Methods, devices and storage media for predicting vehicle-machine interaction intent
By acquiring and processing current driving scenario data in the vehicle's AI assistant, enhancing user interaction commands, and performing vector conversion and scoring model filtering, the problem of the vehicle's AI assistant being unable to match the user's true intent is solved, achieving accurate response and efficient user experience improvement.
Patent Information
- Application Number
- CN202611001007.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-06
- Publication Date
- 2026-11-10
AI Technical Summary
Existing in-vehicle AI assistants cannot guarantee that their understanding of user commands matches the user's true intentions, leading to erroneous responses and a degraded user experience.
By acquiring target vehicle data that represents the semantics of the current driving scenario, the user interaction commands are enhanced to generate target enhanced commands, which are then vectorized. A preset memory vector library and a scoring model are used to select candidate memory vectors that have the highest semantic matching degree with the current driving scenario, and target interaction content is generated.
It achieves accurate response to the user's current interaction command, improving the user experience. Through a dual filtering mechanism, it ensures that the recalled candidate memory vectors match the real intent and task objectives in the current driving scenario, significantly improving the accuracy of the response.
Smart Images

Figure CN123070134A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicles, and in particular to a method, apparatus and storage medium for predicting vehicle-machine interaction intentions. Background Technology
[0002] With the continuous development of science and technology, most in-vehicle smart cockpits are equipped with AI assistants, which can understand and respond to user commands. In related technologies, after receiving a user command, the AI assistant can retrieve one or more semantically related interactions from historical content to understand the command. While the retrieved results may be similar to the user command in textual semantics, they cannot guarantee a match with the user's true intent, leading to erroneous responses and a reduced user experience. Summary of the Invention
[0003] This invention provides a method, device, and storage medium for predicting vehicle-to-everything (V2X) interaction intent, so as to achieve accurate response to the user's current interaction command and make the final generated interaction content match the user's current true intent.
[0004] In a first aspect, embodiments of the present invention provide a method for predicting vehicle-machine interaction intent, comprising: If the user's current interaction command is received, target vehicle system data that represents the semantics of the current driving scenario of the vehicle is obtained, and the current interaction command is enhanced based on the target vehicle system data to obtain the target enhanced command; The target enhancement command is vector-transformed to obtain the target vector; M candidate memory vectors associated with the target vector are retrieved from a preset memory vector library, where M is a positive integer. The preset memory vector library contains vectors corresponding to the historical interaction content between the user and the vehicle. Based on a preset scoring model, the matching degree between each candidate memory vector and the target vector in the current driving scenario semantics is scored, and the K candidate memory vectors with the highest scores are selected, where K is a positive integer; Based on the K candidate memory vectors, the target interactive content corresponding to the current interactive instruction is generated.
[0005] In some implementations, acquiring target vehicle-mounted system data for characterizing the semantics of the current driving scenario includes: The system acquires the current time when the current interaction command is received, the target position of the vehicle at the current time, the target speed of the vehicle at the current time, and the user's schedule information within a preset time period, wherein the time difference between the start time of the preset time period and the current time is less than a threshold. The enhancement processing of the current interaction command based on the target vehicle system data to obtain the target enhanced command includes: The target enhancement instruction is obtained by concatenating the current interaction instruction based on the current time, the target location, the target vehicle speed, and the schedule information.
[0006] In some implementations, retrieving M candidate memory vectors associated with the target vector from a preset memory vector library includes: The HNSW algorithm is used to determine the M vectors that are closest to the target vector from the memory vector library as the M candidate memory vectors.
[0007] In some implementations, the preset scoring model is constructed in the following manner: The user's N historical interaction commands are transformed into vectors to obtain the corresponding N historical interaction vectors, where N is a positive integer; For each historical interaction vector, retrieve M memory vectors associated with that historical interaction vector from the preset memory vector library; For each historical interaction vector, determine the supervision label for each memory vector corresponding to that historical interaction vector, wherein the M memory vectors and M supervision labels corresponding to each historical interaction vector are used as a set of training samples; The initial preset scoring model is trained based on the training samples, and the trained scoring model is used as the preset scoring model.
[0008] In some implementations, determining the supervision label for each memory vector corresponding to each historical interaction vector includes: For each historical interaction vector, the historical interaction vector and its corresponding M memory vectors are input into a preset large language model. The preset large language model is used to analyze the matching score between each memory vector and the true intent of the corresponding historical interaction command under the semantics of the corresponding driving scenario. Based on the matching score, the supervision label of each memory vector is determined.
[0009] In some embodiments, the method further includes: Determine the target label corresponding to each of the K candidate memory vectors; Based on the target enhancement instruction, the K candidate memory vectors, and the target label corresponding to each candidate memory vector in the K candidate memory vectors, a target sample is constructed; The preset scoring model is updated based on the target sample.
[0010] In some implementations, determining the target label corresponding to each of the K candidate memory vectors includes: If feedback is received from the user regarding the target interactive content, a user behavior tag is determined based on the feedback. By analyzing the matching score between each of the K candidate memory vectors and the true intent of the current interaction command under the semantics of the current driving scenario using a pre-set large language model, a model judgment label corresponding to each of the K candidate memory vectors is obtained. For each of the K candidate memory vectors, the model judgment label corresponding to the candidate memory vector is fused with the user behavior label to obtain the target label corresponding to the candidate memory vector.
[0011] Secondly, embodiments of the present invention provide a vehicle-machine interaction intent prediction device, comprising: The instruction enhancement module is used to obtain target vehicle system data that characterizes the semantics of the current driving scenario of the vehicle when it receives the user's current interaction instruction, and to enhance the current interaction instruction based on the target vehicle system data to obtain the target enhanced instruction; A vector conversion module is used to perform vector conversion on the target enhancement command to obtain a target vector; The retrieval module is used to retrieve M candidate memory vectors associated with the target vector from a preset memory vector library, where M is a positive integer, and the preset memory vector library contains vectors corresponding to the historical interaction content between the user and the vehicle. The first processing module is used to score the matching degree between each candidate memory vector and the target vector under the current driving scenario semantics based on a preset scoring model, and to select the K candidate memory vectors with the highest scores, where K is a positive integer; The second processing module is used to generate target interaction content corresponding to the current interaction instruction based on the K candidate memory vectors.
[0012] Thirdly, embodiments of the present invention provide a vehicle-machine interaction device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the above-described vehicle-machine interaction method.
[0013] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described vehicle-machine interaction method.
[0014] The above-described one or more technical solutions in the embodiments of this application have at least the following technical effects: The vehicle-to-everything (V2X) interaction intent prediction method provided in this specification, upon receiving a user's current interaction command, acquires target V2X data representing the semantics of the current driving scenario. Based on the target V2X data, the current interaction command is enhanced to obtain a target enhanced command, which is then vectorized to obtain a target vector. M candidate memory vectors associated with the target vector are retrieved from a preset memory vector library, which contains vectors corresponding to historical user-vehicle interactions. Based on a preset scoring model, the matching degree between each candidate memory vector and the target vector under the current driving scenario semantics is scored, and the K candidate memory vectors with the highest scores are selected. Based on the K candidate memory vectors, target interaction content corresponding to the current interaction command is generated. This solution enhances the current interaction command using target V2X data, integrating the current driving scenario semantics into the user command. Therefore, when using the target vector corresponding to the enhanced command for retrieval, it can recall candidate memory vectors related to the current driving scenario semantics; that is, the recalled candidate memory vectors are more closely aligned with the user's implicit true intent and task objectives in the current driving scenario. Furthermore, this solution employs a two-stage retrieval process. In the first stage, M candidate memory vectors semantically related to the target vector are selected. In the second stage, the matching degree between each candidate memory vector and the target vector under the current driving scenario semantics is evaluated. This process determines the effectiveness of each candidate memory vector under the current task objective, achieving dual filtering based on semantic relevance and task effectiveness. This significantly improves the effectiveness of candidate memory vector recall. Consequently, the target interactive content generated based on K candidate memory vectors can accurately respond to the user's current interactive commands, thereby enhancing the user experience. Attached Figure Description
[0015] Figure 1 A flowchart illustrating a vehicle-machine interaction intent prediction method provided in an embodiment of this specification; Figure 2 This is a schematic diagram of the functional modules of a vehicle-machine interaction intent prediction device provided in the embodiments of this specification; Figure 3 This is a schematic diagram of the structure of a vehicle-machine interaction intent prediction device provided in the embodiments of this specification. Detailed Implementation
[0016] The overall technical solution of this specification embodiment is as follows: If a user's current interaction command is received, target vehicle-mounted data representing the semantics of the current driving scenario is obtained, and the current interaction command is enhanced based on the target vehicle-mounted data to obtain a target enhanced command; the target enhanced command is vector-converted to obtain a target vector; M candidate memory vectors associated with the target vector are retrieved from a preset memory vector library, where M is a positive integer, and the preset memory vector library contains vectors corresponding to the user's historical interaction content with the vehicle; based on a preset scoring model, the matching degree between each candidate memory vector and the target vector under the current driving scenario semantics is scored, and the K candidate memory vectors with the highest scores are selected, where K is a positive integer; based on the K candidate memory vectors, target interaction content corresponding to the current interaction command is generated.
[0017] In this embodiment, the current interaction command is enhanced using target vehicle data, integrating the semantics of the current driving scenario into the user command. Therefore, when searching using the target vector corresponding to the enhanced command, candidate memory vectors related to the semantics of the current driving scenario can be recalled. This means the recalled candidate memory vectors better match the user's implicit true intent and task objective in the current driving scenario. Furthermore, this solution employs a two-stage retrieval process. The first stage filters out M candidate memory vectors semantically related to the target vector. The second stage evaluates the matching degree between each candidate memory vector and the target vector in the current driving scenario semantics, determining the effectiveness of each candidate memory vector in the current task objective. This achieves dual filtering based on semantic relevance and task effectiveness, significantly improving the effectiveness of candidate memory vector recall. Consequently, the target interaction content generated based on K candidate memory vectors can accurately respond to the user's current interaction command, improving the user experience.
[0018] To better understand the above technical solutions, the technical solutions of the embodiments of this specification will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of this specification and the specific features in the embodiments are detailed descriptions of the technical solutions of the embodiments of this specification, rather than limitations on the technical solutions of this specification. Unless otherwise specified, the embodiments of this specification and the technical features in the embodiments can be combined with each other.
[0019] First, it should be clarified that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0020] This specification provides an embodiment of a method for predicting vehicle-to-machine interaction intent, such as... Figure 1 As shown, the method includes the following steps: Step S101: If the user's current interaction command is received, obtain the target vehicle system data used to characterize the semantics of the current driving scenario of the vehicle, and enhance the current interaction command based on the target vehicle system data to obtain the target enhanced command; Step S102: Perform vector transformation on the target enhancement command to obtain the target vector; Step S103: Retrieve M candidate memory vectors associated with the target vector from the preset memory vector library, where M is a positive integer, and the preset memory vector library contains vectors corresponding to the historical interaction content between the user and the vehicle. Step S104: Based on the preset scoring model, score the matching degree between each candidate memory vector and the target vector in the current driving scenario semantics, and select the K candidate memory vectors with the highest scores, where K is a positive integer; Step S105: Based on the K candidate memory vectors, generate the target interaction content corresponding to the current interaction command.
[0021] The methods provided in the embodiments of this specification can be applied to vehicles, cloud servers that are connected to vehicles, or systems consisting of vehicles and cloud servers; no limitations are imposed here.
[0022] In step S101, the user's current interaction command can be the user's original command or a command obtained after preprocessing the user's original command; there is no limitation here. The user's original command can include, but is not limited to, voice commands and text commands. In some embodiments, taking the user's original command as a voice command and the current interaction command as a command obtained after preprocessing the original command as an example, after receiving the original voice command, speech recognition can be performed on the voice command to obtain the corresponding speech text. Then, the speech text can be preprocessed, including but not limited to filtering out meaningless words in the speech text and normalizing the text (such as normalizing the uppercase and lowercase of numbers and letters). The preprocessed command is then used as the current interaction command.
[0023] Target vehicle infotainment data, also known as vehicle infotainment context, is data used to characterize the semantics of the current driving scenario of the vehicle. In some embodiments, target vehicle infotainment data can be obtained in the following ways: obtaining the current time when the current interaction command is received, the target position of the vehicle at the current time, the target speed of the vehicle at the current time, and the user's schedule information within a preset time period, wherein the time difference between the start time of the preset time period and the current time is less than a threshold.
[0024] It should be noted that the schedule information can be whether the user has any scheduled activities during the preset time period, such as meetings or travel plans. The duration and threshold of the preset time period can be set according to actual needs. For example, the duration of the preset time period can be half an hour, one hour, etc., and the threshold can be 10 minutes, half an hour, etc. There are no restrictions here.
[0025] Of course, in addition to the current time, target location, target speed, and schedule information mentioned above, the target vehicle data can also include other data, such as weather and user fatigue status.
[0026] After acquiring the target vehicle infotainment system data, the current interaction command can be enhanced using the target vehicle infotainment system data to obtain the target enhanced command. In some embodiments, taking the target vehicle infotainment system data as the current time, target location, target vehicle speed, and schedule information as an example, the enhancement processing of the current interaction command can be: concatenating the current interaction command based on the current time, target location, target vehicle speed, and schedule information to obtain the target enhanced command.
[0027] The following example illustrates the process of generating target enhancement instructions.
[0028] If the current interaction command received from the user is "I'm hungry", the current time is 23:00, the target location is G4 Expressway, the target speed is 110, and the schedule information is no schedule, then the target vehicle system data is concatenated with the current interaction execution to obtain the target enhanced command: "[Time: 23:00][Location: G4 Expressway][Speed: 110][Schedule: No schedule] I'm hungry".
[0029] In step S102, the vector transformation of the target enhancement instruction can be achieved in various ways. In some embodiments, the BGE (BAAI General Embedding) semantic embedding model can be used to output the target vector corresponding to the target enhancement instruction. Of course, other vector generation models or algorithms can also be used for vector transformation, such as Word2Vec (Word to Vector), TF-IDF, etc., which are not limited here. Taking the BGE semantic embedding model as an example, the target enhancement instruction can be converted into a 1024-dimensional vector.
[0030] In the embodiments described in this specification, in order to understand the user's true intention, relevant candidate memory vectors can be retrieved from a preset memory vector library based on the target vector, and then the candidate memory vectors can be injected into the large language model to generate an accurate and reliable answer.
[0031] It should be noted that, compared to the traditional method of directly using the user's current interaction command for memory retrieval, the embodiments in this specification use enhanced commands, which incorporate the semantics of the current driving scenario, significantly improving the relevance of memory retrieval. For example, if the user's current interaction command is "I'm hungry," the traditional method only considers semantic matching on the text. When performing memory retrieval, it might recall the memory "I'm hungry, order me some beef noodles to be delivered to my home." The recalled data is only semantically related to the current interaction command, but the location may be outdated. However, in the implementation of this specification, the command is enhanced through target vehicle system data. For example, the enhanced command is "[Time: 23:00][Location: G4 Expressway][Speed: 110][Schedule: No Schedule] I'm hungry," which can recall the memory "You chose AA convenience store bento boxes three times in the past at the G4 service area." Although the text does not contain the word "hungry," it can accurately match the behavioral pattern of "eating quickly after a long drive," achieving a qualitative change from "textual semantic relevance" to "task usefulness."
[0032] In the embodiments of this specification, when searching the preset memory vector library, a two-stage retrieval can be achieved based on steps S103 and S104. Step S103 is the first stage of retrieval based on semantic relevance, and step S104 is the second stage of retrieval based on task validity.
[0033] It should be noted that the preset memory vector library contains vectors corresponding to the user's historical interactions with the vehicle. When building the preset memory vector library, to avoid adding invalid memory vectors to the library, you can set memory inclusion criteria or exclusion conditions according to actual needs.
[0034] In some embodiments, memories (i.e., historical interaction content) can be categorized into the following types: semantic memory, procedural memory, episodic memory, and active memory. Semantic memory can be a user's stable, real-time knowledge about themselves, others, preferences, or the world. Procedural memory can be user-defined, repeatable automated workflows or conditional rules. Episodic memory is a user's memory of specific events experienced at a specific time and place, such as "Last Wednesday, my friend and I went to that newly opened restaurant." Active memory is content that the user actively requests to remember.
[0035] For semantic memory, the inclusion criteria may include: ① basic user information (occupation, identity); ② names and relationships of family members, partners, and pets; ③ home address and company address; ④ long-term stable personal preferences (music preferences, food tastes, favorite brands, restaurant types, etc.). Exclusion criteria may include temporary or one-time information.
[0036] For procedural memory, the criteria for inclusion can include preset key logical structures, such as the logical structure corresponding to a conditional trigger action (e.g., "From now on, when you say 'go for it,' play this song"), or the logical result corresponding to a scenario-based operation sequence (e.g., "Every time you get in the car, turn on the air conditioning first, then play music"). In some embodiments, whether it is a preset key logical result can be determined based on a preset sentence pattern. The preset sentence pattern can be set according to actual needs, such as "if / as / when…then…", "whenever / every time…".
[0037] For contextual memories, the criteria for inclusion in the database may include: ① Type matching: belonging to the "important interaction" category of events (such as personal stories, memories, etc.); ② Judgment key: narrative in nature, describing the user's experience at a specific event location.
[0038] For active memory, the inclusion criteria may include: ① display preferences and emphasis (such as "Remember I don't eat bitter melon" or "Don't play slow songs like this for me in the future"); ② mandatory binding of key entity information (such as "Remember my license plate number is Zhejiang AXXXXX"); ③ long-term effective personal memos / habit settings (such as "Remember I have to pick up my child every Friday afternoon").
[0039] By using the aforementioned memory storage criteria and / or exclusion conditions, the historical interaction content between the user and the vehicle is filtered to select the historical interaction content that can be stored in the database. The selected historical interaction content is then converted into vectors and added to a preset memory vector. In some embodiments, the vector conversion method for historical interaction content is the same as the vector conversion method for target enhancement instructions. For example, both use the BGE semantic embedding model to convert them into 1024-dimensional embedding vectors.
[0040] In step S103, the first stage of retrieval is achieved by retrieving M candidate memory vectors associated with the target vector from a preset memory vector library. It should be noted that the retrieval method for the preset memory vector library can be selected according to actual needs and is not limited here.
[0041] In some embodiments, the Hierarchical Navigable Small Worlds (HNSW) algorithm is used to determine the M vectors closest to the target vector from the memory vector database as the M candidate memory vectors. Specifically, HNSW is a high-dimensional vector approximate nearest neighbor retrieval algorithm. Its core is to construct a multi-level graph structure, using high-level sparse nodes for fast global coarse localization, and then descending layer by layer to the low-level dense nodes for fine-grained local search, thereby achieving extremely fast retrieval speed and high recall accuracy in massive high-dimensional datasets.
[0042] In this embodiment of the specification, after retrieving the M candidate memory vectors that are closest to the target vector using HNSW, the task effectiveness of each candidate memory vector is evaluated in step S104.
[0043] In some embodiments, a preset scoring model is used to further filter the M candidate memory vectors. The preset scoring model can be a multilayer perceptron (MLP) model, or other models, which are not limited here.
[0044] For ease of explanation, we take the MLP model as the default scoring model as an example. The input dimension of the default scoring model can be 2048 dimensions, which is composed of a 1024-dimensional target vector and a 1024-dimensional candidate memory vector. The hidden layer dimension is 256, and the output is a single score scalar, which is the score of the matching degree between the input target vector and the candidate memory vector on the task.
[0045] For example, the target vector The target vector belongs to a 1024-dimensional vector space and is generated by the target enhancement command (such as "[Time: 23:00][Location: G4 Expressway][Speed: 110][Schedule: No Schedule] I'm hungry") through the BGE semantic embedding model. This is for candidate memory vectors. When adding a pre-defined memory vector library, the same BGE semantic inference model is used to perform vector transformation and storage on each memory entry. The model's output is a task relevance score between the candidate memory vector and the target vector. The candidate memory vectors are sorted by their corresponding scores, and the K candidate memory vectors with the highest scores are selected as the final vector retrieval results.
[0046] In some embodiments, the construction process of the preset rating model can be as follows: N historical interaction commands of the user are vectorized to obtain N corresponding historical interaction vectors, where N is a positive integer; for each historical interaction vector, M memory vectors associated with that historical interaction vector are retrieved from the preset memory vector library; for each historical interaction vector, a supervision label is determined for each memory vector corresponding to that historical interaction vector, wherein the M memory vectors and M supervision labels corresponding to each historical interaction vector are used as a set of training samples; the initial preset rating model is trained based on the training samples, and the trained rating model is used as the preset rating model.
[0047] Specifically, when constructing a preset scoring model, it is necessary to first construct a training sample set. In some embodiments, N historical interaction vectors between the user and the vehicle can be obtained. These N historical interaction vectors can be vectors corresponding to interaction instructions received within a preset historical period, or they can be N historical interaction vectors randomly selected, or they can be N historical interaction vectors determined by other means. No limitation is made here.
[0048] For each historical interaction vector, HNSW can be used to retrieve M memory vectors from a pre-defined memory vector library. For each memory vector The corresponding supervision label is determined. In some embodiments, the supervision label can be generated through the following steps: For each historical interaction vector, the historical interaction vector and the corresponding M memory vectors are input into a preset large language model, so as to analyze the matching score between each memory vector and the true intent of the corresponding historical interaction command under the corresponding driving scenario semantics through the preset large language model, and determine the supervision label of each memory vector based on the matching score.
[0049] Specifically, for each historical interaction vector, this historical interaction vector and its corresponding M memory vectors are used as input data for a pre-defined large language model. The pre-defined large language model analyzes each memory vector based on pre-defined prompts to obtain a matching score between each memory vector and the true intent of the corresponding historical interaction command. To better understand the analysis basis of the pre-defined large language model, examples of the prompts in the pre-defined large language model are provided below.
[0050] In some embodiments, the prompt words for the preset large language model can be: You are a memory retrieval scoring model. Your task is to output a relevance score for each memory vector based on the target vector of a given user query, the memory vector, and the vehicle context data, and provide a brief explanation.
[0051] Scoring objective: To determine the degree of matching between each memory vector and the target vector of the current user query and the vehicle context data. The higher the score, the more relevant it is and the more likely it is to be the memory that the user is currently trying to retrieve.
[0052] Please consider the following factors: (1) Semantic relevance: whether the memory vector is consistent with the user's query topic; (2) Contextual consistency: whether the memory vector is consistent with the current time, location, vehicle speed, schedule and other vehicle context data; (3) Task availability: whether the memory vector is appropriate to be returned to the user in the current scenario; (4) Specificity: whether the memory vector can clearly answer the user's query, rather than giving an overly general answer; (5) User preference consistency: whether the memory vector is consistent with the user's historical behavior or preferences.
[0053] Scoring rules: 1-2 points: completely irrelevant; 3-4 points: extremely weak relevance; 5-6 points: some relevance, but not enough to support the search; 7-8 points: moderately relevant, possibly useful; 9-10 points: highly relevant, suitable for return.
[0054] Output requirements: Please output the following fields for each candidate memory: (1) memory_id: memory vector number; (2) score: an integer between 1 and 10; (3) reason: a brief reason of no more than 50 words. Please strictly output in JSON format and do not output any extra content.
[0055] Based on the aforementioned prompts, for each historical interaction vector, the preset large language model can output M scores from the corresponding M memory vectors. In some embodiments, the raw scores output by the preset large language model can be... Normalization is performed, for example, by linear normalization to the [0,1] interval, and the result is used as a supervision label. .
[0056] When training a pre-defined rating model, mean squared error can be used as the loss function. If the output rating of the pre-defined rating model is... The loss function L can be expressed as:
[0057] In some embodiments, the parameters of the scoring model can be updated using the Adam optimizer, and the trained scoring model can be used as the preset scoring model after the scoring model has been trained.
[0058] In this embodiment of the specification, a preset scoring model can be used to obtain scores for M candidate memory vectors related to the target vector. Furthermore, the scores of the M candidate memory vectors can be sorted in order of magnitude, and the K candidate memory vectors with the highest scores are used as vectors related to the target vector task. K can be set to any value according to actual needs, such as K being 10, 15, etc.
[0059] In step S105, the final target interaction content can be generated based on K candidate memory vectors and the target vector. In some embodiments, the K candidate memory vectors and the target vector can be input into a large language model, which can use the candidate memory to output the target interaction content that has the highest matching degree with the target vector and best matches the user's intent in the current driving scenario semantics.
[0060] Considering the common problem of feedback sparsity in user interactions (such as users not displaying or correcting the vehicle system's feedback), this embodiment of the specification can construct multi-source heterogeneous supervision labels to continuously update and optimize the preset scoring model, thereby improving the accuracy of user-vehicle interaction. Therefore, in some embodiments, the following steps can also be performed: determining the target label corresponding to each of the K candidate memory vectors; constructing target samples based on the target enhancement instruction, the K candidate memory vectors, and the target labels corresponding to each of the K candidate memory vectors; and updating the preset scoring model based on the target samples.
[0061] Specifically, for the K candidate memory vectors, the target label for each candidate memory vector comes from two aspects: user behavior feedback and large language model scoring. In some embodiments, the K candidates and the target labels of the vectors can be determined through the following steps: if feedback results from the user regarding the target interaction content are received, a user behavior label is determined based on the feedback results; a preset large language model is used to analyze the matching score between each of the K candidate memory vectors and the true intent of the current interaction command under the semantics of the current driving scenario, obtaining a model judgment label corresponding to each of the K candidate memory vectors; for each candidate memory vector, the model judgment label corresponding to the candidate memory vector is fused with the user behavior label to obtain the target label corresponding to the candidate memory vector.
[0062] Firstly, regarding user behavior tags, if a user adopts the target interaction content, such as confirming the execution of the target interaction content through clicking or voice, the user behavior tag can be set to 1. If the user corrects or ignores the target interaction content, the user behavior tag can be set to 0.
[0063] Secondly, regarding the model's decision label, K candidate memory vectors and the target vector can be input into a pre-defined large language model. The pre-defined large language model generates a score for each candidate memory vector. The scoring process of the pre-defined large language model can refer to the method described above. The output score of the pre-defined large language model is 1-10 points. The score is converted to the [0,1] interval through linear normalization and used as the model's decision label.
[0064] Furthermore, user behavior labels and model decision labels are fused. In some embodiments, different strategies can be selected based on the presence or absence of user labels. For example, if user behavior exists, the user behavior label is the primary factor, and the model decision label is secondary. For instance, the weight of the user behavior label is set to 0.8, and the weight of the model decision label is set to 0.2. The final target label of the corresponding candidate memory vector is obtained by weighted summation. If no user behavior exists, the model decision label can be used entirely.
[0065] In the embodiments of this specification, the accumulated training samples can be uploaded to the cloud periodically. For example, the daily accumulated training samples (target vector, candidate memory vector, and target label of candidate memory vector) can be uploaded to the cloud to update the parameters of the preset scoring model and realize the periodic optimization of the model.
[0066] In summary, by injecting vehicle context data into the real-time state of the vehicle system and using the target vector corresponding to the enhanced command to detect candidate memory vectors, the recalled candidate memory vectors are more closely aligned with the user's implicit true intent and task objectives in the current driving scenario, rather than relying solely on semantic similarity. This ensures that the retrieved memory vectors can serve the task objectives, effectively improving the recall rate of useful memories and reducing erroneous feedback content caused by invalid memories.
[0067] In the embodiments of this specification, a cascaded retrieval paradigm that combines vehicle context, HNSW coarse screening, and precise sorting using a preset scoring model breaks through the limitations of traditional RAG which relies solely on semantic similarity. It achieves dual filtering from semantic relevance to task usefulness, significantly improving the effectiveness of memory recall in in-vehicle scenarios.
[0068] In addition, the embodiments in this specification can automatically adjust the sorting of candidate memory vectors based on the vehicle context data through a preset scoring model, without manual intervention. In mixed task scenarios, the fluctuation of memory retrieval accuracy is reduced, and the system's generalization ability is significantly enhanced.
[0069] In the embodiments of this specification, multi-source heterogeneous supervision labels are constructed by fusing user behavior labels and model judgment labels. These labels are used to periodically optimize the parameters of the preset scoring model, which has continuous learning capabilities and realizes periodic iterative optimization of the detection strategy. This improves the accuracy of the preset scoring model in recalling useful memories and ensures the long-term service consistency and scenario adaptability of the system.
[0070] Based on the same inventive concept, embodiments of this specification also provide a vehicle-machine interaction intent prediction device, such as... Figure 2 As shown, the device includes: The instruction enhancement module 201 is used to obtain target vehicle system data that characterizes the semantics of the current driving scenario of the vehicle when it receives the user's current interaction instruction, and to enhance the current interaction instruction based on the target vehicle system data to obtain the target enhanced instruction; Vector conversion module 202 is used to perform vector conversion on the target enhancement command to obtain the target vector; The retrieval module 203 is used to retrieve M candidate memory vectors associated with the target vector from a preset memory vector library, where M is a positive integer, and the preset memory vector library contains vectors corresponding to the historical interaction content between the user and the vehicle. The first processing module 204 is used to score the matching degree between each candidate memory vector and the target vector in the current driving scenario semantics based on a preset scoring model, and to select the K candidate memory vectors with the highest scores, where K is a positive integer; The second processing module 205 is used to generate target interactive content corresponding to the current interactive instruction based on the K candidate memory vectors.
[0071] Regarding the above-mentioned device, the specific implementation of each step has been described in detail in the embodiment of the vehicle-machine interaction intent prediction method provided in the specification, and will not be elaborated here.
[0072] Based on the same inventive concept, embodiments of the present invention also provide a vehicle-machine interaction intent prediction device, such as... Figure 3 The aforementioned includes a memory 304, a processor 302, and a computer program stored in the memory 304 and executable on the processor 302. When the processor 302 executes the program, it implements any of the above-described methods for predicting vehicle-machine interaction intentions.
[0073] Among them, Figure 3 In this document, a bus architecture (represented by bus 300) is used. Bus 300 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 302 and memory represented by memory 304. Bus 300 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 305 provides an interface between bus 300 and receiver 301 and transmitter 303. Receiver 301 and transmitter 303 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 302 is responsible for managing bus 300 and general processing, while memory 304 can be used to store data used by processor 302 during operation.
[0074] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0075] Based on the same inventive concept, embodiments of this specification provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described vehicle-machine interaction intent prediction method.
[0076] Based on the same inventive concept, embodiments of this specification provide a computer program product, which includes a computer program that, when executed by a processor, loads and executes the steps of the above-described vehicle-machine interaction intent prediction method.
[0077] The functions described herein can be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions can be stored as one or more instructions or code on or transmitted via a computer-readable medium. Other examples and embodiments are within the scope of this invention and the appended claims. For example, due to the nature of software, the functions described above can be implemented using software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Furthermore, the functional units can be integrated into a single processing unit, or each unit can exist physically separately, or two or more units can be integrated into a single unit.
[0078] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.
[0079] The units described as separate components may or may not be physically separate. Similarly, the components of the control device may or may not be physical units; they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0080] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.
[0081] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of the claims of the present invention.
Claims
1. A method for predicting vehicle-machine interaction intent, characterized in that, include: If the user's current interaction command is received, target vehicle system data that represents the semantics of the current driving scenario of the vehicle is obtained, and the current interaction command is enhanced based on the target vehicle system data to obtain the target enhanced command; The target enhancement command is vector-transformed to obtain the target vector; M candidate memory vectors associated with the target vector are retrieved from a preset memory vector library, where M is a positive integer. The preset memory vector library contains vectors corresponding to the historical interaction content between the user and the vehicle. Based on a preset scoring model, the matching degree between each candidate memory vector and the target vector in the current driving scenario semantics is scored, and the K candidate memory vectors with the highest scores are selected, where K is a positive integer; Based on the K candidate memory vectors, the target interactive content corresponding to the current interactive instruction is generated.
2. The method as described in claim 1, characterized in that, The acquisition of target vehicle system data used to characterize the semantics of the current driving scenario includes: The system acquires the current time when the current interaction command is received, the target position of the vehicle at the current time, the target speed of the vehicle at the current time, and the user's schedule information within a preset time period, wherein the time difference between the start time of the preset time period and the current time is less than a threshold. The enhancement processing of the current interaction command based on the target vehicle system data to obtain the target enhanced command includes: The target enhancement instruction is obtained by concatenating the current interaction instruction based on the current time, the target location, the target vehicle speed, and the schedule information.
3. The method as described in claim 1, characterized in that, The step of retrieving M candidate memory vectors associated with the target vector from a preset memory vector library includes: The HNSW algorithm is used to determine the M vectors that are closest to the target vector from the memory vector library as the M candidate memory vectors.
4. The method as described in claim 1, characterized in that, The preset scoring model is constructed in the following way: The user's N historical interaction commands are transformed into vectors to obtain the corresponding N historical interaction vectors, where N is a positive integer; For each historical interaction vector, retrieve M memory vectors associated with that historical interaction vector from the preset memory vector library; For each historical interaction vector, determine the supervision label for each memory vector corresponding to that historical interaction vector, wherein the M memory vectors and M supervision labels corresponding to each historical interaction vector are used as a set of training samples; The initial preset scoring model is trained based on the training samples, and the trained scoring model is used as the preset scoring model.
5. The method as described in claim 4, characterized in that, The step of determining the supervision label for each memory vector corresponding to each historical interaction vector includes: For each historical interaction vector, the historical interaction vector and its corresponding M memory vectors are input into a preset large language model. The preset large language model is used to analyze the matching score between each memory vector and the true intent of the corresponding historical interaction command under the semantics of the corresponding driving scenario. Based on the matching score, the supervision label of each memory vector is determined.
6. The method as described in claim 1, characterized in that, The method further includes: Determine the target label corresponding to each of the K candidate memory vectors; Based on the target enhancement instruction, the K candidate memory vectors, and the target label corresponding to each candidate memory vector in the K candidate memory vectors, a target sample is constructed; The preset scoring model is updated based on the target sample.
7. The method as described in claim 6, characterized in that, Determining the target label corresponding to each of the K candidate memory vectors includes: If feedback is received from the user regarding the target interactive content, a user behavior tag is determined based on the feedback. By analyzing the matching score between each of the K candidate memory vectors and the true intent of the current interaction command under the semantics of the current driving scenario using a pre-set large language model, a model judgment label corresponding to each of the K candidate memory vectors is obtained. For each of the K candidate memory vectors, the model judgment label corresponding to the candidate memory vector is fused with the user behavior label to obtain the target label corresponding to the candidate memory vector.
8. A vehicle-machine interaction intent prediction device, characterized in that, include: The instruction enhancement module is used to obtain target vehicle system data that characterizes the semantics of the current driving scenario of the vehicle when it receives the user's current interaction instruction, and to enhance the current interaction instruction based on the target vehicle system data to obtain the target enhanced instruction; A vector conversion module is used to perform vector conversion on the target enhancement command to obtain a target vector; The retrieval module is used to retrieve M candidate memory vectors associated with the target vector from a preset memory vector library, where M is a positive integer, and the preset memory vector library contains vectors corresponding to the historical interaction content between the user and the vehicle. The first processing module is used to score the matching degree between each candidate memory vector and the target vector under the current driving scenario semantics based on a preset scoring model, and to select the K candidate memory vectors with the highest scores, where K is a positive integer; The second processing module is used to generate target interaction content corresponding to the current interaction instruction based on the K candidate memory vectors.
9. A vehicle-mounted infotainment system (V2S) interaction device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-7.
