Intention recognition method and device based on large model thought chain reasoning, storage medium and electronic equipment

Abstract

This application discloses an intent recognition method, device, storage medium, and electronic device based on large-model thinking chain reasoning, relating to the field of smart home technology. The intent recognition method based on large-model thinking chain reasoning includes: acquiring semantic elements corresponding to the instruction information of the target object, and inputting the semantic elements and prompt words indicating the sample reasoning process information of the first model into the first model, so that the first model outputs the intent set corresponding to the instruction information and the reasoning process information of the first model; determining a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score; verifying each intent information according to the comprehensive score of each intent information to obtain a verification result; and determining the target intent set based on the verification result, the reasoning process information, and the intent set.

Description

Technical Field

[0001] This application relates to the field of smart home technology, and more specifically, to an intent recognition method and apparatus, storage medium and electronic device based on large model thinking chain reasoning. Background Technology

[0002] With the widespread adoption of smart home systems, voice interaction has become the primary entry point for users to control home appliances. Users expect to complete the coordinated operation of multiple devices at once through natural language commands (such as "I'm a bit hot, set the living room air conditioner to 26 degrees, and also dim the bedroom lights"), achieving the convenient experience of "controlling the whole house with just one sentence." However, current mainstream voice assistants generally suffer from inaccurate intent recognition when processing such complex, multi-intent voice commands, seriously affecting user experience and system reliability.

[0003] In related technologies, user commands are matched using manually predefined grammatical templates (such as "open [A] and adjust [B]"), and device and parameters are extracted by slot filling. However, this method relies heavily on fixed sentence patterns, and when faced with diverse expressions in natural language (such as "It's too hot, help me save some electricity" or "The living room is too bright, can you dim it?"), it is prone to parsing failure due to semantic deviation from the template.

[0004] There is still no effective solution to the problem of low accuracy in recognizing the intent of instructions in related technologies. Summary of the Invention

[0005] This application provides an intent recognition method, apparatus, storage medium, and electronic device based on large model thought chain reasoning, in order to at least solve the problem of low accuracy in intent recognition of instruction information in related technologies.

[0006] According to one embodiment of this application, an intent recognition method based on large model thought chain reasoning is provided, comprising: acquiring semantic elements corresponding to instruction information of a target object, and inputting the semantic elements and prompt words indicating sample reasoning process information of a first model into a first model, so that the first model outputs an intent set corresponding to the instruction information and reasoning process information of the first model; determining a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score; verifying each intent information according to the comprehensive score of each intent information to obtain a verification result; and determining a target intent set according to the verification result, the reasoning process information, and the intent set.

[0007] According to another embodiment of this application, an intent recognition device based on large model thought chain reasoning is also provided, comprising: a first determining module, configured to acquire semantic elements corresponding to instruction information of a target object, and input the semantic elements and prompt word information indicating sample reasoning process information of a first model into the first model, so that the first model outputs an intent set corresponding to the instruction information and reasoning process information of the first model; a second determining module, configured to determine a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score; a verification module, configured to verify each intent information according to the comprehensive score of each intent information to obtain a verification result; and a third determining module, configured to determine a target intent set according to the verification result, the reasoning process information, and the intent set.

[0008] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, and the computer program is configured to execute the above-described intention recognition method based on large model thought chain reasoning at runtime.

[0009] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described intention recognition method based on large model thought chain reasoning through the computer program.

[0010] In this embodiment, semantic elements are first extracted from the user's original voice commands. These semantic elements, along with a set of preset sample reasoning prompts, are then input into a first model. The model outputs an intent set and reasoning process information accordingly. After obtaining the original intent set, a multi-dimensional comprehensive scoring mechanism is used to assess the quality of each intent. The comprehensive score is calculated based on a reasonableness score and / or a device accessibility score and / or a state consistency score. After scoring, a verification process is performed on each intent. The verification results, along with the reasoning process information generated by the model, serve as the decision-making basis, ultimately determining the target intent set. This solves the problem of low intent recognition accuracy for command information in related technologies, significantly improving the intent recognition accuracy and system execution reliability under complex commands. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the hardware environment for an intent recognition method based on large model thinking chain reasoning according to an embodiment of this application;

[0014] Figure 2 This is a flowchart of an intent recognition method based on large model thought chain reasoning according to an embodiment of this application;

[0015] Figure 3 This is a schematic diagram of an intent recognition method based on large model thinking chain reasoning according to an embodiment of this application;

[0016] Figure 4 This is a structural block diagram of an intent recognition device based on large model thinking chain reasoning according to an embodiment of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] According to one aspect of the embodiments of this application, an intent recognition method based on large-model thinking chain reasoning is provided. This intent recognition method based on large-model thinking chain reasoning is widely applicable to digital control scenarios such as smart homes, smart home ecosystems, whole-house intelligence, digital homes, networked homes, automated homes, electronic homes, digital homes, and digital homes. It relies on technologies such as the Internet of Things (IoT), cloud computing, artificial intelligence (AI), digital technology, digital twins, and AR / VR to construct intelligent, automated, personalized, and customized digital ecosystems. Optionally, in this embodiment, the above-mentioned intent recognition method based on large-model thinking chain reasoning can be applied to, for example... Figure 1 The hardware environment shown consists of terminal device 102 and server 104. For example... Figure 1 As shown, server 104 is connected to terminal device 102 via a network and can be used to provide services (such as application services) to the terminal or clients installed on the terminal. A database can be set up on the server or independently of the server to provide data storage services for server 104. Cloud computing and / or edge computing services can be configured on the server or independently of the server to provide data processing services for server 104.

[0020] The aforementioned network communication architecture includes, but is not limited to, two types of links: wired transmission and wireless connection. Among them, the wired network can cover wide area networks, metropolitan area networks, and local area networks, while the wireless network supports communication standards such as WIFI, Bluetooth, and 5G. It can adapt to the digital technology transmission needs of the Internet of Things (IoT), the Internet of Things, and the Internet of Things blockchain, ensuring stable interaction of digital information, enabling devices to make autonomous, accurate, precise, and accurate judgments and adjustments, and supporting personalized and customized intelligent capabilities such as voice interaction, remote control, and self-control. Terminal devices 102 include, but are not limited to, PCs, mobile phones, tablets, and all categories of digital home appliances, digital electrical appliances, smart home appliances, and whole-house smart devices. Specifically, they include: smart air conditioners, air conditioning equipment, smart range hoods, smart refrigerators, smart ovens, grilling and cooking equipment, smart stoves, smart washing machines, washing, rinsing, drying and garment processing equipment, smart water heaters, smart washing equipment, smart dishwashers, smart projectors, smart TVs, smart clothes racks, smart curtains, smart audio-visual equipment, smart sockets, smart speakers, smart speakers, smart fresh air systems, smart kitchen and bathroom systems, kitchens, cabinets, kitchen cabinet equipment, smart bathroom equipment, robot vacuum cleaners, robot window cleaners, robot mopping, air purifiers, smart steam ovens, smart microwave ovens, smart water heaters, smart air purifiers, smart water dispensers, smart door locks, smart lighting fixtures, smart light bulbs, smart air detectors, smart remote controls, and smart gateways. They also cover smart terminals for security, lighting, and other digital home furnishings.

[0021] This embodiment provides an intent recognition method based on large-model thought chain reasoning, which is applied to the aforementioned terminal device or server. Figure 2 This is a flowchart of an intent recognition method based on large-model thought chain reasoning according to an embodiment of this application. The process includes the following steps:

[0022] Step S202: Obtain the semantic elements corresponding to the instruction information of the target object, and input the semantic elements and the prompt words indicating the sample reasoning process information of the first model into the first model, so that the first model outputs the intent set corresponding to the instruction information and the reasoning process information of the first model;

[0023] Step S202 extracts semantic elements from the user-input instruction information (e.g., original voice instructions). The semantic elements include: action predicates (e.g., "turn on" or "turn down"), target devices (e.g., "living room air conditioner" or "bedroom light"), parameter values ​​(e.g., "26℃" or "50% brightness"), and implicit context (e.g., "I'm a little hot") and other structured semantic components.

[0024] The sample reasoning process information contains one or more example reasoning chains, such as: "The user says 'I'm hot, turn the air conditioner down a bit, and dim the lights,' indicating that the user wants to improve comfort. It should be broken down into: 1) 'Adjust the air conditioner temperature to a lower value' corresponding device: living room air conditioner, parameter: 24℃; 2) 'Reduce the lighting brightness' corresponding device: bedroom light, parameter: 40%. The two do not conflict and can be executed in parallel."

[0025] Step S204: Determine a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score;

[0026] Step S206: Verify each intent information based on the comprehensive score of each intent information to obtain a verification result;

[0027] Optionally, if the overall score is lower than a preset threshold (e.g., 0.6), it is marked as "low confidence intent".

[0028] Step S208: Determine the target intent set based on the verification result, the reasoning process information, and the intent set.

[0029] Through the above steps, semantic elements are first extracted from the user's original voice commands. These semantic elements, along with a set of preset sample reasoning prompts, are then input into the first model. Based on this, the model outputs an intent set and reasoning process information. After obtaining the original intent set, a multi-dimensional comprehensive scoring mechanism is used to assess the quality of each intent. The comprehensive score is calculated based on a reasonableness score and / or a device accessibility score and / or a state consistency score. After scoring, a verification process is performed on each intent. The verification results, together with the reasoning process information generated by the model, serve as the decision-making basis, ultimately determining the target intent set. This solves the problem of low intent recognition accuracy for command information in related technologies, significantly improving the intent recognition accuracy and system execution reliability under complex commands.

[0030] In an exemplary embodiment, determining a comprehensive score for each intent information in the intent set includes: determining an indicator score for each intent information in the intent set based on an evaluation index, wherein the indicator score includes at least one of the following: a reasonableness score, a device accessibility score, and a state consistency score; determining a fuzzy correction factor corresponding to each indicator score; determining a first number of conflicting intent information in the intent set and a second number of total intent information in the intent set, and determining a conflict compensation factor based on the ratio of the first number to the second number; and determining a comprehensive score for each intent information based on the indicator score corresponding to each evaluation index, the corresponding fuzzy correction factor, and the conflict compensation factor.

[0031] It should be noted that the reasonableness score is used to indicate whether the schematic diagram information conforms to the semantic intent of the instruction information.

[0032] The equipment reachability score is used to indicate whether the target equipment is in a controllable state, and is based on a comprehensive evaluation of real-time status parameters such as equipment online status and equipment collaboration information.

[0033] State consistency score is used to indicate whether the executed schematic diagram contradicts or is redundant with the current environmental state. For example, if the current air conditioner temperature is 23℃, and the intention command is "adjust to 24℃", then the state consistency score = 0.9 (reasonable fine-tuning); if the air conditioner is currently off, and the command is "turn off the air conditioner", then the state consistency score = 0.3 (redundant operation); if the command is "turn on the air conditioner" and the ambient temperature is 32℃, then the state consistency score = 0.95 (necessary operation).

[0034] Optionally, in this embodiment of the application, a corresponding fuzzy correction factor is determined for each evaluation index. The fuzzy correction factor is used to dynamically adjust the weight of the index score to address boundary value or uncertainty scenarios. The fuzzy correction factor is determined based on the following formula for each index score: Where k∈[1,h], and h is the number of evaluation indicators. S is the dimensional fuzziness coefficient. k The score for the k-th evaluation indicator. This is the fuzzy correction factor corresponding to the index score of the kth evaluation index.

[0035] The fuzzy correction factor is determined based on historical operational data and a pre-set fuzzy rule base using expert experience. It can be dynamically updated. For example, the fuzzy correction factor corresponding to the rationality score... Fuzzy correction factor corresponding to equipment accessibility score State consistency score, state consistency score .

[0036] In this embodiment, the conflict compensation factor is used to indicate the residual risk after conflict resolution. Assuming the total number of intents in the intent set is N (second quantity), and the number of intents identified as having semantic or execution conflicts is M (first quantity), then: conflict compensation factor = M / N. When there is no conflict, the conflict compensation factor = 0. The more conflicts there are and the lower the reliability of resolution, the larger the conflict compensation factor and the lower the confidence.

[0037] The embodiments of this application significantly improve the robustness and accuracy of intent recognition in complex, ambiguous, and ever-changing home scenarios.

[0038] In an exemplary embodiment, determining a comprehensive score for each piece of intent information based on the index score corresponding to each evaluation index, the corresponding fuzzy correction factor, and the conflict compensation factor includes: determining the comprehensive score Conf for each piece of intent information based on the following formula:

[0039] Where, k∈[1, h], S k The score for the k-th evaluation indicator. Here, h is the fuzzy correction factor corresponding to the score of the k-th evaluation indicator, and h is the number of evaluation indicators. The conflict compensation factor is... The conflict sensitivity coefficient, This is the minimum offset. This is the dimensional balance coefficient.

[0040] It should be noted that the minimum offset The value is 10 -3 To avoid scoring S in a certain dimension k =0 results in the product term being 0, ensuring the robustness of the model.

[0041] The dimensionality balance coefficient λ is set to 2.3 (determined experimentally) to adjust the distribution range of the overall confidence level, making the output more concentrated in the effective range of [0,1].

[0042] In an exemplary embodiment, before determining the comprehensive score for each intent information in the intent set, the method further includes: performing dependency parsing matching on verb information in each intent information and verb information in the instruction information to generate an action matching score for each intent information and the instruction information; performing fuzzy named entity matching on noun information in each intent information and noun information in the instruction information to generate an entity matching score for each intent information and the instruction information; performing semantic slot alignment on parameter value information in each intent information and parameter value information in the instruction information to generate a parameter value matching score for each intent information and the instruction information; and determining the reasonableness score based on the action matching score, the entity matching score, and the parameter value matching score.

[0043] In this embodiment, dependency parsing is performed on the original instruction text to extract the predicate-object structure and modification relationships, forming an action semantic graph of the instruction. Simultaneously, the verbs contained in each intent in the intent set (such as "turn on", "turn down", "turn off") are matched. If the verb contained in the intent and the core verb in the instruction have a direct dominance or synonymous extension relationship in the dependency tree (e.g., the instruction is "help me dim the light a little", and the intent verb is "dim"), the action matching score is recorded as 0.9–1.0; if the verbs are synonymous but have different semantic strengths (e.g., the instruction is "turn off the light", and the intent is "turn off the light"), the score is 0.7–0.8; if the verbs deviate (e.g., the instruction is "it's too hot", and the intent verb is "open the window"), the score is 0.3–0.5, indicating a semantic shift between the intended action and the user's semantic intent.

[0044] Fuzzy named entity recognition matching is performed on nominal entities (devices, spaces, objects) in the instructions and intents. For example: if "living room light" and "master bedroom light" are predefined as different devices in the device ontology library, and "light" is a generalized category; if the instruction is "dim the living room light" and the intent is "living room light", then it matches and scores 1.0; if the intent is "living room lighting" or "main light", then semantic similarity is calculated through word vector similarity (such as Word2Vec). If the semantic similarity is >0.85, the score is 0.85; if the intent is "bedroom light", although it is a similar device, the space does not match, so the score is 0.4; if the intent is "refrigerator", then the semantics are completely unrelated, so the score is 0.1.

[0045] Semantic slot alignment is performed on the parameter values ​​in the command, mapping them to a standardized parameter value range. For example, the command "dim it a little" is aligned to "reduce brightness by 10%–20%"; the command "it's a little hot" is aligned to "lower the temperature by 2℃–4℃"; and the default parameter for the command "turn on the air conditioner" is "cooling mode, 26℃".

[0046] If the intent explicitly provides parameters (such as "adjust to 50% brightness"), determine whether the parameters are within the reasonable range of the instruction semantic slot: if the instruction is "darken a little" and the intent is "adjust to 60%", then because "60%" is relatively bright, it is the opposite of the intent, and the score is 0.2; if the intent is "adjust to 30%", it is outside of "10%–20%" but is within the reasonable brightness reduction range, and the score is 0.7; if the intent is "adjust to 15%", it is within the slot, and the score is 0.95.

[0047] In an exemplary embodiment, before determining the comprehensive score for each intent information in the intent set, the method further includes: querying a device state map corresponding to the target object, wherein the device state map is used to indicate the device state of devices in the device set corresponding to the target object; determining the device state of the target device corresponding to each intent information based on the device state map, and determining device collaboration information of the target device, wherein the device collaboration information is used to indicate whether the target device needs the collaboration of other devices when executing the control command corresponding to each intent information; and determining a device reachability score corresponding to each intent information based on the device state, the device collaboration information, and a first conditional response function.

[0048] In this embodiment, an updatable device state graph is maintained on a local or edge server. The device state graph is a directed weighted graph structure. The nodes in the device state graph are smart devices registered in the home (such as air conditioners, lights, curtains, fresh air systems, humidifiers, etc.), and the edges represent spatial relationships, energy consumption linkages, logical dependencies, or security constraints between devices. Each node contains the following attribute information:

[0049] Device status: on / off status, operating mode, current parameters (such as temperature, brightness, wind speed), online status, network latency, access control (such as child lock, elderly mode).

[0050] Device capabilities: Supported action set, parameter range, response latency.

[0051] Collaborative information: Pre- and post-device dependencies required to execute a certain intention. For example, "turning on the air conditioner" requires "closing the windows" to save energy; "dimming the lights" requires "closing the curtains" to ensure environmental comfort.

[0052] For each intent in the intent set, based on the target device specified in the intent (e.g., "living room air conditioner"), the current state information of the target device is determined in the device state graph. Based on the edge relationships in the device state graph, it is determined whether the execution of the intent requires the cooperation of other devices. For example:

[0053] If the intention is to "start the air conditioner for cooling", and there is an edge in the graph that reads "Air conditioner ← needs to be closed → window", then it means that the window must be closed first.

[0054] If the intention is to "dim the bedroom light", and the diagram contains the edge "Bedroom light ← suggested to synchronize → curtains", it means that it is recommended to close the curtains to enhance the atmosphere.

[0055] Finally, based on the first conditional response function, and considering both the device status and collaborative constraints, the final device reachability score is output.

[0056] Optionally, determining the device reachability score corresponding to each intent information based on the device status, the device collaboration information, and the first conditional response function includes: when the device status of the target device corresponding to each intent information is offline, determining a first sub-function matching the device status in the first conditional response function, and determining the device reachability score based on the first sub-function; when the device status of the target device corresponding to each intent information is online, and the device collaboration information instructs the target device to execute the control command corresponding to each intent information without requiring collaboration from other devices, determining a first sub-function matching the device status, the device collaboration information, and the first conditional response function in the first conditional response function. The system includes a second sub-function for matching device collaboration information, and determines the device reachability score based on the second sub-function. When the target device corresponding to each intent information is in an online state, and the device collaboration information instructs the target device to execute the control command corresponding to each intent information, and other device collaboration is required, the system determines the ratio of the total number of online collaborative devices to the total number of collaborative devices. It also determines a third sub-function in the first conditional response function that matches the device state and the device collaboration information, and determines the device reachability score based on the ratio and the third sub-function. The total number of collaborative devices is the sum of the total number of online collaborative devices and the total number of offline and online collaborative devices.

[0057] This application embodiment calls different sub-functions to perform scoring calculations based on the online status of the target device and the device collaboration requirements. Specifically, the first conditional response function contains three mutually exclusive sub-functions, each corresponding to one of three cases:

[0058] 1. When the target device corresponding to each intent information is in an offline state in the device state graph (i.e., device heartbeat timeout ≥ 3 times, network disconnection, power off, or firmware abnormality), it is determined that the device is currently uncontrollable, and the first sub-function is called in the first conditional response function:

[0059] A device reachability score of 0 means that commands to offline devices are prohibited to avoid invalid requests consuming communication resources or causing user misjudgment.

[0060] 2. When the target device is online and executing the current intent does not require the cooperation of other devices, the second sub-function is called in the first conditional response function:

[0061] A device reachability score of 1 indicates that the device is in good condition, has no external dependencies, and can execute commands. For example, the command "Turn on the desk lamp" calls the second sub-function when there is no window control linkage or interference from other devices.

[0062] 3. When the target device is online, but the execution of the current intention requires the cooperation of other devices (such as "turn on the air conditioner" requires "close the window"), call the third sub-function: Device reachability score = total number of online cooperating devices / total number of cooperating devices.

[0063] In an exemplary embodiment, before determining the comprehensive score of each intent information in the intent set, the method further includes: determining the state information of the target device after the target device executes the control instruction corresponding to each intent information, and determining the current state information of the target device, wherein the target device is the target device corresponding to each intent information; matching the state information with the current state information through a multi-dimensional rule engine; determining the conflict type between the state information and the current state information based on the matching result, and determining the state consistency score according to the conflict type.

[0064] Optionally, determining the state consistency score based on the conflict type includes: determining a second conditional response function corresponding to the state consistency score; if the conflict type is a first conflict type, determining a fourth sub-function corresponding to the conflict type in the second conditional response function, and determining the state consistency score based on the fourth sub-function; if the conflict type is a second conflict type, determining a fifth sub-function corresponding to the conflict type in the second conditional response function, and determining the state consistency score based on the fifth sub-function; if the conflict type is a third conflict type, determining a sixth sub-function corresponding to the conflict type in the second conditional response function, and determining the state consistency score based on the sixth sub-function, wherein the conflict severity of the first conflict type is higher than that of the second conflict type, and the conflict severity of the second conflict type is higher than that of the third conflict type.

[0065] This application embodiment includes a multi-dimensional rule engine, which can be built based on the experience of experts in the field of home appliance control and historical execution logs, and includes the following three conflict judgment dimensions:

[0066] The first type of conflict includes:

[0067] Mode conflict: This means that modes are mutually exclusive. For example, "cooling" and "dehumidification" cannot coexist (on the same device).

[0068] Redundancy in actions: This refers to actions that are repeated, such as: "Turn off the air conditioner" when the air conditioner is already off; "Turn down the brightness" when the brightness is already 0%;

[0069] Safety conflict: Safety constraints, such as "Turn off the kitchen ventilation fan" when "Turn on the gas stove".

[0070] The second type of conflict includes:

[0071] Parameter out of range: This means that the parameter is out of range, for example, the temperature setting value exceeds the range supported by the device (such as <16℃ or >30℃).

[0072] Energy conflict: Energy consumption contradictions, such as "turning on the air conditioner" and "opening all windows" at the same time;

[0073] The expected state and the current state are taken as input. Pattern matching is performed on each dimension, and a set of conflict types is output. If there is no conflict, it is marked as "no conflict"; if there is a conflict, its type is labeled (such as "pattern conflict", "energy consumption conflict", etc.).

[0074] The third type of conflict includes: no conflict.

[0075] In the absence of conflict, i.e., when the expected state is consistent with the current state, or when there are only minor adjustments to non-critical parameters, the sixth sub-function is called to determine the state consistency score as 0.95–1.0.

[0076] In cases where parameters exceed limits (i.e., the set value exceeds the equipment's capacity) or where there is an energy consumption conflict (i.e., it contradicts the energy-saving strategy), the state consistency score is determined to be 0.6–0.8.

[0077] In cases of redundant operations (i.e., the intended action is repeated with the current state), or in cases of mode conflict (i.e., the intended action is mutually exclusive with the current mode, such as "cooling" vs. "dehumidifying"), or in cases of safety conflict (i.e., involving personal / property safety, such as turning off the fan and turning on the gas stove), the state consistency score is determined to be 0-0.3.

[0078] The rating value is a range, and the specific value can be dynamically adjusted according to the type of equipment.

[0079] In an exemplary embodiment, after determining the target intent set based on the verification result, the inference process information, and the intent set, the method further includes: determining feature information for each target intent information in the target intent set, wherein the feature information includes: timing constraint information, energy consumption type, and device collaboration relationship; cyclically executing the determination step to obtain the dependency of each target intent information on other target intent information, wherein the determination step includes: determining the dependency of the i-th target intent information on the j-th target intent information based on the feature information of the i-th target intent information and the feature information of the j-th target intent information, where i and j are positive integers, and i is not equal to j; establishing a target directed acyclic graph based on the dependency of each target intent information on other target intent information, and generating control instructions based on the target directed acyclic graph, wherein the target directed acyclic graph is used to indicate the execution order of the target intent information.

[0080] In this embodiment, after determining the final target intent set based on the verification results, inference process information, and initial intent set, the execution order of the target intents is determined through feature analysis and dependency modeling. First, for each target intent in the target intent set, key feature information is extracted: timing constraint information, energy consumption type, and device collaboration relationship. The timing constraint information indicates whether the intent requires execution before or after other intents, such as a user explicitly stating "close the window before turning on the air conditioner" or implicit physical logic such as "the window must be closed before turning on the air conditioner." The energy consumption type refers to the energy consumption level of the device driven by the schematic diagram, such as high energy consumption (air conditioner, underfloor heating), medium energy consumption (fan, fresh air system), or low energy consumption (lighting, sensors). The device collaboration relationship indicates whether the execution of the schematic diagram depends on changes in the state of other devices, including mandatory collaboration (such as "the window must be closed before turning on the air conditioner") or recommended collaboration (such as "it is recommended to close the curtains when dimming the lights").

[0081] Subsequently, the dependency determination step is executed iteratively. For each distinct pair of intents i and j (i≠j) in the target intent set, the dependency of the i-th intent on the j-th intent is calculated based on the feature information of the i-th intent and the j-th intent. The dependency of the i-th intent on the j-th intent is used to indicate whether the execution of the i-th intent should take precedence over the j-th intent. After calculating the dependencies between all intent pairs, a weighted directed graph is constructed based on the directed edges. Topological sorting is performed on the directed acyclic graph to obtain the execution sequence. Furthermore, intents without dependencies are executed in parallel, and the output is a sequence of instructions divided by steps and executed concurrently.

[0082] In an exemplary embodiment, determining the dependency of the i-th target intent information on the j-th target intent information based on the feature information of the i-th target intent information and the feature information of the j-th target intent information includes: determining the time constraint value of the i-th target intent information on the j-th target intent information based on the time constraint information of the i-th target intent information and the time constraint information of the j-th target intent information; and determining the energy consumption value of the i-th target intent information based on the energy consumption type of the i-th target intent information, and determining the dependency of the i-th target intent information on the j-th target intent information based on the energy consumption type of the j-th target intent information. The energy consumption value of the j-th target intent information is used to determine the energy consumption correlation value between the i-th target intent information and the j-th target intent information based on the maximum value between the i-th target intent information and the j-th target intent information; the device collaboration value between the i-th target intent information and the j-th target intent information is determined based on the device collaboration relationship between the i-th target intent information and the j-th target intent information; and the dependency of the i-th target intent information on the j-th target intent information is determined based on the timing constraint value, the energy consumption correlation value, and the device collaboration value.

[0083] In this embodiment, the timing constraint information of the i-th intention and the j-th intention are analyzed to determine whether there is an explicit requirement for their sequential execution. If the timing constraint of the i-th intention includes "must be executed before the j-th intention", then the timing constraint value is 1; if the i-th intention requires "must be executed after the j-th intention", then the timing constraint value is zero; if there is no explicit or implicit timing relationship between the two, then the timing constraint value is 0.5.

[0084] Based on the energy consumption types of the i-th and j-th intentions, corresponding energy consumption values ​​are determined. Energy consumption types are divided into three levels: high, medium, and low, each with a corresponding numerical range. For example, high energy consumption is 0.8, medium energy consumption is 0.5, and low energy consumption is 0.2. Then, the maximum energy consumption value between the i-th and j-th intentions is taken as the energy consumption correlation value between them. For example, if the i-th intention is low energy consumption (0.2) and the j-th intention is high energy consumption (0.8), then the energy consumption correlation value is 0.8.

[0085] Based on the device collaboration relationship between the i-th intent and the j-th intent, determine whether there is a cross-intent device dependency relationship. If the device involved in the j-th intent is a prerequisite for the execution of the i-th intent, such as "closing the window" being a mandatory collaborative device for "turning on the air conditioner", then the device collaboration value is determined to be 1; if the two devices have no sharing, linkage, or logical dependency, then the device collaboration value is zero.

[0086] Finally, the timing constraint value, energy consumption correlation value, and device coordination value are comprehensively calculated through a preset weighted fusion function to obtain the overall dependency of the i-th intention on the j-th intention.

[0087] In an exemplary embodiment, establishing a target directed acyclic graph (DAG) based on the dependency of each target intent information on other target intent information includes: establishing nodes of the DAG based on each target intent information, wherein the nodes are used to indicate target intent information; cyclically executing the establishment step to establish directed edges in the DAG, wherein the establishment step includes: when the dependency of the i-th target intent information on the j-th target intent information is greater than a preset dependency, establishing a first directed edge in the DAG, wherein the source node of the first directed edge is the node corresponding to the i-th target intent information, and the target node of the first directed edge is the node corresponding to the j-th target intent information, i and j are positive integers, and i is not equal to j; when a cycle exists in the DAG, deleting the directed edge with the smallest weight in the cycle to obtain the target DAG, wherein the weight of the directed edge is used to indicate the dependency of the target intent information corresponding to the source node of the directed edge on the target intent information corresponding to the target node of the directed edge.

[0088] A directed acyclic graph of targets is constructed based on the dependence of each target intent information on other target intent information, including the following steps: First, a unique node is created for each target intent information in the target intent set, and each node represents an independent control intent.

[0089] The process of establishing directed edges is executed iteratively, traversing all target intent pairs (i, j), where i and j are positive integers and i ≠ j. For each intent pair, it is determined whether the dependency of the i-th intent on the j-th intent exceeds a preset dependency threshold. If the dependency exceeds the threshold, a directed edge is established in the graph from the node corresponding to the i-th intent to the node corresponding to the j-th intent. The source node of this directed edge is the i-th intent node, and the target node is the j-th intent node. Each directed edge has a weight value, which is the dependency value.

[0090] After constructing all directed edges, perform cycle detection on the directed graph. If a circular dependency is detected in the graph, such as intent A depending on intent B, intent B depending on intent C, and intent C depending on intent A, forming a closed cycle, then traverse all directed edges that constitute the cycle, determine the edge with the smallest weight, and delete the edge with the smallest weight from the graph.

[0091] The directed edge acyclic graph structure not only fully preserves the priority constraints between intents, but also identifies the dependency strength through weights, providing a precise structural basis for subsequent topology sorting and parallel scheduling. This ensures that the generation of control instruction sequences satisfies the semantic order of user intents while avoiding execution deadlock, thus achieving robustness of control logic in complex multi-intent scenarios.

[0092] In an exemplary embodiment, before obtaining the semantic elements corresponding to the instruction information of the target object, the method further includes: segmenting the audio signal of the target object into multiple consecutive frames; inputting each frame of audio data into a second model so that the second model outputs type information of each frame of audio data, wherein the second model is trained with audio sample data labeled with speech and non-speech, and the type information includes one of the following: speech information, non-speech information; and calculating the energy intensity of each frame of audio data, comparing the energy intensity with a preset energy intensity, and determining the energy feature of each frame of audio data through the comparison result; determining the start and end points of the target audio data in the audio data according to the type information and the energy feature; determining the target audio data in the audio data according to the start and end points of the target audio data, and determining the instruction information corresponding to the target audio data.

[0093] This application preprocesses the acquired raw audio signal to determine the effective time period of the user's voice command, thereby suppressing the interference of environmental noise and irrelevant speech and improving the accuracy of subsequent semantic parsing.

[0094] Specifically, the continuous audio signal of the target object is processed into frames with a fixed duration (e.g., 10 milliseconds) to form temporally continuous audio frames. Each frame of audio data is input into a second model, which is a speech activity detector built based on a deep neural network. The training samples of the second model are a large amount of audio data labeled "speech" and "non-speech," including various background noises in the home environment (such as television sound, air conditioner fan, human conversation, etc.) and real speech command samples. The second model outputs the type information of each frame of audio data, identifying each frame as "speech" or "non-speech."

[0095] The short-time energy intensity of each frame of audio data is calculated by calculating the root mean square value of the sampling points within the frame. The short-time energy intensity is compared with a preset energy threshold. If the energy exceeds the threshold, the audio data of that frame is determined to have acoustic activity.

[0096] The start and end frames of the target audio data are determined based on the "type information" output by the second model and the "energy characteristics" calculated by energy. When a frame satisfies the condition that "the model determines it to be speech" and "the energy intensity exceeds the threshold", it is identified as a candidate point for the start of speech. If several consecutive frames (such as 5 consecutive frames) are determined to be "non-speech" or "low energy", the first non-speech frame is identified as the end point of speech.

[0097] In an exemplary embodiment, after determining the target intent set based on the verification result, the reasoning process information, and the intent set, the method further includes: determining whether there is first intent information and second intent information in the target intent set that do not conform to the target rules; if there is first intent information and second intent information that do not conform to the target rules, determining the lower priority intent information among the first intent information and the second intent information according to the target rules; deleting the lower priority intent information from the target intent set to obtain an adjusted target intent set.

[0098] In this embodiment, each intent information in the target intent set is compared with a predefined "target rule". The target rule can be a security policy, energy efficiency policy, or user preference policy of a smart home system, such as: "When a gas leak detection is triggered, prohibit any device from starting", "Prohibit high-brightness lighting from being turned on in night mode", "Automatically turn off the air conditioner if the window is open while it is running", etc. The intent pairs in the target intent set are checked to determine whether there are one or more pairs of intent information (i.e., the first intent information and the second intent information) that simultaneously meet the triggering conditions, but whose execution behaviors conflict logically or in terms of rules.

[0099] For example, if the target intent set contains both "turn on the living room air conditioner" and "open all windows", and the target rule includes "do not open windows while the air conditioner is running to ensure energy efficiency", then "turn on the living room air conditioner" and "open all windows" are determined to be inconsistent with the target rule.

[0100] Upon detecting conflicting intent information, the conflicting intent information is adjusted according to the preset priority determination rules in the target rules. These priority determination rules include: safety rules take precedence over comfort rules, energy-saving rules take precedence over convenience rules, explicit user commands take precedence over system default policies, and emergency commands take precedence over routine controls. Based on these priority determination rules, the intent information with lower priority is determined between the first and second intent information. For example, in a conflict between "opening the window" and "turning on the air conditioner," if the rule explicitly states "energy saving takes precedence," then "opening the window" has lower priority. The lower-priority intent information is then deleted from the target intent set, completing a dynamic correction of the intent set. This ensures that the final output intent set conforms to both the user's intent and the preset rules.

[0101] In one exemplary embodiment, determining a target intent set based on the verification result, the inference process information, and the intent set includes: determining the intent set as the target intent set when the verification result indicates that each intent information passes the verification; and correcting any intent information based on the inference process information when the verification result indicates that any intent information fails the verification, thereby obtaining the target intent set.

[0102] Optionally, modifying any intention information based on the reasoning process information to obtain the target intention set includes: determining erroneous reasoning information corresponding to any intention information in the reasoning process information; determining correct intention information corresponding to the erroneous reasoning information; and replacing any intention information with the correct intention information to obtain the target intention set.

[0103] For each intent in the initial intent set, a consistency check is performed. If the check indicates that any intent has failed the check (e.g., an intent to "set the air conditioner temperature to 10℃"—lower than the device's minimum allowed setting), the "erroneous inference information" associated with the failed intent is identified in the inference process information. Erroneous inference information may include: over-interpreting the user's ambiguous expression (e.g., misinterpreting "it's a little hot" as "turn it to the lowest temperature"), misjudging the device's capabilities (e.g., mistaking "dimmerable light" for "only supports on / off"), and ignoring the context (e.g., generating the "close the window" intent even when the window is already closed).

[0104] After identifying erroneous inference information, the correct intent information is derived by combining contextual information in the inference chain, user history preferences, real-time device status, and a preset rule base. The original, unvalidated intent information is then replaced with the correctly inferred intent information. After replacement, the corrected intent is written into the intent set and re-validated to ensure that all constraints are met.

[0105] To better understand the process of the above-described intent recognition method based on large model thinking chain reasoning, the implementation flow of the above-described intent recognition method based on large model thinking chain reasoning will be described below in conjunction with optional embodiments, but this is not intended to limit the technical solution of the embodiments of this application.

[0106] This embodiment provides an intent recognition method based on large-scale model thought chain reasoning. Figure 3 This is a schematic diagram of an intent recognition method based on large-model thought chain reasoning according to an embodiment of this application, such as... Figure 3 As shown, the specific steps are as follows:

[0107] Step S301: Analyze the input audio and locate speech segments using a fusion strategy of deep neural network and energy detection technology.

[0108] Step S301 filters out ambient noise, ensuring that subsequent processing modules receive clear voice signals.

[0109] Step S302: The automatic speech recognition module receives the processed speech signal and uses an end-to-end model based on the Transformer architecture to directly map the audio features into a text sequence.

[0110] Step S302 simplifies the recognition process, improves robustness to natural language, and ensures that even colloquial instructions can be accurately converted.

[0111] Step S303: Send the text information to the multi-intent semantic parsing module, and identify key actions, device entities and attribute values ​​through semantic role labeling and slot filling technology.

[0112] Step S304: The finely tuned large language model uses Few-shot Chain-of-Thought prompting engineering to decompose complex instructions into a series of atomic intentions.

[0113] Each intention represents an independent control action.

[0114] Specifically, multiple sets of reasoning examples for home scenarios are pre-built. Each set of examples includes "user instructions, structured thought chain reasoning process information, atomic intent set, and verification result". The reasoning examples serve as reference templates for the large language model to perform reasoning. The number of examples is set to 3 to 5 sets. For example, for the compound instruction "I'm hot and the living room is too bright," the reasoning process information in the prompt template is as follows: First, semantic dependency analysis identifies "hot" and "the living room is too bright" as parallel structures, with no sequential dependency between them, corresponding to two independent needs: temperature control and lighting adjustment (decision basis: semantic function and semantic role labeling results of the conjunction "and"); Second, the device type corresponding to the temperature control need is matched, mapping the temperature control request "hot" to the air conditioner (default priority is higher than the fan); Third, the device corresponding to the lighting adjustment need is matched, mapping "too bright" to the living room lighting device; Fourth, combined with common sense action mapping relationships, "hot" corresponds to "lower the temperature," and "too bright" corresponds to "dimming the brightness"; Fifth, finally, two independent atomic intents are generated: "lower the living room air conditioner temperature" and "dimming the living room lights," ensuring that each intent contains only a single device and a single action, conforming to the minimum semantic unit specification. The verification result is labeled as "no missing intents, no logical conflicts, and reasonable device and action matching."

[0115] In the semantic dependency-guided reasoning and explicit reasoning recording stage, the structured information (including key action words, device entities, attribute values, and conjunctions or logical connectives in the sentence) output from the front-end speech recognition and preliminary semantic parsing is input into the fine-tuned large language model (i.e., the first model in the above embodiment). The model first parses the syntactic structure of the user's original command based on semantic dependency analysis technology, identifying the syntactic structure as coordinate, progressive, causal, or conditional, etc. For example, when processing the command "The bedroom is cold, and I also want to close the living room curtains," the model identifies the word "also" between "the bedroom is cold" and "close the living room curtains" as a coordinate conjunction. The model, referencing the reasoning process in the prompt template, gradually generates a clear and traceable thought chain: First, it confirms that the syntactic structure is a parallel relationship with no dependency (decision basis: conjunction "still" + dependency parsing results); second, it identifies "bedroom cold" as a temperature control need, combining the location word "bedroom" with the device-need mapping rule to determine the target device as the bedroom air conditioner, with the action direction being "adjust the temperature" (decision basis: semantic association between the temperature perception word "cold" and the temperature control action); third, it identifies "closing the living room curtains" as a curtain control need, based on the location word "living room" and the inherent matching rule between the action "close" and the curtain device, determining the target device as the living room curtains, with the action direction being "close"; fourth, it generates multiple intent information; fifth, it verifies that each intent information points to a different device, has no semantic overlap, and conforms to the independence principle of atomic intents; sixth, it generates a structured reasoning log, fully recording the reasoning path and the grammatical rules used in each of the above steps.

[0116] Step S305: Determine the comprehensive score for each intent message, and verify each intent message based on the comprehensive score.

[0117] Specifically, the formula for calculating the overall score is as follows:

[0118] Where, k∈[1, h], S k The score for the k-th evaluation indicator. Let h be the fuzzy correction factor corresponding to the score of the k-th evaluation indicator, h be the number of evaluation indicators, Δ be the conflict compensation factor, and γ be the conflict sensitivity coefficient. λ is the minimum offset, and λ is the dimensionality balance coefficient.

[0119] For example, when h=3, the evaluation indicators include: reasonableness, equipment accessibility, and state consistency.

[0120] S1: Semantic rationality score (range [0,1]): Calculated based on the pre-trained semantic consistency model (i.e., the second one in the above embodiment), with the input being "atomic intent set + original composite instruction", and the output being semantic matching degree (e.g., the matching degree between "lower the air conditioner temperature" and "I am hot" is 0.95, and the matching degree between "open the curtains" and "I am hot" is 0.1).

[0121] S2: Device Reachability Score (range [0,1]): Query the device status graph. If the target device is online and supports the current action, the score is 1; if the device is offline, the score is 0; if the device is online but needs to rely on other devices for collaboration (e.g., "starting linkage mode" requires both air conditioner and humidifier to be online), then the score = number of online collaborative devices / total number of collaborative devices required.

[0122] S3: State Consistency Score (range [0,1]): Calculates the degree of conflict between the device state after the execution of the atomic intent and the current state. No conflict is scored as 1 (e.g., the device is currently off, and the intent is "on"). Slight conflict is scored as 0.6-0.8 (e.g., the device is currently at 26℃, and the intent is "adjust to 25℃"). Severe conflict is scored as 0-0.3 (e.g., the device is currently on, and the intent is "on" with no parameter adjustment).

[0123] Fuzzy correction factor (k=1,2,3: Based on fuzzy logic, the scores of each dimension are adjusted for scene adaptation, with values ​​ranging from [0.9,1.1]. The calculation formula is:)

[0124] ,in For dimensional fuzziness coefficients (empirical values: , , ), used to amplify the correction magnitude of the score extreme value (such as S) k When =1, S k When =0, ).

[0125] Minimum offset The value is 10. -3 To avoid scoring S in a certain dimension k =0 results in the product term being 0, ensuring the robustness of the model.

[0126] Dimensional balance coefficient The value is 2.3 (determined experimentally), used to adjust the distribution range of the overall confidence level, so that the output is more concentrated in the effective range of [0,1].

[0127] Conflict compensation factor : Reflects the residual risk after conflict resolution, calculated using the following formula: Wherein, "conflict resolution confidence" is the decision reliability output by the conflict resolution module (range [0,1]), when there is no conflict. The more conflicts there are, the lower the reliability of their resolution. The larger the value, the lower the final confidence level.

[0128] Activation function: The Sigmoid function is used to map the comprehensive calculation results to the [0,1] interval to ensure that the confidence level is a standardized probability value (0 = completely unreliable, 1 = completely reliable).

[0129] Set confidence threshold (Calibrated using data from 100,000 home control commands, balancing accuracy and interaction frequency).

[0130] like Execute step S306, if Generate targeted clarification requests using the T5-based NLG module, example:

[0131] Low semantic rationality (S1<0.5): "Your control needs are not clearly defined. Do you need to adjust the air conditioning temperature or the brightness of the lights?"

[0132] Low device accessibility (S2<0.5): "The living room air conditioner is currently offline. Is it still necessary to perform other operations?"

[0133] Low consistency (S3<0.5): "The bedroom lights have been adjusted to 50% brightness. Is further adjustment necessary?"

[0134] Step S306: Through the conflict detection and resolution module, predefined priority rules are applied to identify and resolve conflicts between intentions, thereby avoiding the execution of chaotic or unreasonable instructions.

[0135] Define a series of conflict rules, such as "safety rules take precedence" (the "gas leak" command has the highest priority) and "energy-saving rules take precedence" ("close windows" takes precedence over "turn on air conditioning"). When an intent conflict is detected (such as "open all windows" versus "turn on air conditioning"), provide the optimal execution suggestion based on the rule base or issue a clarification to the user.

[0136] Step S307: Generate a directed acyclic graph based on intent dependency.

[0137] Specifically, the quantitative modeling and logical determination of dependencies between intents: First, define core parameters to provide a foundation for dependency analysis: Let the set of atomic intents after decomposing a compound instruction be I={I1,I2,...,In}, where each intent I... k Associate three features:

[0138] Timing identifier T kThe value can be {0, 1} (0 = no explicit timing requirements, 1 = includes explicit timing constraints such as "first / last").

[0139] Resource correlation degree R k Values ​​range from [0,1] (calculated based on the energy consumption type of the device; for high-energy-consuming devices such as air conditioners and water heaters, use 0.8-1.0; for low-energy-consuming devices such as lights and curtains, use 0.1-0.3).

[0140] Device Collaboration Tag C k The value is {0,1} (1 = there is a functional relationship with other devices, such as air conditioner-curtains, humidifier-air conditioner; 0 = no functional relationship).

[0141] Definition Intent I i to I j Dependence W i,j (That is, the degree to which Ij must be executed before Ii), the formula is as follows: W i,j =0.5×T i,j +0.3×R i,j +0.2×C i,j .

[0142] T i,j Let I be the time-series constraint factor. i Containing "after" or "afterwards" and other terms pointing to I j The temporal word, or I j Containing words like "first" and "before" that point to I i The time sequence word, then T i,j =1; otherwise T i,j =0.

[0143] R i,j As the resource correlation factor, take I. i with I j The maximum resource correlation degree, i.e., R i,j =max(R i ,R j (Resource association between high-energy-consuming devices has higher priority).

[0144] C i,j Let I be the equipment coordination factor. i with I j If the target device is a collaborative device (such as an air conditioner and curtains), then C i,j =1; otherwise C i,j =0.

[0145] When W i,j When ≥0.4, I is determined. i Dependence I j That is, there exists a directed edge I. j →I i When Wi,j When the value is less than 0.4, it is determined that there is no dependency.

[0146] The DAG generation process is as follows: Construct a node set V=I (where the atomic intent is a node), and an edge set E= ; Traverse all intention pairs (I i, I j (i=j), calculate W using the above formula i,j If W i,j If the value is ≥0.4, then add a directed edge (I). j ,I i A simplified topological sorting method is used; if a cycle is detected (e.g., I1→I2→I1), the weight W is retained. i,j Larger edges are removed, and smaller edges are deleted to ensure no loops. Nodes with no incoming edges (no prior dependencies) are assigned to the "initial layer" and executed first. Nodes that depend only on nodes in the initial layer are assigned to the "second layer" and executed after the initial layer has finished. Nodes with no dependencies are assigned to the same layer to support parallel execution.

[0147] For example, the set of atomic intents I = {I1 = Close the living room curtains, I2 = Turn on the living room air conditioner, I3 = Dim the living room lights} after decomposing a compound instruction has the following characteristics:

[0148] I1: T1=0, R1=0.2, C1=1 (in conjunction with air conditioning).

[0149] I2: T2=0, R2=0.9, C2=1 (in conjunction with the curtains);

[0150] I3: T3=0, R3=0.1, C3=0.

[0151] Calculate the dependency between each intent:

[0152] W 2,1 =0.5×0+0.3×max(0.9,0.2)+0.2×1=0+0.27+0.2=0.47≥0.4, add edge I1→I2;

[0153] W 2,3 =0.5×0+0.3×max(0.9,0.1)+0.2×0=0.27<0.4, no dependent edges;

[0154] W 3,1 =0.5×0+0.3×max(0.1,0.2)+0.2×0=0.06<0.4, no dependent edges.

[0155] The execution order from the initial layer (I1) to the second layer (I2,I3) is as follows: first, close the living room curtains, then start the living room air conditioner and dim the living room lights in parallel. This satisfies both the need for collaboration and energy saving, while also ensuring execution efficiency.

[0156] Step S308: Generate a JSON instruction sequence based on the directed acyclic graph.

[0157] Step S309: The executor receives the JSON instruction sequence and sends control commands to the specific device via lightweight communication protocols such as MQTT / CoAP. Simultaneously, the monitoring module continuously listens for device feedback and automatically initiates an exponential backoff and retry strategy when execution fails.

[0158] Step S310: After the instruction is executed, summarize the execution results and use the T5-based natural language generation module to convert the details of the control operation into natural language, such as "The living room air conditioner has been adjusted to a lower temperature for you, and the bedroom lights have been dimmed."

[0159] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0160] Figure 4 This is a structural block diagram of an intent recognition device based on large-model thought chain reasoning according to an embodiment of this application; as shown... Figure 4 As shown, it includes:

[0161] The first determining module 42 is used to obtain the semantic elements corresponding to the instruction information of the target object, and input the semantic elements and the prompt word information indicating the sample reasoning process information of the first model into the first model, so that the first model outputs the intent set corresponding to the instruction information and the reasoning process information of the first model;

[0162] The second determining module 44 is used to determine a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score.

[0163] Verification module 46 is used to verify each intent information based on the comprehensive score of each intent information to obtain a verification result;

[0164] The third determining module 48 is used to determine the target intent set based on the verification result, the reasoning process information, and the intent set.

[0165] Through the aforementioned device, composite voice commands are precisely decomposed into structured intent information using a target model and instruction prompts. A device state graph dynamically assesses the feasibility of each intent in the current environment, ensuring that the instructions match the actual device state and capabilities, thus preventing abnormal execution. After confirming no intent conflicts, the execution order is automatically adjusted according to preset rules, constructing a reasonable device control flow, such as "close the window" before "turn on the air conditioner." This achieves logical optimization and orderly control among multiple intents, significantly improving the execution efficiency and user experience of the smart home system, and solving the problem of low intent recognition accuracy in related technologies.

[0166] In an exemplary embodiment, the second determining module is configured to determine an indicator score for each intent information in the intent set based on evaluation indicators, wherein the indicator score includes at least one of the following: a reasonableness score, a device accessibility score, and a state consistency score; determine a fuzzy correction factor corresponding to each indicator score; determine a first number of conflicting intent information in the intent set and a second number of total intent information in the intent set, and determine a conflict compensation factor based on the ratio of the first number to the second number; and determine a comprehensive score for each intent information based on the indicator score corresponding to each evaluation indicator, the corresponding fuzzy correction factor, and the conflict compensation factor.

[0167] In an exemplary embodiment, the second determining module is configured to determine the fuzzy correction factor corresponding to each index score based on the following formula: Where k∈[1,h], and h is the number of evaluation indicators. S is the dimensional fuzziness coefficient. k The score for the k-th evaluation indicator. This is the fuzzy correction factor corresponding to the index score of the kth evaluation index.

[0168] In one exemplary embodiment, the second determining module is configured to determine a comprehensive score Conf for each intent piece of information based on the following formula:

[0169] Where, k∈[1, h], S k The score for the k-th evaluation indicator. Here, h is the fuzzy correction factor corresponding to the score of the k-th evaluation indicator, and h is the number of evaluation indicators. The conflict compensation factor is... The conflict sensitivity coefficient, This is the minimum offset. This is the dimensional balance coefficient.

[0170] In one exemplary embodiment, the second determining module is configured to perform dependency parsing matching on verb information in each intent information and verb information in the instruction information to generate an action matching score for each intent information and the instruction information; perform fuzzy named entity matching on noun information in each intent information and noun information in the instruction information to generate an entity matching score for each intent information and the instruction information; perform semantic slot alignment on parameter value information in each intent information and parameter value information in the instruction information to generate a parameter value matching score for each intent information and the instruction information; and determine the reasonableness score based on the action matching score, the entity matching score, and the parameter value matching score.

[0171] In an exemplary embodiment, the second determining module is configured to query a device state map corresponding to the target object, wherein the device state map is used to indicate the device state of devices in the device set corresponding to the target object; determine the device state of the target device corresponding to each intent information based on the device state map, and determine the device collaboration information of the target device, wherein the device collaboration information is used to indicate whether the target device needs the collaboration of other devices when executing the control command corresponding to each intent information; and determine the device reachability score corresponding to each intent information based on the device state, the device collaboration information, and a first conditional response function.

[0172] In an exemplary embodiment, the second determining module is configured to: when the device state of the target device corresponding to each intent information is offline, determine a first sub-function matching the device state in the first conditional response function, and determine the device reachability score based on the first sub-function; when the device state of the target device corresponding to each intent information is online, and the device collaboration information instructs the target device to execute the control command corresponding to each intent information without the need for other device collaboration, determine a second sub-function matching the device state and the device collaboration information in the first conditional response function, and determine the device reachability score based on the second sub-function; when the device state of the target device corresponding to each intent information is online, and the device collaboration information instructs the target device to execute the control command corresponding to each intent information without the need for other device collaboration, determine a ratio of the total number of online collaborative devices to the total number of collaborative devices, and determine a third sub-function matching the device state and the device collaboration information in the first conditional response function, and determine the device reachability score based on the ratio and the third sub-function, wherein the total number of collaborative devices is the sum of the total number of online collaborative devices and the total number of offline and online collaborative devices.

[0173] In an exemplary embodiment, the second determining module is configured to determine the state information of the target device after the target device executes the control instruction corresponding to each intent information, and to determine the current state information of the target device, wherein the target device is the target device corresponding to each intent information; match the state information with the current state information through a multi-dimensional rule engine; determine the conflict type between the state information and the current state information based on the matching result, and determine the state consistency score according to the conflict type.

[0174] In one exemplary embodiment, the second determining module is configured to: determine a second conditional response function corresponding to the state consistency score; if the conflict type is a first conflict type, determine a fourth sub-function corresponding to the conflict type in the second conditional response function, and determine the state consistency score based on the fourth sub-function; if the conflict type is a second conflict type, determine a fifth sub-function corresponding to the conflict type in the second conditional response function, and determine the state consistency score based on the fifth sub-function; if the conflict type is a third conflict type, determine a sixth sub-function corresponding to the conflict type in the second conditional response function, and determine the state consistency score based on the sixth sub-function, wherein the conflict severity of the first conflict type is higher than that of the second conflict type, and the conflict severity of the second conflict type is higher than that of the third conflict type.

[0175] In an exemplary embodiment, the above-described apparatus further includes: a generation module, configured to determine feature information of each target intent information in the target intent set, wherein the feature information includes: timing constraint information, energy consumption type, and device collaboration relationship; cyclically executing a determination step to obtain the dependency of each target intent information on other target intent information, wherein the determination step includes: determining the dependency of the i-th target intent information on the j-th target intent information based on the feature information of the i-th target intent information and the feature information of the j-th target intent information, where i and j are positive integers and i is not equal to j; establishing a target directed acyclic graph based on the dependency of each target intent information on other target intent information, and generating control instructions based on the target directed acyclic graph, wherein the target directed acyclic graph is used to indicate the execution order of the target intent information.

[0176] In one exemplary embodiment, the generation module is configured to: determine a timing constraint value of the i-th target intent information on the j-th target intent information based on the timing constraint information of the i-th target intent information and the timing constraint information of the j-th target intent information; determine an energy consumption value of the i-th target intent information based on its energy consumption type, determine an energy consumption value of the j-th target intent information based on its energy consumption type, determine an energy consumption association value of the i-th target intent information on the j-th target intent information based on the maximum value of the energy consumption values ​​of the i-th and j-th target intent information; determine a device collaboration value of the i-th target intent information on the j-th target intent information based on the device collaboration relationship of the i-th target intent information and the device collaboration relationship of the j-th target intent information; and determine the dependency of the i-th target intent information on the j-th target intent information based on the timing constraint value, the energy consumption association value, and the device collaboration value.

[0177] In an exemplary embodiment, a generation module is configured to establish nodes of a directed acyclic graph based on each target intent information, wherein the nodes are used to indicate target intent information; and to cyclically execute a building step to establish directed edges in the directed acyclic graph, wherein the building step includes: if the dependency of the i-th target intent information on the j-th target intent information is greater than a preset dependency, establishing a first directed edge in the directed acyclic graph, wherein the source node of the first directed edge is the node corresponding to the i-th target intent information, and the target node of the first directed edge is the node corresponding to the j-th target intent information, i and j are positive integers, and i is not equal to j; if a cycle exists in the directed acyclic graph, deleting the directed edge with the smallest weight in the cycle to obtain the target directed acyclic graph, wherein the weight of the directed edge is used to indicate the dependency of the target intent information corresponding to the source node of the directed edge on the target intent information corresponding to the target node of the directed edge.

[0178] In an exemplary embodiment, a first determining module is configured to: segment the audio signal of the target object into multiple consecutive frames; input each frame of audio data into a second model, so that the second model outputs type information of each frame of audio data, wherein the second model is trained on audio sample data labeled with speech and non-speech, and the type information includes one of the following: speech information, non-speech information; calculate the energy intensity of each frame of audio data, compare the energy intensity with a preset energy intensity, and determine the energy characteristics of each frame of audio data based on the comparison result; determine the start and end points of the target audio data in the audio data according to the type information and the energy characteristics; determine the target audio data in the audio data according to the start and end points of the target audio data, and determine the instruction information corresponding to the target audio data.

[0179] In an exemplary embodiment, the third determining module is configured to determine whether there is first intent information and second intent information in the target intent set that do not conform to the target rules; if there is first intent information and second intent information that do not conform to the target rules, determine the intent information with lower priority among the first intent information and the second intent information according to the target rules; delete the intent information with lower priority from the target intent set to obtain an adjusted target intent set.

[0180] In one exemplary embodiment, the third determining module is configured to determine the intent set as the target intent set when the verification result indicates that each intent information has passed the verification; and to modify any intent information according to the reasoning process information when the verification result indicates that any intent information has failed the verification, so as to obtain the target intent set.

[0181] In an exemplary embodiment, the third determining module is configured to determine erroneous reasoning information corresponding to any intention information in the reasoning process information; determine correct intention information corresponding to the erroneous reasoning information; and replace any intention information with the correct intention information to obtain the target intention set.

[0182] Embodiments of this application also provide a storage medium including a stored program, wherein the program executes any of the methods described above when it is run.

[0183] Optionally, in this embodiment, the storage medium may be configured to store program code for performing the following steps:

[0184] S1, obtain the semantic elements corresponding to the instruction information of the target object, and input the semantic elements and the prompt words indicating the sample reasoning process information of the first model into the first model, so that the first model outputs the intent set corresponding to the instruction information and the reasoning process information of the first model;

[0185] S2, determine a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score;

[0186] S3, Verify each intent information based on the comprehensive score of each intent information to obtain a verification result;

[0187] S4. Determine the target intent set based on the verification result, the reasoning process information, and the intent set.

[0188] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0189] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0190] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0191] S1, obtain the semantic elements corresponding to the instruction information of the target object, and input the semantic elements and the prompt words indicating the sample reasoning process information of the first model into the first model, so that the first model outputs the intent set corresponding to the instruction information and the reasoning process information of the first model;

[0192] S2, determine a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score;

[0193] S3, Verify each intent information based on the comprehensive score of each intent information to obtain a verification result;

[0194] S4. Determine the target intent set based on the verification result, the reasoning process information, and the intent set.

[0195] Optionally, in this embodiment, the storage medium may include, but is not limited to, 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.

[0196] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0197] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0198] The embodiments described herein also provide a computer program that includes computer instructions stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in any of the above method embodiments.

[0199] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0200] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0201] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An intent recognition method based on large-scale model thinking chain reasoning, characterized in that, include: Obtain the semantic elements corresponding to the instruction information of the target object, and input the semantic elements and the prompt words that indicate the sample reasoning process information of the first model into the first model, so that the first model outputs the intent set corresponding to the instruction information and the reasoning process information of the first model; Determine a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score; Each intent information is verified based on a comprehensive score for each intent information to obtain a verification result. The target intent set is determined based on the verification result, the reasoning process information, and the intent set.

2. The intent recognition method based on large-scale model thinking chain reasoning according to claim 1, characterized in that, Determining a comprehensive score for each intent in the intent set includes: The evaluation criteria determine the score for each intent information in the intent set, wherein the score includes at least one of the following: reasonableness score, device accessibility score, and state consistency score. Determine the fuzzy correction factor corresponding to the score of each indicator; Determine a first number of conflicting intent information in the intent set and a second number of total intent information in the intent set, and determine a conflict compensation factor based on the ratio of the first number to the second number; The comprehensive score for each intent information is determined based on the index score corresponding to each evaluation indicator, the corresponding fuzzy correction factor, and the conflict compensation factor.

3. The intent recognition method based on large-scale model thinking chain reasoning according to claim 2, characterized in that, Determine the fuzzy correction factor corresponding to the score of each indicator, including: The fuzzy correction factor corresponding to the score of each indicator is determined based on the following formula: Where k∈[1,h], and h is the number of evaluation indicators. S is the dimensional fuzziness coefficient. k The score for the k-th evaluation indicator. This is the fuzzy correction factor corresponding to the index score of the kth evaluation index.

4. The intent recognition method based on large-scale model thinking chain reasoning according to claim 2, characterized in that, Based on the score of each evaluation indicator, the corresponding fuzzy correction factor, and the conflict compensation factor, a comprehensive score for each intent information is determined, including: The overall score for each intent piece of information is determined based on the following formula: Where, k∈[1, h], S k The score for the k-th evaluation indicator. Here, h is the fuzzy correction factor corresponding to the score of the k-th evaluation indicator, and h is the number of evaluation indicators. The conflict compensation factor is... The conflict sensitivity coefficient, This is the minimum offset. This is the dimensional balance coefficient.

5. The intent recognition method based on large-scale model thinking chain reasoning according to claim 1, characterized in that, Before determining the overall score for each intent in the intent set, the method further includes: Dependency syntax matching is performed on the verb information in each intent information and the verb information in the instruction information to generate an action matching score for each intent information and the instruction information. Fuzzy named entity matching is performed on the noun information in each intent information and the noun information in the instruction information to generate an entity matching score for each intent information and the instruction information; Semantic slot alignment is performed between the parameter value information in each intent information and the parameter value information in the instruction information to generate a parameter value matching score between each intent information and the instruction information. The reasonableness score is determined based on the action matching score, the entity matching score, and the parameter value matching score.

6. The intent recognition method based on large-scale model thinking chain reasoning according to claim 1, characterized in that, Before determining the overall score for each intent in the intent set, the method further includes: Query the device status map corresponding to the target object, wherein the device status map is used to indicate the device status of the devices in the device set corresponding to the target object; Based on the device status map, the device status of the target device corresponding to each intent information is determined, and the device collaboration information of the target device is determined, wherein the device collaboration information is used to indicate whether the target device needs the collaboration of other devices when executing the control command corresponding to each intent information; The device reachability score corresponding to each intent information is determined based on the device status, the device collaboration information, and the first conditional response function.

7. The intent recognition method based on large-scale model thought chain reasoning according to claim 6, characterized in that, Determining the device reachability score corresponding to each intent based on the device status, the device collaboration information, and the first conditional response function includes: If the device status of the target device corresponding to each intent information is offline, a first sub-function matching the device status is determined in the first conditional response function, and the device reachability score is determined based on the first sub-function. When the target device corresponding to each intent information is in an online state, and the device collaboration information instructs the target device to execute the control command corresponding to each intent information, and no other device collaboration is required, a second sub-function matching the device state and the device collaboration information is determined in the first conditional response function, and the device reachability score is determined based on the second sub-function. When the target device corresponding to each intent information is in an online state, and the device collaboration information instructs the target device to execute the control command corresponding to each intent information, and other device collaboration is required, the ratio of the total number of online collaborative devices to the total number of collaborative devices is determined. A third sub-function matching the device state and the device collaboration information is determined in the first conditional response function. The device reachability score is then determined based on the ratio and the third sub-function. The total number of collaborative devices is the sum of the total number of online collaborative devices and the total number of offline and online collaborative devices.

8. The intent recognition method based on large-scale model thinking chain reasoning according to claim 1, characterized in that, Before determining the overall score for each intent in the intent set, the method further includes: After determining that the target device executes the control command corresponding to each intent information, the status information of the target device and the current status information of the target device are determined, wherein the target device is the target device corresponding to each intent information; The state information is matched with the current state information using a multi-dimensional rule engine; Based on the matching results, the conflict type between the state information and the current state information is determined, and the state consistency score is determined according to the conflict type.

9. The intent recognition method based on large-scale model thinking chain reasoning according to claim 8, characterized in that, Determining the state consistency score based on the conflict type includes: Determine the second conditional response function corresponding to the state consistency score; When the conflict type is the first conflict type, a fourth sub-function corresponding to the conflict type is determined in the second conditional response function, and the state consistency score is determined based on the fourth sub-function; In the case that the conflict type is the second conflict type, a fifth sub-function corresponding to the conflict type is determined in the second conditional response function, and the state consistency score is determined based on the fifth sub-function; In the case that the conflict type is the third conflict type, a sixth sub-function corresponding to the conflict type is determined in the second conditional response function, and the state consistency score is determined based on the sixth sub-function, wherein the conflict severity of the first conflict type is higher than that of the second conflict type, and the conflict severity of the second conflict type is higher than that of the third conflict type.

10. The intent recognition method based on large-scale model thought chain reasoning according to claim 1, characterized in that, After determining the target intent set based on the verification result, the reasoning process information, and the intent set, the method further includes: Determine the feature information of each target intent information in the target intent set, wherein the feature information includes: time-series constraint information, energy consumption type, and device collaboration relationship; The determination step is executed iteratively to obtain the dependency of each target intent information on other target intent information. The determination step includes: determining the dependency of the i-th target intent information on the j-th target intent information based on the feature information of the i-th target intent information and the feature information of the j-th target intent information, where i and j are positive integers and i is not equal to j. A target directed acyclic graph is established based on the dependency of each target intent information on other target intent information, and control instructions are generated based on the target directed acyclic graph, wherein the target directed acyclic graph is used to indicate the execution order of the target intent information.

11. The intent recognition method based on large-scale model thought chain reasoning according to claim 10, characterized in that, Determining the dependency of the i-th target intent information on the j-th target intent information based on the feature information of the i-th target intent information and the feature information of the j-th target intent information includes: Based on the temporal constraint information of the i-th target intent information and the temporal constraint information of the j-th target intent information, determine the temporal constraint value of the i-th target intent information on the j-th target intent information; and... The energy consumption value of the i-th target intent information is determined based on the energy consumption type of the i-th target intent information; the energy consumption value of the j-th target intent information is determined based on the energy consumption type of the j-th target intent information; and the energy consumption correlation value between the i-th target intent information and the j-th target intent information is determined based on the maximum value between the energy consumption values ​​of the i-th and j-th target intent information; and, Based on the device collaboration relationship of the i-th target intent information and the device collaboration relationship of the j-th target intent information, determine the device collaboration value of the i-th target intent information to the j-th target intent information; The dependency of the i-th target intent information on the j-th target intent information is determined based on the time-series constraint value, the energy consumption correlation value, and the device coordination value.

12. The intent recognition method based on large-scale model thinking chain reasoning according to claim 10, characterized in that, A directed acyclic graph of targets is constructed based on the dependency of each target intent information on other target intent information, including: Based on each target intent information, a directed acyclic graph is constructed with nodes, wherein the nodes are used to indicate the target intent information; The establishment steps are executed repeatedly to establish directed edges in the directed acyclic graph. The establishment steps include: when the dependency of the i-th target intent information on the j-th target intent information is greater than a preset dependency, establishing a first directed edge in the directed acyclic graph, where the source node of the first directed edge is the node corresponding to the i-th target intent information, and the target node of the first directed edge is the node corresponding to the j-th target intent information, where i and j are positive integers, and i is not equal to j. If a cycle exists in the directed acyclic graph, the directed edge with the smallest weight in the cycle is deleted to obtain the target directed acyclic graph. The weight of the directed edge is used to indicate the dependence of the target intent information corresponding to the source node of the directed edge on the target intent information corresponding to the target node of the directed edge.

13. The intent recognition method based on large-scale model thinking chain reasoning according to claim 1, characterized in that, Before obtaining the semantic elements corresponding to the instruction information of the target object, the method further includes: The audio signal of the target object is divided into multiple consecutive frames; Each frame of audio data is input into a second model, which outputs type information for each frame of audio data. The second model is trained using audio sample data labeled with both speech and non-speech information. The type information includes one of the following: speech information, non-speech information; and... The energy intensity of each frame of audio data is calculated, and the energy intensity is compared with a preset energy intensity. The energy characteristics of each frame of audio data are determined by the comparison result. Based on the type information and the energy characteristics, the start and end points of the target audio data are determined in the audio data; The target audio data is determined in the audio data based on the start and end points of the target audio data, and the instruction information corresponding to the target audio data is determined.

14. The intent recognition method based on large-scale model thinking chain reasoning according to claim 1, characterized in that, After determining the target intent set based on the verification result, the reasoning process information, and the intent set, the method further includes: Determine whether there are first intent information and second intent information in the target intent set that do not conform to the target rules; In the case where there are first intent information and second intent information that do not conform to the target rule, the intent information with lower priority is determined from the first intent information and the second intent information according to the target rule; The low-priority intent information is deleted from the target intent set to obtain the adjusted target intent set.

15. The intent recognition method based on large-scale model thought chain reasoning according to claim 1, characterized in that, Determining the target intent set based on the verification result, the reasoning process information, and the intent set includes: If the verification result indicates that each intent information passes the verification, the intent set is determined as the target intent set; If the verification result indicates that any intent information fails the verification, the intent information is corrected according to the reasoning process information to obtain the target intent set.

16. The intent recognition method based on large-scale model thought chain reasoning according to claim 15, characterized in that, Based on the reasoning process information, any of the intent information is modified to obtain the target intent set, including: Determine the erroneous reasoning information corresponding to any intention information from the reasoning process information; Determine the correct intent information corresponding to the erroneous inference information, and replace any of the intent information with the correct intent information to obtain the target intent set.

17. An intent recognition device based on large-scale model thought chain reasoning, characterized in that, include: The first determining module is used to obtain the semantic elements corresponding to the instruction information of the target object, and input the semantic elements and the prompt words that indicate the sample reasoning process information of the first model into the first model, so that the first model outputs the intent set corresponding to the instruction information and the reasoning process information of the first model. The second determining module is used to determine a comprehensive score for each intent information in the intent set, wherein the comprehensive score is determined based on at least one of the following: reasonableness score, device accessibility score, and state consistency score; The verification module is used to verify each intent information based on the comprehensive score of each intent information in order to obtain the verification result. The third determining module is used to determine the target intent set based on the verification result, the reasoning process information, and the intent set.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed, performs the method described in any one of claims 1 to 16.

19. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 16 through the computer program.

20. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 16.

Images