the spatial attention map and the adjusted environmental feature vector to select the feature spaces that are more important for understanding the environment, forming the dominant environmental feature branch.

[0123] Step S362: Perform contextual demand correlation analysis on the semantic dominant feature branch to extract explicit user demand features; wherein, the contextual demand correlation analysis is achieved by traversing the semantic units in the semantic dominant feature branch and matching them with the predefined set of traffic service keywords.

[0124] Contextual demand correlation analysis analyzes semantic units within the semantically dominant feature branch to identify semantic units that match a predefined set of transportation service keywords, thereby extracting explicit user demand features. The predefined set of transportation service keywords is a predefined group of keywords related to transportation services, such as "route," "bus," "subway," and "parking lot."

[0125] During contextual demand correlation analysis, each semantic unit in the semantically dominant feature branch is traversed and matched against a predefined set of transportation service keywords. If a semantic unit matches a keyword in the keyword set, it is considered relevant to transportation service demand. By comprehensively analyzing all matching semantic units, explicit user demand features are extracted. For example, if the semantically dominant feature branch contains "route from home to work," the explicit user demand feature of "route query" is extracted by matching it against the predefined set of transportation service keywords.

[0126] Step S363: Perform dynamic scene adaptation analysis on the dominant environmental feature branch to generate implicit scene adaptation features; wherein, the dynamic scene adaptation analysis includes: determining the dynamic constraints of traffic resources on the scene in which the target user is located based on the real-time traffic state parameters in the dominant environmental feature branch.

[0127] Dynamic scene adaptation analysis analyzes the scene in which the target user is located based on real-time traffic state parameters in the dominant environmental feature branch, determines the dynamic constraints of this scene on traffic resources, and thus generates implicit scene adaptation features. Real-time traffic state parameters include real-time traffic flow, road congestion, and public transportation operation status.

[0128] When performing dynamic scenario adaptation analysis, the first step is to determine the target user's scenario based on real-time traffic status parameters, such as whether it's peak or off-peak traffic, or whether the road is congested or unobstructed. Then, based on different scenarios, dynamic constraints on traffic resources are determined. For example, during peak traffic hours, road capacity may be limited, and public transportation capacity may be insufficient; in congested areas, alternative routes may be necessary. By analyzing these dynamic constraints, implicit scenario adaptation features are generated. For instance, if real-time traffic status parameters indicate congestion on a certain road segment, the generated implicit scenario adaptation feature might suggest alternative routes or modes of transportation to the user.

[0129] Step S364: Input the user's explicit demand features and implicit scene adaptation features into the cross-modal interaction network to generate potential demand compensation features; wherein, the cross-modal interaction network achieves feature compensation by alternately performing the projection mapping from semantic features to environmental features and the feedback correction from environmental features to semantic features.

[0130] Cross-modal interaction networks are neural network models used to handle interactions between features of different modalities. User explicit demand features and implicit scene adaptation features are input into the cross-modal interaction network. By alternately performing the projection mapping from semantic features to environmental features and the feedback correction from environmental features to semantic features, feature compensation is achieved, generating latent demand compensation features.

[0131] The projection mapping from semantic features to environmental features projects explicit user needs into the environmental feature space, enabling the environmental features to better understand the user's needs. The feedback correction from environmental features to semantic features modifies the explicit user needs based on the environmental features, taking into account the impact of the real-time traffic environment on user needs. Cross-modal interaction networks can be implemented using bidirectional recurrent neural networks (Bi-RNN) or Transformer architectures. During network training, the network parameters are continuously optimized by alternately executing projection mapping and feedback correction operations, ensuring that the generated latent demand compensation features better supplement the deficiencies of explicit user needs features and implicit scene adaptation features.

[0132] Step S365: Perform orthogonalization processing on the explicit user requirement features, implicit scenario adaptation features, and potential requirement compensation features to generate an orthogonal requirement feature set after eliminating redundant information.

[0133] Feature orthogonalization aims to eliminate redundant information among explicit user requirement features, implicit scenario adaptation features, and potential requirement compensation features, making these features independent of each other. The Gram-Schmidt orthogonalization method can be used for feature orthogonalization. The Gram-Schmidt orthogonalization method transforms a set of linearly independent vectors into a set of orthogonal vectors. The specific steps are as follows: First, consider the explicit user requirement features, implicit scenario adaptation features, and potential requirement compensation features as a set of vectors. Then, select one vector as the initial orthogonal vector. Next, orthogonalize the other vectors sequentially by subtracting the projection of the newly generated vector onto the already orthogonalized vectors, making the newly generated vector orthogonal to the already orthogonalized vectors. Repeat this process until all vectors are orthogonalized. Finally, a set of orthogonal requirement features is obtained. Feature orthogonalization eliminates redundant information between features, improving their independence and effectiveness.

[0134] Step S366: Based on each feature vector in the orthogonal demand feature set, construct a multi-dimensional demand distribution matrix, and extend the feature dimension of the multi-dimensional demand distribution matrix to generate multi-dimensional intent features that include explicit demands, implicit scene adaptation, and real-time resource constraints; wherein, the feature dimension extension is achieved by tensor concatenating the multi-dimensional demand distribution matrix with the pre-trained resource constraint embedding vector.

[0135] A multidimensional demand distribution matrix is ​​a matrix composed of eigenvectors from an orthogonal demand feature set. It can intuitively represent the distribution of users' multidimensional demands. When constructing a multidimensional demand distribution matrix, each eigenvector from the orthogonal demand feature set is used as a row or column of the matrix, forming a single matrix.

[0136] Feature dimension expansion aims to incorporate real-time resource constraint information into the multi-dimensional demand distribution matrix, generating multi-dimensional intent features that include explicit demands, implicit scene adaptations, and real-time resource constraints. The pre-trained resource constraint embedding vector is a pre-trained vector representing real-time traffic resource constraints, such as road capacity and public transport capacity. By tensor concatenating the multi-dimensional demand distribution matrix with the pre-trained resource constraint embedding vector, the matrix's dimensions are expanded, generating multi-dimensional intent features containing more information. For example, the pre-trained resource constraint embedding vector can be used as a column or row of the matrix and concatenated with the multi-dimensional demand distribution matrix to obtain the expanded multi-dimensional intent features.

[0137] Step S400: Based on multi-dimensional intent features, optimize the service path for the target user's traffic query request and generate a comprehensive response strategy.

[0138] Service path optimization selects the optimal transportation service path based on multi-dimensional intent characteristics, considering explicit user needs, implicit scenario adaptation, and real-time resource constraints, to meet users' transportation query requirements. The comprehensive response strategy is a response plan generated based on the optimized service path and other user needs, including path guidance information and resource adaptation strategies.

[0139] Service route optimization requires consideration of multiple factors, such as traffic flow, road conditions, travel time, and resource consumption. Multi-dimensional intent features provide rich information, including user needs and real-time traffic conditions, offering a basis for service route optimization. Through analysis and processing of these multi-dimensional intent features, the optimal service route is selected, and a comprehensive response strategy is generated based on that route.

[0140] As one implementation method, step S400 involves optimizing the service path for the target user's traffic query request based on multi-dimensional intent characteristics to generate a comprehensive response strategy. This may specifically include the following steps S410-S460:

[0141] Step S410: Decompose the service requirements of the multi-dimensional intent features to obtain the core service requirement features and auxiliary service requirement features.

[0142] Service demand decomposition involves breaking down multi-dimensional intent features to extract core service demand features and auxiliary service demand features. Core service demand features are those directly related to the user's primary transportation query needs, such as the user's route origin and destination, and mode of transportation. Auxiliary service demand features are other auxiliary demand features related to the core service demands, such as the user's requirements for travel time and comfort.

[0143] When decomposing service requirements, analysis can be performed based on the different dimensions and semantic information of multi-dimensional intent features. For example, features related to route planning can be extracted as core service requirement features, while features related to travel time and comfort can be extracted as auxiliary service requirement features. By decomposing service requirements, complex multi-dimensional intent features are broken down into more easily processed core and auxiliary service requirement features, providing a foundation for subsequent service path optimization.

[0144] Step S420: Based on the real-time traffic database, obtain an initial set of service paths that match the core service demand characteristics.

[0145] A real-time traffic database is a database that stores real-time traffic information, including road network information, traffic flow information, and public transportation operation information. Based on the real-time traffic database, and according to the characteristics of core service needs, a set of initial service paths that match them is found.

[0146] For example, if the core service requirement is a route query from point A to point B, all possible routes from point A to point B can be found using road network information and traffic flow information in a real-time traffic database, forming an initial set of service paths. In practical applications, graph search algorithms, such as Dijkstra's algorithm and A* algorithm, can be used to search for paths that satisfy the core service requirement in the road network of the real-time traffic database.

[0147] Step S430: Perform constraint matching on each service path in the initial service path set to determine the resource consumption parameters and path adaptation score corresponding to each service path.

[0148] Constraint matching involves matching each service path in the initial service path set with constraints related to the real-time traffic environment and user needs, evaluating the feasibility and applicability of each service path. Resource consumption parameters refer to the resources consumed by each service path during execution, such as time, energy, and costs. Path adaptation score assesses the degree of matching between each service path and user needs and the real-time traffic environment. A higher score indicates that the service path better meets user needs and the real-time traffic environment. In one implementation, for resource consumption parameters, energy consumption for each path can be calculated based on traffic flow and road conditions of different road segments; the estimated travel time for each path can be estimated based on real-time traffic conditions as time consumption; and if toll roads exist, their costs can be calculated as cost consumption. When calculating the path adaptation score, quantitative standards can be set for each consideration factor of user needs and the real-time traffic environment. For the travel time dimension, the absolute value of the difference between the estimated travel time of the service route and the user's expected travel time is calculated. This difference is then divided into intervals: 0-5 minutes, 8-10 points; 5-10 minutes, 5-7 points; and over 10 minutes, 1-4 points. For the travel mode preference dimension, a perfect match to the user's preferences earns 8-10 points; partial match, 4-7 points; and no match, 1-3 points. For the comfort dimension, levels are assigned based on factors such as the number of transfers and road conditions. For the traffic environment dimension, levels are assigned based on traffic flow and congestion levels. Based on actual needs, corresponding weights are assigned to each dimension. The route suitability score is obtained by multiplying each dimension's score by its corresponding weight and then summing the results.

[0149] When performing constraint matching, multiple factors need to be considered, such as traffic flow, road conditions, travel time, and resource consumption. For example, if a service path passes through congested sections, its travel time will increase, resource consumption will increase accordingly, and the path fit score will decrease. Resource consumption parameters and path fit scores for each service path can be calculated using real-time traffic flow and road condition information from a traffic database.

[0150] Step S440: Construct a dynamic optimization objective function based on the characteristics of ancillary service demand; wherein, the dynamic optimization objective function is used to balance the comprehensive optimization weights of resource consumption parameters and path adaptation scores.

[0151] The dynamic optimization objective function is a function constructed based on the characteristics of ancillary service demand. It is used to balance the comprehensive optimization weights of resource consumption parameters and path suitability scores. The characteristics of ancillary service demand reflect users' requirements for travel time, comfort, etc. By using the dynamic optimization objective function, the weights of resource consumption parameters and path suitability scores can be adjusted according to different user needs.

[0152] For example, if users have high requirements for travel time, the dynamic optimization objective function can increase the weight of the time factor in the resource consumption parameter and decrease the weight of other factors in the path suitability score, in order to prioritize the service path with the shortest travel time. The dynamic optimization objective function can be expressed as: F = w1R + w2S, where F is the value of the dynamic optimization objective function, R is the resource consumption parameter, S is the path suitability score, and w1 and w2 are weight coefficients that are dynamically adjusted according to the characteristics of ancillary service demand.

[0153] Step S450: Invoke the path optimization algorithm to iteratively filter the initial service path set based on the dynamic optimization objective function and generate the optimal service path.

[0154] Path optimization algorithms are used to select the optimal path from multiple service paths. They iteratively filter an initial set of service paths based on a dynamic optimization objective function, continuously comparing the dynamic optimization objective function values ​​of each service path to select the optimal one.

[0155] As one implementation method, the path optimization algorithm performs iterative selection through the following steps S451~S455:

[0156] Step S451: Initialize the path filtering queue and add all service paths from the initial service path set to the path filtering queue.

[0157] The path filtering queue is a queue used to store service paths to be filtered. During initialization, all service paths from the initial service path set are added to the queue sequentially. In subsequent iterative filtering processes, service paths are retrieved from the queue for processing.

[0158] Step S452: Extract the current path from the path filtering queue and calculate the objective function value corresponding to the current path.

[0159] A service path is retrieved from the path filtering queue as the current path, and then the objective function value corresponding to this path is calculated based on the dynamic optimization objective function. The dynamic optimization objective function comprehensively considers resource consumption parameters and path adaptation scores. By calculating the objective function value, the merits of the current path can be evaluated.

[0160] Step S453: If the objective function value is greater than the preset threshold, then add the current path to the candidate path set.

[0161] The preset threshold is a pre-defined threshold for the objective function value, used to determine whether the current path meets the preset requirements. If the objective function value of the current path is greater than the preset threshold, it indicates that the path performs well in terms of resource consumption and path adaptation, and it is added to the candidate path set. The candidate path set is a collection used to store service paths that meet the set requirements.

[0162] Step S454: If the objective function value is less than or equal to the preset threshold, perform path segmentation optimization on the current path, generate an optimized sub-path set, and add the sub-path set to the path filtering queue.

[0163] If the objective function value of the current path is less than or equal to a preset threshold, it indicates that the path performs poorly in terms of resource consumption and path adaptation, and needs to be optimized. Path segmentation optimization involves dividing the current path into multiple sub-paths and then optimizing each sub-path individually. Local search algorithms, such as greedy algorithms and simulated annealing algorithms, can be used to optimize each sub-path. After generating the set of optimized sub-paths, they are added to the path filtering queue for further filtering in subsequent iterations.

[0164] Step S455: Repeat the above extraction, calculation and processing steps until the path filtering queue is empty or the maximum number of iterations is reached, and select the service path with the largest objective function value from the candidate path set as the optimal service path.

[0165] Repeat steps S452-S454, continuously extracting service paths from the path filtering queue for processing, until the path filtering queue is empty or the maximum number of iterations is reached. The maximum number of iterations is a pre-set upper limit used to control the algorithm's running time. When the path filtering queue is empty or the maximum number of iterations is reached, the service path with the largest objective function value is selected from the candidate path set as the optimal service path.

[0166] Step S460: Generate a comprehensive response strategy that includes path guidance information and resource adaptation strategies based on the optimal service path and auxiliary service demand characteristics.

[0167] A comprehensive response strategy is generated based on the optimal service path and the characteristics of ancillary service demand. Path guidance information provides detailed information on how to travel along the optimal service path, such as the specific route, locations passed through, and transfer information. Resource adaptation strategy is a strategy for rationally allocating transportation resources based on the characteristics of ancillary service demand, such as selecting appropriate modes of transportation and scheduling suitable travel times.

[0168] When generating a comprehensive response strategy, the optimal service path and ancillary service demand characteristics are combined, integrating route guidance information and resource adaptation strategies. For example, if the ancillary service demand characteristics indicate that the user has high requirements for travel time, the resource adaptation strategy can select the fastest mode of transportation and arrange a suitable travel time; the route guidance information can provide the specific mode of transportation and route. By generating a comprehensive response strategy that includes route guidance information and resource adaptation strategies, the user's transportation query needs are met.

[0169] Step S500: Transform the comprehensive response strategy into interactive feedback information, dynamically adapt the device parameters in the real-time environment-related information, and execute feedback operations to complete the optimized response to the traffic query request.

[0170] Transforming the comprehensive response strategy into interactive feedback information allows target users to easily access and understand traffic query results. Interactive feedback information can be in text, voice, or graphical formats. Dynamically adapting device parameters in real-time environmental information ensures that feedback information can be displayed and interacted with correctly on different devices. Executing the feedback operation sends the interactive feedback information to the target user, completing the optimized response to the traffic query request.

[0171] As one implementation method, step S500 involves converting the comprehensive response strategy into interactive feedback information and dynamically adapting it to the device parameters in the real-time environmental information to perform feedback operations and complete the optimized response to the traffic query request. Specifically, this may include the following steps S510~S580:

[0172] Step S510: Decompose the path guidance information and resource adaptation strategy in the integrated response strategy into multi-channel encoding to generate text feedback data stream and device control data stream.

[0173] Multichannel coding decomposition breaks down the path guidance information and resource adaptation strategy in the integrated response strategy into text feedback data streams and device control data streams, respectively. The text feedback data stream provides users with text-based feedback information, such as route instructions and transportation selection suggestions. The device control data stream controls related devices, such as controlling navigation devices to display routes and controlling smart transit cards for payment.

[0174] When performing multi-channel encoding decomposition, the encoding can be categorized according to the different content and uses of path guidance information and resource adaptation strategies. For example, the text description portion of the path guidance information can be encoded as a text feedback data stream, and the instruction portion of the resource adaptation strategy involving device control can be encoded as a device control data stream.

[0175] Step S520: Reorganize the text feedback data stream using natural language structure to generate a contextually coherent text sequence that conforms to user interaction habits.

[0176] Natural language restructuring (NLP) processes text feedback data streams to make their linguistic structure and expression more consistent with user interaction habits. A context-coherent text sequence refers to text content that is semantically and logically consistent and easy for users to understand. NLP techniques, such as syntactic analysis and semantic understanding, can be used in NLP restructuring. First, syntactic analysis is performed on the text feedback data stream to understand the sentence structure and grammatical relationships. Then, based on user interaction habits and language expression norms, the sentences are adjusted and optimized to make the text content more fluent and understandable. For example, complex sentences can be broken down into simpler sentences, and the sentence order can be adjusted to make the text's logic clearer.

[0177] Step S530: Perform device type matching processing on the device control data stream. Based on the device type parameters in the real-time environment association information, extract the target device communication protocol rules from the predefined protocol adaptation library.

[0178] Device type matching determines the type of the target device based on the device type parameters in the real-time environment association information. The predefined protocol adaptation library is a library that stores communication protocol rules for different devices, containing communication protocol information for various devices.

[0179] Based on the device type parameter in the real-time environment association information, the target device communication protocol rules are extracted from a predefined protocol adapter library. For example, if the target device is a smartphone, protocol rules related to smartphone communication, such as Bluetooth and Wi-Fi protocols, are extracted from the protocol adapter library. Through device type matching and protocol rule extraction, it is ensured that the device control data stream can communicate correctly with the target device.

[0180] Step S540: Based on the target device communication protocol rules, perform instruction format conversion on the device control data stream to generate a set of device execution instructions compatible with the target device hardware interface.

[0181] Instruction format conversion is the process of converting the device control data stream into an instruction format compatible with the target device's hardware interface, based on the target device's communication protocol rules. Different devices may have different hardware interfaces and communication protocols, requiring corresponding conversions of the device control data stream to be correctly recognized and executed by the target device. During instruction format conversion, the instructions in the device control data stream are formatted and encoded according to the target device's communication protocol rules. For example, if the target device's hardware interface requires instructions to use a binary encoding format, the instructions in the device control data stream are converted to that binary encoding format. Through instruction format conversion, a set of device execution instructions compatible with the target device's hardware interface is generated.

[0182] Step S550: Perform real-time alignment processing on the contextually coherent text sequence and the set of device execution instructions. Based on the device response delay parameters in the real-time environment association information, calculate the synchronization offset between the text display timing and the device control timing.

[0183] Real-time alignment is performed to ensure that the display of a context-coherent text sequence and the execution of a set of device commands are synchronized in time. Based on the device response latency parameter in the real-time environment association information, the synchronization offset between the text display timing and the device control timing is calculated. The device response latency parameter refers to the device's response time to commands; different devices may have different response latencies. During real-time alignment, the device response latency parameter is first obtained from the real-time environment association information. Then, based on the device response latency parameter, the synchronization offset between the text display timing and the device control timing is calculated. For example, if the device response latency is 1 second, the text display timing needs to be delayed by 1 second accordingly to ensure synchronization between text display and device control.

[0184] Step S560: Based on the synchronization offset, dynamically calibrate the push time nodes of the context-coherent text sequence to generate a timeline-aligned text push queue.

[0185] Based on the calculated synchronization offset, the push time of the context-coherent text sequence is dynamically calibrated. Dynamic calibration refers to adjusting the push time of the context-coherent text sequence according to the synchronization offset to synchronize it with the device control timing. A timeline-aligned text push queue is generated, and the calibrated context-coherent text sequence is arranged in the queue in chronological order. In subsequent feedback operations, text information is pushed sequentially according to the timeline-aligned text push queue to ensure synchronization between text display and device control.

[0186] Step S570: Encapsulate the timeline-aligned text push queue with the device execution instruction set across channels to generate a multimodal interaction instruction package.

[0187] Cross-channel data encapsulation integrates and encapsulates data from different channels—namely, text data represented by time-aligned text push queues and device control data represented by device execution command sets—so that it can be delivered to the target user's device in an orderly and efficient manner during the same interaction. This process needs to consider data format compatibility, transmission stability, and parsability across different devices.

[0188] When encapsulating cross-channel data, the first step is to define a unified data structure to accommodate text data and device control data. Formats such as JSON (JavaScriptObjectNotation) or XML (eXtensibleMarkupLanguage) can be used. Taking JSON as an example, create a JSON object containing two main fields: one field to store information about the text push queue, and the other field to store the set of device execution commands. Each field then organizes the data according to predefined rules. For instance, the text push queue can store each text message and its corresponding push time in chronological order as an array, while the device execution command set can store each command and its related parameters as an object in an array.

[0189] In practical applications, suppose there are three text messages in the text push queue: "Please turn right at the next intersection," "The next stop is XX station," and "2 kilometers away from the destination," with corresponding push times of t1, t2, and t3, respectively. The device execution instruction set contains two instructions: one is to control the navigation device to display the route, and the other is to control the smart speaker to play voice prompts.

[0190] Step S580: The multimodal interaction instruction packet is sent in parallel to the display terminal and device control terminal corresponding to the target user through the distributed communication interface, and the display terminal is triggered to perform a phased text rendering operation, while the device control terminal is triggered to execute the control commands in the device execution instruction set.

[0191] A distributed communication interface is used for data transmission between different devices, enabling efficient and stable data transmission. In this step, the distributed communication interface is used to send multimodal interaction command packets in parallel to the target user's corresponding display terminal and device control terminal. Parallel transmission improves data transmission efficiency and ensures that the display terminal and device control terminal can receive the required data simultaneously.

[0192] A display terminal refers to a device used to display information to users, such as a smartphone screen or an in-vehicle display. When a display terminal receives a multimodal interaction command packet, it performs a phased text rendering operation based on the text push queue information within it. The phased text rendering operation means displaying text information sequentially on the screen according to the push time nodes set in the text push queue. For example, at time t1, it displays "Please turn right at the next intersection"; at time t2, it displays "The next stop is XX station"; and at time t3, it displays "2 kilometers to the destination".

[0193] A device control terminal is used to control various devices, such as smart speakers and navigation devices. When a device control terminal receives a multimodal interaction command packet, it parses the set of device execution instructions and executes the corresponding control commands. For example, it controls a navigation device to display a specified route and set an appropriate zoom level; it controls a smart speaker to play preset voice prompts and adjust the volume, etc.

[0194] Through the above steps, the comprehensive response strategy is transformed into interactive feedback information, dynamically adapting to device parameters in the real-time environment and executing feedback operations, thus completing the optimized response to traffic query requests. The entire comprehensive traffic query optimization method, by acquiring user traffic query requests and performing a series of steps including multimodal feature analysis, feature fusion, service path optimization, and feedback information generation, fully considers user needs and the real-time traffic environment, providing users with accurate, efficient, and personalized traffic query services.

[0195] It is understood that the various algorithms, models, and network layers involved in the above descriptions of the embodiments of the present invention, such as greedy algorithms and simulated annealing algorithms, can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in this application embodiment. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solution of this application. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience, or business scenario requirements, and train the model based on a general model training method, etc. This application will not provide redundant descriptions of overly detailed implementation processes here.

[0196] Furthermore, those skilled in the art can make optimized modifications to this solution based on their own knowledge, addressing some unstated aspects. For example, after feature orthogonalization, a feature compatibility verification step can be added to calculate the correlation coefficient between the newly added vector and the orthogonal basis. When the correlation coefficient exceeds a threshold, the resource-constrained embedding vector can be subjected to Schur complement orthogonalization to eliminate potential conflicts between the linear independence of orthogonal features and the correlation of the newly added vector. When training the weight model, it can be optimized to adopt a two-stage weight prediction mechanism: in the offline stage, the basic weight allocation model is pre-trained, and in the online stage, the weight parameters are updated in real time through Kalman filtering. As for some network structures, those skilled in the art can make specific structural selections according to actual needs. For example, the temporal convolutional layer can adopt causal dilated convolution, and the spatial graph attention can adopt a multi-head attention mechanism, etc.

[0197] Please see Figure 2 , Figure 2 This is a schematic diagram of a human-computer interaction system provided in an embodiment of the present invention. The human-computer interaction system, such as a smartphone or in-vehicle terminal, includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the human-computer interaction system, capable of parsing various instructions and processing various data within the system. The communication interface 102 may optionally include standard wired or wireless interfaces (such as Wi-Fi or mobile communication interfaces), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the human-computer interaction system. The memory 103 is a storage device in the human-computer interaction system used to store programs and data. It is understood that the memory 103 here can include the built-in memory of the human-computer interaction system, or it can include extended memory supported by the system. The memory 103 provides storage space for storing the operating system of the human-computer interaction system, which may include, but is not limited to, Android system, iOS system, Windows Phone system, etc. The present invention does not limit this.

[0198] In one embodiment, the processor 101 executes the traffic comprehensive query optimization method for intelligent human-computer interaction provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A traffic comprehensive query optimization method applied to intelligent human-computer interaction, characterized in that, include: Obtain the traffic query request input by the target user in the interaction scenario, the traffic query request including natural language text and real-time environmental information; The traffic query request is subjected to multimodal feature parsing to generate semantic understanding features and scene association features. The scene association features are used to indicate the real-time traffic environment status related to the traffic query request. Based on a pre-defined language model, the semantic understanding features and the scene association features are dynamically fused to generate multi-dimensional intent features. Specifically, this includes: mapping the semantic understanding features to a first feature vector and mapping the scene association features to a second feature vector; performing dynamic weight allocation processing on the first feature vector and the second feature vector to determine semantic weight coefficients and environment weight coefficients; adjusting the weights of the first feature vector according to the semantic weight coefficients to obtain an adjusted semantic feature vector; adjusting the weights of the second feature vector according to the environment weight coefficients to obtain an adjusted environment feature vector; concatenating the adjusted semantic feature vector with the adjusted environment feature vector to generate an initial fusion feature; calling the intent recognition layer in the language big model to expand the intent dimension of the initial fusion feature to generate the multi-dimensional intent feature; specifically, this includes: decomposing the initial fusion feature into a semantic-dominant feature branch and an environment-dominant feature branch; wherein the semantic-dominant feature branch is composed of features filtered by channel attention from the adjusted semantic feature vector, and the environment-dominant feature branch is composed of features filtered by spatial attention from the adjusted environment feature vector; performing contextual demand association analysis on the semantic-dominant feature branch to extract explicit user demand features; wherein the contextual demand association analysis is performed by traversing the... The matching relationship between semantic units in the semantic-dominant feature branch and a predefined set of traffic service keywords is realized; dynamic scene adaptation analysis is performed on the environment-dominant feature branch to generate implicit scene adaptation features; wherein, the dynamic scene adaptation analysis includes: determining the dynamic constraints of the target user's scene on traffic resources based on real-time traffic state parameters in the environment-dominant feature branch; inputting the user's explicit demand features and the implicit scene adaptation features into a cross-modal interaction network to generate potential demand compensation features; wherein, the cross-modal interaction network alternately executes the projection mapping from semantic features to environmental features and environmental features... Feature compensation is achieved through feedback correction of semantic features; the explicit user demand features, implicit scene adaptation features, and potential demand compensation features are orthogonalized to eliminate redundant information and generate an orthogonal demand feature set; based on each feature vector in the orthogonal demand feature set, a multi-dimensional demand distribution matrix is ​​constructed, and the feature dimension of the multi-dimensional demand distribution matrix is ​​expanded to generate the multi-dimensional intent features containing explicit demands, implicit scene adaptation, and real-time resource constraints; wherein, the feature dimension expansion is achieved by tensor concatenating the multi-dimensional demand distribution matrix with a pre-trained resource constraint embedding vector. Based on the multi-dimensional intent features, the service path for the target user's traffic query request is optimized to generate a comprehensive response strategy; The comprehensive response strategy is transformed into interactive feedback information, and the device parameters in the real-time environment association information are dynamically adapted to execute feedback operations to complete the optimized response to the traffic query request.

2. The method as described in claim 1, characterized in that, The step of performing multimodal feature parsing on the traffic query request to generate semantic understanding features and scene-related features includes: The natural language text is segmented to obtain multiple semantic units, and the real-time environment association information is processed to extract environmental parameters to obtain multiple environmental state parameters. The semantic units are subjected to contextual semantic encoding to generate the semantic understanding features; wherein, the semantic understanding features are used to characterize the core semantic intent of the traffic query request; Spatiotemporal correlation analysis is performed on the multiple environmental state parameters to generate an environmental state coding vector; Based on the environmental state encoding vector and the device type parameter in the real-time environmental association information, the scene association feature is constructed; The scene association features include at least one of the following: real-time traffic flow distribution parameters, device response delay parameters, and spatial distance parameters between the user's location and the target traffic node.

3. The method as described in claim 1, characterized in that, The step of optimizing the service path for the target user's traffic query request based on the multi-dimensional intent features and generating a comprehensive response strategy includes: The multi-dimensional intent features are decomposed into service requirements to obtain core service requirement features and auxiliary service requirement features; Based on the real-time traffic database, an initial set of service paths matching the core service demand characteristics is obtained; Constraint matching is performed on each service path in the initial service path set to determine the resource consumption parameters and path adaptation score corresponding to each service path; Based on the characteristics of the auxiliary service demand, a dynamic optimization objective function is constructed; wherein, the dynamic optimization objective function is used to balance the comprehensive optimization weights of the resource consumption parameters and the path adaptation score; The path optimization algorithm is invoked to iteratively filter the initial service path set based on the dynamic optimization objective function to generate the optimal service path; Based on the optimal service path and the auxiliary service requirement characteristics, a comprehensive response strategy including path guidance information and resource adaptation strategy is generated; The path optimization algorithm implements iterative filtering through the following steps: Initialize the path filtering queue and add all service paths in the initial service path set to the path filtering queue; Extract the current path from the path filtering queue and calculate the objective function value corresponding to the current path; If the objective function value is greater than a preset threshold, then the current path is added to the candidate path set; If the objective function value is less than or equal to the preset threshold, then the current path is subjected to path segmentation optimization to generate an optimized sub-path set, and the sub-path set is added to the path filtering queue. Repeat the above extraction, calculation and processing steps until the path filtering queue is empty or the maximum number of iterations is reached; The service path with the largest objective function value is selected from the candidate path set as the optimal service path.

4. The method as described in claim 3, characterized in that, The step of converting the comprehensive response strategy into interactive feedback information, dynamically adapting to the device parameters in the real-time environment association information, and executing feedback operations to complete the optimized response to the traffic query request includes: The path guidance information and resource adaptation strategy in the comprehensive response strategy are decomposed into multi-channel encoding to generate text feedback data stream and device control data stream; The text feedback data stream is restructured using natural language to generate a context-coherent text sequence that conforms to user interaction habits; The device control data stream is subjected to device type matching processing. Based on the device type parameters in the real-time environment association information, the target device communication protocol rules are extracted from the predefined protocol adaptation library. According to the target device communication protocol rules, the device control data stream is converted into an instruction format to generate a set of device execution instructions compatible with the target device hardware interface; The context-coherent text sequence and the set of device execution instructions are aligned in real time. Based on the device response delay parameter in the real-time environment association information, the synchronization offset between the text display timing and the device control timing is calculated. Based on the synchronization offset, the push time nodes of the context-coherent text sequence are dynamically calibrated to generate a timeline-aligned text push queue. The time-axis-aligned text push queue and the device execution instruction set are cross-channel data encapsulated to generate a multimodal interaction instruction package; The multimodal interaction instruction package is sent in parallel to the display terminal and device control terminal corresponding to the target user through a distributed communication interface, triggering the display terminal to perform a phased text rendering operation, and simultaneously triggering the device control terminal to execute the control commands in the device execution instruction set.

5. The method as described in claim 2, characterized in that, The step of performing spatiotemporal correlation analysis on the multiple environmental state parameters to generate an environmental state encoding vector includes: Extract timestamp data and geographic location data from the real-time environment association information; Based on the timestamp data, determine the temporal distribution characteristics of the environmental state parameters; Based on the geographic location data, determine the spatial topological relationship of the environmental state parameters; The temporal distribution features and the spatial topological relationships are input into a spatiotemporal coding network to generate spatiotemporal correlation features; The spatiotemporal correlation features are subjected to dimensionality reduction to obtain the environmental state encoding vector; wherein, the spatiotemporal encoding network includes a temporal convolutional layer and a spatial graph attention layer, the temporal convolutional layer is used to extract time series dependencies, and the spatial graph attention layer is used to extract the correlation weights between geographical locations.

6. The method as described in claim 1, characterized in that, The dynamic weight allocation process for the first feature vector and the second feature vector, determining the semantic weight coefficient and the environment weight coefficient, includes: Obtain the set of sample semantic feature vectors and the set of sample environmental feature vectors from the historical interaction dataset. Each sample semantic feature vector is associated with the influence weight value of semantic features in actual decision-making, and each sample environmental feature vector is associated with the influence weight value of environmental features in actual decision-making. The sample semantic feature vector and sample environment feature vector are input into the initial weight prediction model to generate the corresponding semantic weight prediction sequence and environment weight prediction sequence. The error loss is calculated for each predicted value in the semantic weight prediction sequence and its associated actual impact weight value to generate a semantic weight error set. The error loss is calculated for each predicted value in the environmental weight prediction sequence and its associated actual impact weight value to generate an environmental weight error set. The semantic weight error set and the environmental weight error set are weighted and fused to generate a total error loss value. The parameters of the initial weight prediction model are iteratively adjusted according to the total error loss value until the total error loss value converges to a stable range, thereby generating a trained weight allocation model. Input the first feature vector and the second feature vector into the trained weight allocation model, and output the initial semantic weight value and the initial environment weight value before normalization; The initial semantic weight value and the initial environment weight value are subjected to dynamic balance constraint processing so that the sum of the two satisfies a preset constant relationship, thereby generating intermediate semantic weight value and intermediate environment weight value; The intermediate semantic weight values ​​are subjected to nonlinear activation transformation to eliminate abrupt fluctuations in the weight values ​​and generate smooth semantic weight coefficients. The intermediate environment weight values ​​are subjected to nonlinear activation transformation to eliminate abrupt fluctuations in the weight values ​​and generate smooth environment weight coefficients. The smooth semantic weight coefficients and the smooth environment weight coefficients are jointly verified to ensure that they match the interaction contribution of the first feature vector and the second feature vector, thereby generating the final usable semantic weight coefficients and environment weight coefficients.

7. The method as described in claim 1, characterized in that, The process of obtaining the traffic query request input by the target user in the interaction scenario includes: Receive the original query information submitted by the target user through voice input, text input, or touch input; The original query information is subjected to noise filtering to remove irrelevant characters or invalid audio segments; Perform intent pre-identification on the filtered original query information to determine whether it contains traffic-related key semantic units; If the original query information contains the key semantic unit, it will be marked as a valid traffic query request, and subsequent processing will be triggered. If the key semantic unit is not included, guidance information is sent to the target user to re-enter or supplement the query content.

8. A human-computer interaction system, characterized in that, include: A memory, wherein a computer program is stored; A processor is configured to load the computer program to implement the traffic integrated query optimization method for intelligent human-computer interaction as described in any one of claims 1-7.

Images