Interaction control system and interaction control method

Abstract

This application discloses an interactive control system and an interactive control method, belonging to the field of interactive control technology. The interactive control system includes an intent understanding module, an intent processing module, and an execution module. The intent understanding module includes at least one large model, used to send the current intent output by any large model to the intent processing module. The intent processing module is used to determine the current attention score of at least a portion of the text fragments in the context of the current intent, select target text fragments from these text fragments in descending order of current attention score, generate a target task based on the current intent and the target text fragments, and convert the target task into control instructions which are then sent to the execution module. The execution module is used to execute the control instructions. This decouples intent understanding and task execution, enabling flexible model switching.

Description

Technical Field

[0001] This application relates to the field of interactive control technology, specifically to interactive control systems and interactive control methods. Background Technology

[0002] With the widespread application of large-scale models in fields such as natural language processing and multimodal interaction, their core bottleneck has gradually shifted from simply scaling up to accurately understanding and efficiently executing user intent. Integrators need to integrate multiple different large-scale models or back up the underlying large-scale model.

[0003] The related large model architecture couples intent understanding and task execution into a single module. If the large model base supplier stops providing services (such as through iterative obsolescence), the integrator needs to reconstruct the entire system, and the model switching lacks flexibility. Summary of the Invention

[0004] This application provides an interactive control system and interactive control method, which aims to solve the problem of lack of flexibility in model switching.

[0005] Firstly, an interactive control system is provided, including an intent understanding module, an intent processing module, and an execution module;

[0006] The intent understanding module includes at least one large model, which is used to send the current intent output by any large model to the intent processing module;

[0007] The intent processing module is used to determine the current attention score of at least some text fragments in the context of the current intent, select target text fragments from the above-mentioned at least some text fragments in descending order of current attention score, generate target tasks based on the current intent and target text fragments, and convert the target tasks into control instructions and send them to the execution module.

[0008] The execution module is used to execute control commands.

[0009] In some of these designs, the intent processing module is used to generate a target task based on the current intent and the target text fragment, including:

[0010] The current intent is transformed into a structured task using a preset structured data format.

[0011] Add the target text fragment to the structured task to obtain the target task.

[0012] In some of these design approaches, the current intent includes the object of the operation and the intent name;

[0013] The intent processing module is used to convert the current intent into a structured task using a preset structured data format, including:

[0014] Determine the action based on the intent name;

[0015] Determine the target delay label based on the target object and the operation action;

[0016] Using a structured data format, a structured task is generated based on the current intent and target delay label.

[0017] In some of these design approaches, the intent processing module is used to determine the target delay label based on the object of the operation and the action performed, including:

[0018] Determine the target safety coefficient and target user experience coefficient based on the target object and the operation action;

[0019] The target delay label is determined based on the target security coefficient and the target user experience coefficient.

[0020] In some of these designs, the intent processing module is used to determine the current attention score of at least a portion of the text fragments preceding the current intent, including:

[0021] The historical attention scores of at least some of the cached text segments are corrected to obtain the current attention scores; wherein the correction process includes: attenuating the historical attention scores.

[0022] In some of these design approaches, the intent processing module includes:

[0023] The transformation unit is used to convert the current intent into a structured task using a structured data format;

[0024] The context management unit is used to determine the current attention score of at least some of the text fragments, and select target text fragments from the at least some text fragments in descending order of the current attention score to add them to the structured task to obtain the target task.

[0025] The sending unit is used to convert the target task into control commands and send them to the execution module.

[0026] In some of these design approaches, the target task includes a target delay label, and the sending unit employs a multi-channel architecture; the intent processing module also includes a scheduling unit, which is used for:

[0027] Based on the target delay label and the pre-defined correspondence between delay labels and scheduling policies, determine the target scheduling policy corresponding to the target delay label;

[0028] Based on the target scheduling strategy, determine the target channel in the multi-channel architecture;

[0029] Add the target task to the task queue corresponding to the target channel.

[0030] In some of these design approaches, the current intent and the structured task also include an intent source, which is used to indicate the larger model that generated the current intent;

[0031] The context management unit is also used to add the current attention score of the target text fragment to the structured task;

[0032] The scheduling unit is used to determine the target scheduling policy corresponding to the target delay label based on the target delay label and the preset correspondence between delay labels and scheduling policies, including:

[0033] Determine the intent source switching based on the current attention score of the target text segment;

[0034] If the intent source is determined not to be switched, the target scheduling strategy is determined based on the target delay label and the corresponding relationship.

[0035] In some of these design approaches, the target task includes an operation object, the execution module includes a target component for controlling the operation object, and the sending unit includes a hardware abstraction layer.

[0036] The sending unit is used to convert the target task into control commands and send them to the execution module, including:

[0037] Retrieve the target task from the task queue;

[0038] Through the hardware abstraction layer, the target task is transformed into control instructions supported by the target component;

[0039] Control commands are sent to the target component via the target channel.

[0040] Secondly, an interactive control method is also provided, applied to an intent processing module in an interactive control system. The interactive control system further includes an intent understanding module and an execution module connected to the intent processing module. The intent understanding module includes at least one large model. The interactive control method includes:

[0041] Obtain the current intent from the output of any large model from the intent understanding module;

[0042] Determine the current attention score of at least a portion of the text fragments in the context of the current intent, and select the target text fragment from the aforementioned at least a portion of the text fragments in descending order of the current attention score;

[0043] Generate the target task based on the current intent and the target text fragment;

[0044] The target task is converted into control instructions and sent to the execution module so that the execution module can execute the control instructions.

[0045] Beneficial effects:

[0046] In the solution provided in this application embodiment, the interactive control system includes an intent understanding module, an intent processing module, and an execution module. The intent understanding module includes at least one large model. The intent understanding module sends the current intent output by any large model to the intent processing module. The intent processing module determines the current attention score of at least a portion of the text fragments in the context of the current intent, selects target text fragments from these text fragments in descending order of current attention score, generates a target task based on the current intent and the target text fragments, and converts the target task into a control command to be sent to the execution module. By designing the intent processing module, intent understanding and task execution can be decoupled, avoiding the need for integrators to reconstruct the entire system when the large model base supplier stops providing services, thus achieving flexibility in model switching. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram of an interactive control system provided in an embodiment of this application;

[0049] Figure 2 This is a timing diagram of the interactive control method provided in an embodiment of this application;

[0050] Figure 3 This is a schematic diagram of the tree structure provided in the embodiments of this application;

[0051] Figure 4 This is another schematic diagram of the interactive control system provided in the embodiments of this application.

[0052] Reference numerals: 10-Interactive control system, 11-Intent understanding module, 12-Intent processing module, 13-Execution module, 121-Conversion unit, 122-Context management unit, 123-Scheduling unit, 124-Sending unit. Detailed Implementation

[0053] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0054] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0055] "A and / or B" includes the following three combinations: A only, B only, and a combination of A and B.

[0056] The use of "applies to" or "configured to" in this application implies open and inclusive language, which does not exclude the applicability to or configuration of devices to perform additional tasks or steps. Additionally, the use of "based on" implies openness and inclusivity, because processes, steps, calculations, or other actions "based on" one or more conditions or values ​​may in practice be based on additional conditions or values ​​beyond those stated.

[0057] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0058] As mentioned earlier, the relevant large model architecture couples intent understanding and task execution into a single module. If the large model base supplier stops providing services (such as through iterative obsolescence), the integrator needs to reconstruct the entire system, and the model switching lacks flexibility.

[0059] This application provides an interactive control system and interactive control method, which aims to solve the problem of lack of flexibility in model switching.

[0060] In some embodiments, the interactive control system in this application includes, but is not limited to, a voice assistant system.

[0061] In some embodiments, the interactive control system in this application is a system mounted on a vehicle. The vehicle may be, for example, a vehicle or an aircraft.

[0062] Figure 1 This is a schematic diagram of an interactive control system provided in an embodiment of this application. For example... Figure 1 As shown, the interactive control system 10 includes an intent understanding module 11, an intent processing module 12, and an execution module 13. The intent processing module 12 is communicatively connected to both the intent understanding module 11 and the execution module 13. The intent understanding module 11 includes at least one large model. Each of these large models can be used to identify user intents. For example, the input to any one of these large models includes the user's current dialogue text, and the output includes the user intent identified based on the current dialogue text. In one example, the intent processing module 12 may be referred to as the intent layer.

[0063] The interaction process between the intent understanding module 11, the intent processing module 12, and the execution module 13 can be as follows: Figure 2 As shown. Among them, Figure 2 This is a timing diagram of the interactive control method provided in an embodiment of this application. For example... Figure 2 As shown, the interactive control method includes:

[0064] S201: The intent understanding module 11 sends the current intent output by any large model to the intent processing module 12;

[0065] S203: The intent processing module 12 determines the current attention score of at least a portion of the text fragments in the context of the current intent;

[0066] S205: The intent processing module 12 selects the target text segment from the above-mentioned at least some text segments in descending order of the current attention score;

[0067] S207: The intent processing module 12 generates a target task based on the current intent and the target text fragment;

[0068] S209: The intent processing module 12 converts the target task into control commands;

[0069] S211: The intent processing module 12 sends the control command to the execution module 13;

[0070] S213: Execution module 13 executes control instructions.

[0071] The solution provided in this application embodiment includes an intent understanding module, an intent processing module, and an execution module. The intent understanding module includes at least one large model. The intent understanding module sends the current intent output by any large model to the intent processing module. The intent processing module determines the current attention score of at least a portion of the text fragments in the context of the current intent, selects a target text fragment from these fragments in descending order of current attention score, generates a target task based on the current intent and the target text fragment, and converts the target task into a control command to be sent to the execution module. By designing the intent processing module, intent understanding and task execution can be decoupled, avoiding the need for integrators to reconstruct the entire system when the large model base supplier stops providing services, thus achieving flexibility in model switching. Moreover, by selecting the target text fragment from the aforementioned at least a portion of the text fragments in descending order of current attention score and generating a target task based on the current intent and the target text fragment, the consistency of intent during cross-model calls can be ensured.

[0072] Steps S201 to S213 will be explained below.

[0073] In step S201, the intent understanding module 11 sends the current intent output by any large model to the intent processing module 12. This current intent can be the user's intent identified by the large model based on the user's current dialogue text. Furthermore, the large model can identify the user's current intent based on the user's current dialogue text and historical dialogue text (such as the previous round of dialogue text).

[0074] It should be noted that the user's dialogue text, such as the current dialogue text and historical dialogue text, can be text entered by the user, or obtained by converting the user's input speech or image into text. In practice, ASR (automatic speech recognition) technology can be used for speech-to-text conversion, and OCR (optical character recognition) technology can be used for image-to-text conversion.

[0075] The current intent may include, for example, the intent source, task type, intent name, target, parameters, and original dialogue text (current dialogue text). Taking the large model that generates the current intent as Model A, the current dialogue text as "set it to 24 degrees", and the previous dialogue text as "set the air conditioner to 20 degrees", the current intent may include, for example, the following: intent source: "Model A", task type: "device control", intent name: "adjust temperature", target: "air conditioner", parameter: 24, original dialogue text: "set it to 24 degrees".

[0076] In step S203, the intent processing module 12 determines the current attention score of at least a portion of the text fragments in the context of the current intent.

[0077] Specifically, the intent processing module 12 can cache at least a portion of the text fragments from the context of the current intent and the historical attention scores of those text fragments. The context includes, but is not limited to, the text from the previous dialogue round. The at least a portion of the text fragments is obtained by segmenting the text using natural language processing techniques. The initial attention score of each text fragment in the context can be determined using a natural language processing model or a large model. This initial attention score can decay as the number of dialogue rounds increases.

[0078] The historical attention scores of at least some of the aforementioned text fragments can be either initial attention scores or attention scores that have decayed before the current dialogue round. In the current dialogue round, the cached historical attention scores of at least some of the aforementioned text fragments can be corrected to obtain the current attention score; this correction process includes decaying the historical attention scores. For example, the historical attention scores can be decayed according to a preset decay value. This decay value can be 0.05, 0.1, or 0.15, etc., and can be set according to actual needs; no specific limitation is made here.

[0079] Taking the context of the current intent as the previous dialogue text, and assuming the previous dialogue text was "Turn the air conditioner to 20 degrees," let's assume the text fragments in the previous dialogue text and their corresponding historical attention scores are: "Air conditioner": 0.6 points, "Turn to": 0.2 points, "20 degrees": 0.2 points. With a preset attenuation value of 0.1, the historical attention scores are attenuated to: "Air conditioner": 0.5 points, "Turn to": 0.1 points, "20 degrees": 0.1 points. In one example, the attenuated historical attention score can be used as the current attention score.

[0080] In practice, to ensure that text fragments from the previous dialogue that are relevant to the current intent are carried over into the following text, the historical attention score of the first text fragment in the previous dialogue can be increased; where the first text fragment is the text fragment from the previous dialogue that is relevant to the text fragment in the current dialogue. For example, the first text fragment can be determined in the previous dialogue based on a pre-defined tree structure. Figure 3 As shown, this tree structure includes multiple objects and connections between them. Figure 3 This is a schematic diagram of the tree structure provided in the embodiments of this application. Figure 3The diagram illustrates objects such as temperature, air conditioning, vehicle controls, windows, and seats. It should be understood that the objects in the tree structure and the connections between them can be set according to actual needs, and no specific limitations are made here. Taking the current dialogue text "Set it to 24 degrees" as an example, "24 degrees" corresponds to the "temperature" object in the tree structure. Based on the "air conditioning" object connected to the "temperature" object, we can see that "air conditioning" in the previous dialogue text is related to "24 degrees" in the current dialogue text, thus identifying "air conditioning" in the previous dialogue text as the first text segment.

[0081] Furthermore, in one embodiment, the above-mentioned correction process further includes: increasing the historical attention score of the first text segment after attenuation. For example, the historical attention score of the first text segment after attenuation can be increased according to a preset increase value. This increase value can be 0.05, 0.1, or 0.15, etc., and can be set according to actual needs, without being specifically limited here. According to the description above, the score of "air conditioner" in the previous round of dialogue text attenuated from 0.6 to 0.5; with a preset increase value of 0.1, the historical attention score of "air conditioner" can be increased from 0.5 to 0.6.

[0082] In another embodiment, the historical attention scores of the cached at least some of the text segments can be corrected to obtain the current attention score; wherein, the correction process includes: increasing the historical attention score of the first text segment among the at least some of the text segments, and decreasing the historical attention scores of the other text segments among the at least some of the text segments besides the first text segment.

[0083] In step S205, the intent processing module 12 selects the target text segment from the aforementioned at least some text segments in descending order of current attention scores. For example, the text segment with the highest current attention score can be selected as the target text segment from the aforementioned at least some text segments.

[0084] In step S207, the intent processing module 12 generates a target task based on the current intent and the target text fragment. The target task may include the intent source, task type, delay label (hereinafter referred to as the target delay label), a deadline set according to the target delay label, the operation object, the operation action, parameters, the original intent (current intent), and the target text fragment. Further, the target task may also include a task identifier, a decay value, and / or the current attention score of the target text fragment. The operation action can be determined based on the intent name. The target delay label can be determined based on the operation object and the operation action.

[0085] The latency labels in this application can be indicators from QoS (Quality of Service) technology. QoS is a technology used to manage and guarantee the performance of network services, typically involving indicators such as latency, bandwidth, jitter, and packet loss rate. In one example, the latency labels in this application can be referred to as priorities. The latency labels in this application can be categorized, for example, as low latency, medium latency, and high latency. Low latency can be represented by P0, medium latency by P1, and high latency by P2. The latency requirement corresponding to P0 can be, for example, "<20ms", the latency requirement corresponding to P1 can be, for example, "20-200ms", and the latency requirement corresponding to P2 can be, for example, ">200ms". Here, ms is the unit, representing milliseconds.

[0086] In one implementation, the intent processing module 12 can use a preset structured data format to convert the current intent into a structured task, and then add the target text fragment to the structured task to obtain the target task. The structured data format may include fields representing the intent source, task type, delay label, deadline, operation object, operation action, and parameters. Furthermore, the structured data format may also include fields representing the task identifier. It should be noted that by using a preset structured data format to convert the current intent into a structured task, the entire process can be based on a unified structured data format, providing high flexibility and scalability for future integration with new large models or hardware devices. In addition, when switching models, carrying a high-attention text fragment as additional input can effectively avoid intent gaps and achieve cross-domain context inheritance.

[0087] Furthermore, the intent processing module 12 can determine the operation action based on the intent name in the current intent, and determine the target delay label based on the operation object and the operation action. Then, it generates a structured task using a preset structured data format based on the current intent and the target delay label. Specifically, the intent name can be standardized according to a preset standardization strategy to obtain the operation action. Additionally, the operation object can be standardized according to the same standardization strategy, and then the target delay label can be determined based on the standardized operation object and the operation action.

[0088] Furthermore, based on the operation object and operation action, the target delay label is determined, including: based on the operation object and operation action, the target security coefficient and target user experience coefficient are determined; based on the target security coefficient and target user experience coefficient, the target delay label is determined.

[0089] In practice, user experience coefficient grading requirements and safety coefficient grading requirements can be set. For example, user experience coefficient grading requirements can be set based on the perceptibility of the cockpit, combined with the operating object and operating action; similarly, safety coefficient grading requirements can be set based on the impact on cockpit and driving safety, combined with the operating object and operating action. Based on this, target safety coefficients and target user experience coefficients can be determined according to the operating object and operating action, as well as the set user experience coefficient grading requirements and safety coefficient grading requirements. Then, target delay labels can be determined based on the target safety coefficients and target user experience coefficients, and the pre-defined correspondence between the combination of safety coefficients and user experience coefficients and delay labels. In one example, this correspondence can be shown in Table 1.

[0090] Table 1: Exemplary correspondence between combinations of security and user experience factors and delay labels

[0091]

[0092] In Table 1, both the security coefficient and the user experience coefficient are categorized as 1, 2, 3, and 4. For the security coefficient, a higher value indicates a higher security requirement. For the user experience coefficient, a higher value indicates a higher level of user attention. As shown in Table 1, when both the security coefficient and the user experience coefficient are 4, the latency label is P0 (low latency); when both are 4, the latency label is P1 (medium latency); and when both are 4, the latency label is P2 (high latency). The correspondence between other combinations of security and user experience coefficients and latency labels can be deduced from the relevant content above and will not be elaborated upon here.

[0093] It should be noted that the content shown in Table 1 is exemplary. The values ​​of the security coefficient and user experience coefficient, as well as the correspondence between the combination of the security coefficient and user experience coefficient and the delay label, can be set according to actual needs, and no specific limitations are made here.

[0094] In step S209, the intent processing module 12 converts the target task into control instructions. The intent processing module 12 may include a hardware abstraction layer, and the execution module 13 includes a target component for controlling the operation object in the target task. The intent processing module 12 can use the hardware abstraction layer to convert the target task into control instructions supported by the target component, i.e., control instructions that the target component can understand. For example, when the operation object is an air conditioner, the target component can be an infotainment system. The hardware abstraction layer can uniformly encapsulate low-level instructions, shielding differences between different hardware interfaces.

[0095] In step S211, the intent processing module 12 sends a control command to the execution module 13, causing the execution module 13 to execute the control command in step S213. Further, in step S211, the control command is sent to the target component in the execution module 13, causing the target component to execute the control command in step S213.

[0096] In some embodiments, such as Figure 4 As shown, the intent processing module 12 includes a conversion unit 121, a context management unit 122, and a sending unit 124. Further, the intent processing module 12 may also include a scheduling unit 123. Figure 4 This is another schematic diagram of the interactive control system provided in this application embodiment. The conversion unit 121 is used to convert the large model intent input, or to require the supplier to encapsulate the interface according to the requirements of the conversion unit 121. The input of the conversion unit 121 includes the intent output by the large model. The scheduling unit 123 can perform unified hierarchical scheduling and resource allocation for the intent output by the large model. The sending unit 124 can implement control command issuance through a hardware abstraction layer, designing a multi-channel architecture. The context management unit 122 can segment and parameterize the preceding text, passing it to the following text; segment the content of requests with limited time and length, assign attention scores, and gradually decay the attention scores according to the rule of time decay. The preceding paragraphs with high attention scores are passed to the scheduling unit 123 as additional content for execution for intent inheritance and task distribution judgment. Through the designed four major units—conversion unit 121, context management unit 122, scheduling unit 123, and sending unit 124—end-to-end decoupling from intent input to execution can be achieved. In one example, the conversion unit 121 can be called the object abstraction layer, the scheduling unit 123 can be called the scheduling layer, and the sending unit 124 can be called the execution layer.

[0097] Specifically, the conversion unit 121 is used to convert the current intent into a structured task using the structured data format described above, and then send the structured task to the context management unit 122. The context management unit 122 is used to determine the current attention score of at least some of the text fragments, and select target text fragments from the at least some text fragments in descending order of current attention score to add them to the structured task, thus obtaining the target task. The sending unit 124 is used to convert the target task into control instructions and send them to the execution module 13.

[0098] In one implementation, the intent processing module 12 includes a scheduling unit 123, and the sending unit 13 adopts a multi-channel architecture. The scheduling unit 123 is configured to: determine the target scheduling policy corresponding to the target delay label based on the target delay label and the preset correspondence between delay labels and scheduling policies; determine the target channel in the multi-channel architecture based on the target scheduling policy; and add the target task to the task queue corresponding to the target channel. Further, the target channel can be determined based on the target scheduling policy and the task type and operation object in the target task.

[0099] The multi-channel architecture may include, for example, a first channel, a second channel, and a third channel. The first channel is the real-time kernel of a real-time operating system (such as QNX). The second channel is a general-purpose operating system (including but not limited to Android). The third channel is a batch data upload channel. QNX is a real-time operating system widely used in embedded systems. It should be noted that the object (also called a component) corresponding to the first channel may include, for example, an ECU (electronic control unit) controller, and the object corresponding to the second channel may include, for example, an infotainment system. The ECU controller and the infotainment system are contained in the execution module 13. This achieves hardware resource isolation.

[0100] The correspondence between the preset delay labels and scheduling strategies can be shown in Table 2.

[0101] Table 2: Exemplary Correspondence Between Delay Labels and Scheduling Policies

[0102]

[0103] Table 2 shows the correspondence between delay labels and scheduling strategies, as well as the delay requirements and typical scenarios corresponding to each delay label. As shown in Table 2, the delay requirement for delay label P0 is "<20ms," and typical scenarios applicable to it include collision warnings and DMS (driver monitoring system) alarms. The scheduling strategy is "preemptive scheduling + hardware pass-through," which instructs tasks to be placed in the task queue corresponding to the first channel. Tasks in this queue can be transmitted through the first channel in a preemptive manner to meet the <20ms delay requirement. The delay requirement for delay label P1 is "20-200ms," and typical scenarios applicable to it include voice commands and touch responses. The scheduling strategy is "priority queue + reserved GPU (graphics processing unit) acceleration," which instructs tasks to be placed in the task queue corresponding to the second channel. Tasks in this queue can be transmitted through the second channel using reserved GPU resources to meet the 20-200ms delay requirement. The reserved GPU resources may include, for example, a target percentage of video memory; the target percentage can be 9%, 10%, or 11%, etc., and can be set according to actual needs, without specific limitations here. The latency requirement for latency label P2 is ">200ms", and typical applicable scenarios include map updates and log uploads. The scheduling strategy is "idle resource aggregation + breakpoint resumption after preemption". This scheduling strategy can instruct tasks to be placed in the task queue corresponding to the third channel. Tasks in this task queue can be transmitted through the third channel using idle resources and breakpoint resumption after preemption to meet the latency requirement of >200ms.

[0104] It should be noted that the correspondence between latency labels and scheduling policies shown in Table 2 is an exemplary correspondence, which can be set according to actual needs and is not specifically limited here. Furthermore, the related interface calls only pass fixed parameters, which cannot dynamically adapt to the latency sensitivity of tasks (e.g., real-time interaction requires low latency, while log uploads can tolerate high latency), leading to wasted GPU resources or timeouts for critical tasks. The solution provided in this application can allocate GPU resource pools according to the latency labels of tasks (e.g., reserving 10% of video memory for low-latency tasks).

[0105] In one implementation, the target task further includes the current attention score of the target text segment; the scheduling unit 123 is used to determine the target scheduling strategy corresponding to the target delay label based on the target delay label and the preset correspondence between the delay label and the scheduling strategy, including: determining the intent source switching situation based on the current attention score of the target text segment; and determining the target scheduling strategy based on the target delay label and the correspondence if it is determined that the intent source will not be switched.

[0106] In determining the intent source switching based on the current attention score of the target text fragment, as one implementation, if the target text fragment corresponds to a specific intent source, and this specific intent source is different from the intent source in the target task, then if the current attention score is greater than a threshold and there is currently no intent source with a higher priority than this specific intent source occupying system resources, it is determined to switch to this specific intent source. The scheduling unit 123 can then interrupt the processing of the target task and send an intent source switching request to the intent understanding module 11, causing the intent understanding module 11 to switch to the large model indicated by this specific intent source for intent understanding. Alternatively, if any of the following conditions exist, it is determined not to switch the intent source, and then the target scheduling strategy is determined based on the target delay label and the correspondence between the delay label and the scheduling strategy: the target text fragment does not correspond to a specific intent source; the target text fragment corresponds to a specific intent source, and this specific intent source is the same as the intent source in the target task; the target text fragment corresponds to a specific intent source, this specific intent source is different from the intent source in the target task, and the current attention score is less than or equal to a threshold, and / or there is currently an intent source with a higher priority than this specific intent source occupying system resources.

[0107] It should be noted that determining the switching of intent sources based on the current attention score of the target text segment helps ensure the accuracy of intent understanding and reduces the risk of semantic bias.

[0108] In one implementation, after determining to switch to the aforementioned specific intent source, the scheduling unit 123 can also send an attention score correction request to the context management unit 122. This attention score correction request can indicate the relevant status of the current intent, such as indicating a switch to a larger model for intent understanding. The context management unit 122 can respond to the attention score correction request, determine whether to perform attention score decay and / or increase, and execute the corresponding attention score correction operation when it is determined that attention score decay and / or increase are necessary. It should be noted that dynamically adjusting the attention score ensures accurate inheritance of the contextual semantics under the new intent source, avoids resource waste, and provides a reliable basis for subsequent task scheduling.

[0109] In one implementation, to avoid processing intents from illegal intent sources, if it is determined that the intent source will not be switched, the scheduling unit 123 can also determine whether it is necessary to block the intent source in the target task according to a preset control policy. For example, the control policy includes a source blacklist; if the intent source in the target task is included in the source blacklist, then it is determined that the intent source in the target task should be blocked. Additionally, if the intent understanding module 11 includes multiple large models, when these multiple large models process intent understanding in parallel, other intent sources besides the intent source in the target task can be blocked. If it is determined that the intent source in the target task needs to be blocked, the current interaction control flow ends. If it is determined that the intent source in the target task does not need to be blocked, then the target scheduling policy corresponding to the target delay label is determined according to the target delay label and the preset correspondence between delay labels and scheduling policies.

[0110] As described above, scheduling unit 123 can add the target task to the task queue corresponding to the target channel. Then, sending unit 124 can retrieve the target task from the task queue, convert it into control instructions supported by the target component through the hardware abstraction layer, and send the control instructions to the target component through the target channel. For example, a function related to the control of the operand in the hardware abstraction layer can be called, using the parameters from the target task as input parameters to that function, thus using that function with those parameters as input parameters as the control instructions.

[0111] The following example illustrates the interactive control method provided in this application. The scenario is set as follows:

[0112] Interactive control system 10: a voice assistant system, including an intent understanding module 11, an intent processing module 12 and an execution module 13. The intent understanding module 11 includes at least one large model. The intent processing module 12 includes a conversion unit 121, a context management unit 122, a scheduling unit 123 and a sending unit 124.

[0113] User: I am driving the vehicle and using the voice assistant system to control the cabin environment;

[0114] The current dialogue text corresponding to the current intent: "Adjust it to 24 degrees";

[0115] Previous dialogue text: "Turn the air conditioner to 20 degrees Celsius";

[0116] Large model: A large model of speech understanding operating in the cockpit (e.g., Model A).

[0117] In the scenario described above, Model A outputs the current intent based on the current dialogue text "set it to 24 degrees" and the previous dialogue text "set the air conditioner to 20 degrees". The current intent includes: intent source: "Model A", task type: "device control", intent name: "adjust temperature", operation object: "air conditioner", parameter: 24, and original dialogue text: "set it to 24 degrees". The intent understanding module 11 sends the current intent output by Model A to the conversion unit 121.

[0118] The conversion unit 121 receives the current intent from Model A and converts it into a unified, structured task object, which can be called a structured task. Specifically, the conversion unit 121 can standardize the intent name "adjust temperature" and the operation object "air conditioner," and use the standardized intent name "set temperature" as the operation action. Based on the operation object "air conditioner" and the operation action "set temperature," the target delay label P1 is automatically injected to obtain the structured task. This structured task may include: task identifier: "task-12345," intent source: "Model A," task type: "device control," delay label: "P1," deadline: 200, operation action: "set temperature," operation object: "air conditioner," parameter: 24, original intent: ###. Here, "###" represents the content of the current intent. Afterward, the conversion unit 121 sends the structured task to the context management unit 122.

[0119] The context management unit 122 receives the structured task, and retrieves the contextual information of the current intent from its internal cache, such as the previous dialogue text, and obtains the initial attention scores of each text segment in the previous dialogue text. Then, the context management unit 122 can refine the initial attention scores of each text segment to obtain the current attention score for each text segment. Finally, the context management unit 122 can attach the target text segment with the highest current attention score to the current structured task to obtain the target task.

[0120] Assume the current attention scores of each text fragment in the previous round of dialogue are: "Air conditioner": 0.6 points, "On to": 0.2 points, "20 degrees": 0.2 points. The context management unit 122 can specifically attach "air conditioner" and its current attention score "0.6 points" to the current structured task. Since the "it" in the current dialogue text has an ambiguous referent, the context management unit 122 can correct and enrich the understanding of the current task by using the high-attention text fragment "air conditioner" from the previous round as the inherited context. The target task obtained by the context management unit 122 may include: task identifier: "task-12345", intent source: "A model", task type: "device control", delay label: "P1", deadline: 200, operation action: "set temperature", operation object: "air conditioner", parameter: 24, high-attention text fragment: (text fragment: "air conditioner", attention score: 0.6), original intent: ###. The context management unit 122 sends the target task to the scheduling unit 123.

[0121] The scheduling unit 123 receives the target task with an attached context from the context management unit 122. The operation of the scheduling unit 123 includes:

[0122] 1. Determine the target scheduling strategy based on the target delay label P1 in the target task;

[0123] 2. Query current system resources (such as GPU / CPU (central processing unit) load);

[0124] 3. Place the target task into the appropriate task queue.

[0125] Specifically, the target delay label P1 indicates a medium delay. According to Table 2 as described above, the target task needs to enter the priority queue. The scheduling unit 123 can perform resource checks. If it finds that the current Android system resources are sufficient, there is no need to preempt GPU resources. Taking the second channel as the Android channel as described above as an example, the scheduling unit 123 can also select a channel, for example, determining the target channel as the Android channel based on the task type "device control" and the operation object "air conditioner". Afterwards, the scheduling unit 123 can deliver the target task to the task queue corresponding to the Android channel.

[0126] The sending unit 124 retrieves the target task from the task queue corresponding to the Android channel, and then converts the target task into a control instruction that the underlying hardware can understand through the hardware abstraction layer. Specifically, the sending unit 124 can parse the operation object, operation action, and parameters from the target task. Based on the operation object and operation action, it calls the function related to air conditioning control in the hardware abstraction layer, such as AirConditionerHAL::set_temperature(), and uses the parameter "24" as the input parameter of the function to obtain the control instruction "AirConditionerHAL::set_temperature(24)". Then, the sending unit 124 can send the control instruction to the execution module 13 through the Android channel. Here, "AirConditionerHAL" can be understood as a class used for air conditioning control, and "set_temperature()" is a member function of this class used to set the temperature.

[0127] After receiving the control command, the air conditioner controller in execution module 13 adjusts the air conditioner temperature from 20 degrees to 24 degrees. Furthermore, execution module 13 can return hardware operation success information to sending unit 124.

[0128] In one implementation, after receiving the hardware operation success information returned by the execution module 13, the sending unit 124 may return a brief execution status report to the intent processing module 12. This execution status report may include the task identifier of the target task, the execution status indicating successful execution, and a timestamp, etc.

[0129] The examples set in the above scenarios clearly demonstrate the core value of the intent processing module 12 of this application:

[0130] 1. Decoupling: The intent output of the large model is clearly separated from the final hardware execution;

[0131] 2. Contextual inheritance: The key information "air conditioner" was successfully inherited from the previous dialogue text, enabling the system to correctly understand the reference of "it";

[0132] 3. Dynamic scheduling: The system automatically selects the appropriate execution path and resources based on the task's priority (P1 priority);

[0133] 4. Standardization: The entire process is based on a unified structured data format, providing high flexibility and scalability for future integration with new large models or hardware devices.

[0134] As described above, the intent processing module 12 can be referred to as the intent layer, and the conversion unit 121, scheduling unit 123, and sending unit 124 can be referred to as the intent abstraction layer, scheduling layer, and execution layer, respectively. The solution provided in this application achieves model-independent intent processing through the division of labor among the intent abstraction layer, context management unit, scheduling layer, and execution layer, supporting dynamic vendor switching. A context segmentation mechanism based on attention decay ensures semantic coherence during cross-model calls, resolving the context discontinuity problem in related technologies. Pre-allocation of video memory (e.g., reserving 10% of resources for P0 tasks) and preemptive scheduling improve the response stability of real-time tasks.

[0135] The technical effects of this application are derived through reasoning as follows:

[0136] Supplier neutrality: Through the structured output of the intent layer, integrators can seamlessly switch between large models and avoid code refactoring; decoupling intent understanding and execution facilitates the management and switching of base suppliers; the large model evolves rapidly, and by controlling the middle layer (intent layer), the car manufacturer can retain control while maintaining competitiveness.

[0137] Improved accuracy of context inheritance: Attention-based decay mechanism prioritizes the inheritance of high semantic segments (such as "air conditioner") in subsequent instructions, reducing ambiguity (such as avoiding "open to" being misinterpreted as "arrive by car").

[0138] Resource utilization optimization: Tiered scheduling strategies enable more rational allocation of GPU resources (e.g., P2 tasks only use idle video memory);

[0139] Middle layer control: The new generation of electronic and electrical architecture atomically encapsulates execution services and also atomically encapsulates intents. However, without the control of the middle layer, it is not conducive to the free access of intents and execution services and application upgrades.

[0140] Iterable architecture: It combines low-latency scheduling for large models with high-latency scheduling to meet the needs of future application evolution.

[0141] This application also provides a means of transportation, including the interactive control system 10 as described above. In one example, the means of transportation includes a vehicle.

[0142] The interactive control system and interactive control method provided in the embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An interactive control system, characterized in that, It includes an intent understanding module, an intent processing module, and an execution module; The intent understanding module includes at least one large model, used to send the current intent output by any one of the large models to the intent processing module. The current intent includes an intent source, which is used to indicate the large model that generated the current intent. The intent processing module is used to determine the current attention score of at least a portion of the text segments in the context of the current intent, select a target text segment from the at least a portion of the text segments in descending order of the current attention score, and generate a target task based on the current intent and the target text segment. The target task includes the intent source, the target delay label, the target text segment, and the current attention score of the target text segment. The intent processing module includes a scheduling unit and a sending unit, and the sending unit adopts a multi-channel architecture. The scheduling unit is configured to determine the intent source switching situation based on the current attention score of the target text segment, and if it is determined that the intent source will not be switched, determine the target scheduling strategy corresponding to the target delay label based on the target delay label and the correspondence between the preset delay label and the scheduling strategy, determine the target channel in the multi-channel architecture based on the target scheduling strategy, and add the target task to the task queue corresponding to the target channel. The sending unit is used to obtain the target task from the task queue, convert the target task into control instructions, and send them to the execution module; The execution module is used to execute the control instructions.

2. The interactive control system according to claim 1, characterized in that, The intent processing module is used to generate a target task based on the current intent and the target text fragment, including: The current intent is converted into a structured task using a preset structured data format. The target text fragment is added to the structured task to obtain the target task.

3. The interactive control system according to claim 2, characterized in that, The current intent includes the operation object and the intent name; The intent processing module is used to convert the current intent into a structured task using a preset structured data format, including: Determine the action based on the stated intent name; The target delay label is determined based on the operation object and the operation action; The structured task is generated using the structured data format, based on the current intent and the target delay label.

4. The interactive control system according to claim 3, characterized in that, The intent processing module is used to determine the target delay label based on the operation object and the operation action, including: Based on the operation object and the operation action, determine the target security coefficient and the target user experience coefficient; The target delay label is determined based on the target security coefficient and the target user experience coefficient.

5. The interactive control system according to claim 1, characterized in that, The intent processing module is used to determine the current attention score of at least a portion of the text fragments in the context of the current intent, including: The historical attention scores of at least a portion of the cached text segments are corrected to obtain the current attention score; wherein the correction process includes: attenuating the historical attention scores.

6. The interactive control system according to claim 2, characterized in that, The intent processing module includes: A conversion unit is used to convert the current intent into the structured task using the structured data format; The context management unit is configured to determine the current attention score of the at least some text segments, select the target text segment from the at least some text segments in descending order of the current attention score, and add the target text segment and the current attention score of the target text segment to the structured task to obtain the target task.

7. The interactive control system according to claim 1, characterized in that, The target task also includes an operation object, the execution module includes a target component for controlling the operation object, and the sending unit includes a hardware abstraction layer; The sending unit is used to convert the target task into the control command and send it to the execution module, including: The target task is converted into control instructions supported by the target component through the hardware abstraction layer. The control command is sent to the target component through the target channel.

8. An interactive control method, characterized in that, An intent processing module is applied in an interactive control system as described in any one of claims 1-7, wherein the interactive control system further includes an intent understanding module and an execution module connected to the intent processing module, wherein the intent understanding module includes at least one large model; the interactive control method includes: Obtain the current intent from any of the outputs of the large model from the intent understanding module; Determine the current attention score of at least a portion of the text fragments in the context of the current intent, and select the target text fragment from the at least a portion of the text fragments in descending order of the current attention score; Generate a target task based on the current intent and the target text fragment; The target task is converted into control instructions and sent to the execution module so that the execution module executes the control instructions.

Images