A method and system for vehicle voice intent prediction based on dialogue history

Abstract

The application discloses a kind of vehicle-mounted voice intention prediction method and system based on dialogue history, it is related to vehicle-mounted voice interaction technical field, the method includes dialogue history storage, slot extraction and association, LSTM intention prediction, confidence determination and question interaction steps: through local cache storage multi-round dialogue slot and intention information, realize the intelligent association of history and current slot;Optimized LSTM model is constructed to extract timing characteristics and output intention probability distribution;Determine the optimal confidence threshold in combination with ROC curve, dynamically decide to directly output intention or trigger targeted follow-up questions;While providing two kinds of implementation variants of light weight and enhanced type, adapt to different computing power vehicle-mounted terminal.The application realizes the accuracy rate of multi-round dialogue intention prediction to improve and delay to reduce, weak network / offline scene is all available, significantly improve the accuracy, real-time performance and user experience of vehicle-mounted voice interaction, suitable for voice control system of various intelligent cars.

Description

Technical Field

[0001] This invention relates to the field of in-vehicle voice interaction technology, and in particular to an in-vehicle voice intent prediction method and system based on dialogue history. Background Technology

[0002] With the rapid development of intelligent vehicles, in-vehicle voice interaction has become the core way for users to control vehicle functions (such as air conditioning, navigation, and multimedia) and obtain information services (such as weather and traffic conditions). Among them, multi-turn dialogue is a typical interaction mode in in-vehicle scenarios. Users often complete complex needs through continuous queries with contextual relationships, such as "turn on the air conditioning → adjust to 25 degrees → foot blowing mode" or "navigate to a certain place → avoid traffic jams → play music along the way".

[0003] However, existing in-vehicle voice intent recognition technology has the following key problems in multi-turn dialogue scenarios:

[0004] 1. Ambiguous user intent and low recognition accuracy in multi-turn dialogues: Existing technologies mostly rely on semantic parsing of single-turn dialogues (such as directly parsing the current query text through NLU), ignoring the intent-related information hidden in historical dialogues. When user queries are ambiguous or lack information (such as "adjust it again" or "open it"), the system cannot infer the user's true intent from historical dialogues, resulting in an intent recognition accuracy of less than 65%. Users need to repeatedly clarify, leading to a poor user experience.

[0005] 2. Reliance on real-time semantic parsing, resulting in low response efficiency: Existing methods require a complete "speech-to-text → NLU semantic parsing → intent matching" process in each round of dialogue. The NLU semantic parsing stage, in particular, requires access to cloud services or complex local models, leading to response latency typically exceeding 500ms. In multi-round dialogue scenarios, the cumulative latency further increases; when a user initiates more than three consecutive rounds of dialogue, the response latency can reach over 800ms, failing to meet the real-time requirements of in-vehicle scenarios (the ideal response latency for in-vehicle voice interaction should be ≤300ms).

[0006] 3. Lack of confidence assessment mechanism, resulting in both false positives and false negatives: Existing technologies do not quantitatively assess the reliability of intent prediction results. They either directly output the predicted intent (leading to a high false positive rate) or trigger meaningless follow-up questions due to excessive caution. Statistical data shows that the false positive rate of existing systems is as high as 22%, and the invalid follow-up question rate exceeds 40%, seriously affecting user interaction efficiency.

[0007] In summary, existing in-vehicle voice intent recognition technologies cannot effectively solve the problems of intent ambiguity and low response efficiency in multi-turn dialogues. There is an urgent need for an in-vehicle voice intent prediction method and system based on dialogue history to solve the existing technical problems. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems existing in the prior art, and proposes a method and system for predicting in-vehicle voice intent based on dialogue history.

[0009] In a first aspect, embodiments of the present invention provide a method for predicting in-vehicle voice intent based on dialogue history, including:

[0010] The dialogue history is stored using a local cache to store the dialogue data from the most recent preset rounds. Each round of dialogue data includes the user's query text, the initial intent of the dialogue, historical slot information, and interaction timestamps, with a set storage validity period.

[0011] Extract and associate historical slot information to obtain the current query text and initial slot information. Combine historical slot information in the dialogue history and generate historical and current associated slot features through preset association rules.

[0012] The intent is predicted by using an LSTM model, taking associated slot features and historical dialogue sequences as input, and outputting the probability distribution of the user's current intent through a pre-trained LSTM model.

[0013] The confidence level is judged by calculating the maximum confidence level in the probability distribution and comparing the maximum confidence level with a preset threshold. If the maximum confidence level is greater than or equal to the preset threshold, the prediction intention is directly output; if the maximum confidence level is less than the preset threshold, the follow-up question interaction step is triggered.

[0014] Furthermore, in the step of extracting and associating slot information, the slot type includes adjustment object, parameter type, parameter value, domain, and time / location, and the slot extraction rules include:

[0015] For the adjustment object, match the in-vehicle function keywords, which include air conditioning, volume, navigation, and windows;

[0016] For parameter types, the matching function corresponds to parameter keywords, which include temperature, wind speed, volume, and destination;

[0017] For parameter values, numerical parameter values ​​are extracted by both the number and the unit, and parameter values ​​of different types are matched with preset options.

[0018] Regarding the domain, based on the domain to which the regulated object belongs, the domain includes vehicle control, navigation, multimedia, and information query;

[0019] For time / location, extract time words and location words.

[0020] Furthermore, in the step of extracting and associating slot information, historical and current associated slot features are generated through preset association rules. The specific method includes:

[0021] The domain is determined as follows: if the domain field of the currently queried slot is clear, then the slots with the same domain in the history are filtered; if the domain field of the currently queried slot is ambiguous, then the weight of each historical domain is calculated, and the weight = the number of occurrences / the sum of the round differences, and the domain with the highest weight is selected as the current domain.

[0022] Define time decay, define "round difference" as "current dialogue round - historical dialogue round", and define historical slot weight = 1 / (round difference + 1);

[0023] For slot conflict judgment and supplementation, if the current slot and the historical slot have the same "parameter type" but different "parameter value", it is judged as a conflict, and the current slot shall prevail; if the current slot is missing key information, the nearest non-conflicting slot is extracted from the historical slots in the same field and supplemented to the current slot.

[0024] Furthermore, in the LSTM intent prediction step, the LSTM model adopts a five-layer structure consisting of an input layer, an embedding layer, an LSTM layer, a fully connected layer, and an output layer, wherein:

[0025] The input data of the input layer includes historical dialogue sequences and associated slot features. The historical dialogue sequence is the user query text of the most recent preset round of dialogue, arranged in chronological order. The total length of the historical dialogue text is less than a preset number of characters, and the slot features are less than a preset number of key-value pairs.

[0026] The text embedding in the embedding layer uses a pre-trained Word2Vec model to convert historical dialogue text into a multi-dimensional vector representation; the slot embedding uses One-Hot encoding for the slot "key" and numerical mapping for the slot "value" to finally generate a multi-dimensional slot vector; the text vector and the slot vector are concatenated to form a multi-dimensional fusion feature vector.

[0027] The LSTM layer consists of two hidden layers. The output of the first layer is used as the input of the second layer to set the initial bias of the forget gate. A dropout layer is added between the two LSTM layers to set the dropout probability.

[0028] The fully connected layer consists of 64 neurons, and the activation function is ReLU to normalize the input features.

[0029] The output dimension of the output layer is equal to the number of intent categories supported by the in-vehicle voice system. The activation function is softmax, which converts the output into a probability distribution.

[0030] Furthermore, in the LSTM intent prediction step, the model training is based on an in-vehicle multi-turn dialogue dataset, and the specific training process includes:

[0031] The data was preprocessed by collecting multi-turn dialogues from the vehicle's voice logs, filtering out invalid dialogues, labeling slots for each turn of dialogue, and dividing the dataset into 80% training set, 10% validation set, and 10% test set.

[0032] The model is trained using the Adam optimizer, with settings for the learning rate and number of training epochs, and the batch size adjusted according to the performance of the training device.

[0033] Model optimization: Hyperparameters are optimized through grid search, an early stopping mechanism is adopted, training is stopped if the validation set loss does not decrease for several consecutive rounds, and model compression is performed using quantization and pruning;

[0034] The model is deployed locally as an edge model, using the TensorFlow framework, and its parameters are updated monthly via OTA.

[0035] Furthermore, the preset threshold is determined based on ROC curve analysis of the validation set, and the specific determination method includes:

[0036] Input a validation set, which contains a preset number of rounds of dialogue with labeled intent;

[0037] Iterate through the candidate thresholds and set the range and step size for each threshold.

[0038] For each threshold, calculate the true positive rate (TPR) and the false positive rate (FPR). TPR = number of correctly predicted intents / number of actual intents, and FPR = number of incorrectly predicted intents / number of non-intents.

[0039] Plot the ROC curve with FPR on the horizontal axis and TPR on the vertical axis; calculate the area under the curve and select the threshold that maximizes the area under the curve as the optimal threshold.

[0040] Furthermore, if the maximum confidence level is less than a preset threshold, a follow-up question interaction step is triggered. The follow-up question interaction step specifically includes: generating targeted slot follow-up questions for scenarios with insufficient confidence, obtaining the slot information supplemented by the user, and then returning to the slot extraction and association step to re-perform the association analysis.

[0041] 8. The intent prediction method according to claim 1, characterized in that it further includes an attention mechanism fused to the LSTM model, the specific fusion method including:

[0042] To enhance the model structure, an attention layer is added between the LSTM layer and the fully connected layer. The attention layer adopts Multi-HeadAttention, and the input of the attention layer is the temporal features output by the LSTM layer and the weight vector of the historical slots.

[0043] The slot association is optimized by introducing "domain weight decay". The weight of historical slots across domains is 1 / (round difference + 2). Multi-slot association is supported. When the current query involves multiple slots, historical slots from multiple domains can be associated at the same time.

[0044] The confidence level assessment is optimized by adjusting the threshold based on the dialogue rounds, and supporting the simultaneous output of two high-confidence intents to trigger parallel execution.

[0045] Secondly, this invention also discloses an in-vehicle voice intent prediction system based on dialogue history, employing the aforementioned intent prediction method, comprising: a dialogue history storage module, a historical slot information extraction and association module, an LSTM model intent prediction module, and a confidence level judgment module; wherein:

[0046] The dialogue history storage module is used to store the dialogue history. It uses local caching to store the dialogue data of the most recent preset rounds. Each round of dialogue data includes the user query text, the initial dialogue intent, historical slot information and interaction timestamps, and sets the storage validity period.

[0047] The historical slot information extraction and association module is used to extract and associate historical slot information, obtain the current query text and initial slot information, combine the historical slot information in the dialogue history, and generate historical and current associated slot features through preset association rules;

[0048] The LSTM model intent prediction module is used to predict intent using an LSTM model. It takes associated slot features and historical dialogue sequences as input and outputs the probability distribution of the user's current intent through a pre-trained LSTM model.

[0049] The confidence level judgment module is used to judge the confidence level, calculate the maximum confidence level in the probability distribution, compare the maximum confidence level with a preset threshold, and directly output the prediction intention if the maximum confidence level is greater than or equal to the preset threshold; if the maximum confidence level is less than the preset threshold, the follow-up question interaction step is triggered.

[0050] Thirdly, the present invention also discloses an electronic device, comprising:

[0051] One or more processors;

[0052] Memory, used to store one or more programs;

[0053] When the one or more programs are executed by the one or more processors, the one or more processors implement the intent prediction method.

[0054] This invention discloses a method and system for predicting in-vehicle voice intent based on dialogue history. The method includes dialogue history storage, slot extraction and association, LSTM intent prediction, confidence assessment, and follow-up questioning steps: It stores slot and intent information from multiple rounds of dialogue locally, and intelligently associates historical data with current slots based on the principles of "domain priority, time decay, and conflict coverage"; it constructs an optimized LSTM model to extract temporal features and outputs the intent probability distribution; it determines the optimal confidence threshold using ROC curves, and dynamically decides whether to directly output the intent or trigger targeted follow-up questions; it also provides two implementation variants, a lightweight version and an enhanced version, to adapt to in-vehicle terminals with different computing power. This invention improves the accuracy and reduces latency of multi-turn dialogue intent prediction, is fully usable in weak network / offline scenarios, significantly improves the accuracy, real-time performance, and user experience of in-vehicle voice interaction, and is applicable to voice control systems of various intelligent vehicles. Attached Figure Description

[0055] Figure 1 A flowchart illustrating a method for predicting in-vehicle voice intent based on dialogue history, provided in an embodiment of the present invention;

[0056] Figure 2 A structural block diagram of an in-vehicle voice intent prediction system based on dialogue history is provided in an embodiment of the present invention;

[0057] Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0058] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0059] Where there is no conflict, the various embodiments of the present invention and the features thereof may be combined with each other.

[0060] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0061] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used herein, the singular forms “a” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded. Terms such as “connected” or “linked” are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0062] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having the meaning consistent with their meaning in the context of the relevant art and the invention, and will not be interpreted as having an idealized or overly formal meaning unless expressly so defined herein.

[0063] In the technical solution of this invention, the collection, storage, use, processing, transmission, provision, and disclosure of user personal information all comply with relevant laws and regulations and do not violate public order and good morals. The use of user data in this technical solution follows relevant national laws and regulations (e.g., the "Information Security Technology - Personal Information Security Specification"). For example: appropriate measures are taken for personal information access control; restrictions are imposed on the display of personal information; the purpose of using personal information does not exceed the scope of direct or reasonable association; and explicit identity targeting is eliminated when using personal information to avoid precisely locating a specific individual.

[0064] To address at least one of the technical problems existing in the aforementioned related technologies, the present invention provides a method and system for predicting in-vehicle voice intent based on dialogue history;

[0065] This embodiment discloses a method for predicting in-vehicle voice intent based on dialogue history, such as... Figure 1 ,include:

[0066] S100. Store the dialogue history, using local caching to store the dialogue data of the most recent preset rounds. Each round of dialogue data includes the user query text, the initial dialogue intent, historical slot information, and interaction timestamps, and sets the storage validity period.

[0067] In this embodiment, a local Redis cache module is used to store the most recent 5 rounds of user dialogue data. Each round of data includes the user query text, NLU initial intent, slot information (adjustment object, parameter type, parameter value, domain, time / location), and interaction timestamp. The storage validity period is set to 30 minutes, which perfectly covers the multi-round dialogue requirements during a single driving session, while avoiding redundant data occupying storage resources. The local storage design ensures data accessibility in weak network / offline scenarios without relying on network transmission.

[0068] S200. Extract and associate historical slot information, obtain the current query text and initial slot information, combine historical slot information in the dialogue history, and generate historical and current associated slot features through preset association rules;

[0069] In the slot information extraction and association steps of this implementation, as shown in Table 1, the slot types include adjustment object, parameter type, parameter value, domain, and time / location. The slot extraction rules include:

[0070] For the adjustment object, match the in-vehicle function keywords, which include air conditioning, volume, navigation, and windows;

[0071] For parameter types, the matching function corresponds to parameter keywords, which include temperature, wind speed, volume, and destination;

[0072] For parameter values, numerical parameter values ​​are extracted by both the number and the unit, and parameter values ​​of different types are matched with preset options.

[0073] Regarding the domain, based on the domain to which the regulated object belongs, the domain includes vehicle control, navigation, multimedia, and information query;

[0074] For time / location, extract time words and location words.

[0075] Table 1 Slot Extraction Rules

[0076] In this embodiment, the process of extracting and associating slot information involves generating historical and current associated slot features using preset association rules. The specific method includes:

[0077] The domain is determined. If the domain field of the currently queried slot is clear (e.g., "adjustment object = air conditioner" → domain = vehicle control), then the slots with the same domain in the history are filtered. If the domain field of the currently queried slot is ambiguous (e.g., "adjust again" does not include the adjustment object), then the weight of each historical domain is calculated. The weight = the number of occurrences / the sum of the round differences, and the domain with the highest weight (e.g., vehicle control) is selected as the current domain.

[0078] The time decay is defined as follows: the "round difference" is defined as "current dialogue round - historical dialogue round" (e.g., the current round is the 5th round, and the round difference of the historical 3rd round is 2). The weight of the historical slot is defined as 1 / (round difference + 1). For example, the weight of the 4th round (round difference = 1) is 0.5, and the weight of the 3rd round (round difference = 2) is 0.33, to ensure that the most recent historical slot has a greater impact on the current association.

[0079] For slot conflict judgment and supplementation, if the current slot and the historical slot have the same "parameter type" but different "parameter value" (e.g., current "temperature 28℃" vs. historical "temperature 25℃"), it is judged as a conflict, and the current slot shall prevail; if the current slot is missing key information (e.g., "a little higher" is missing "adjustment object"), the nearest non-conflicting slot in the same domain (e.g., historical "adjustment object = air conditioner") shall be extracted and supplemented to the current slot.

[0080] S300. Intent is predicted using an LSTM model, taking associated slot features and historical dialogue sequences as input, and outputting the probability distribution of the user's current intent through a pre-trained LSTM model.

[0081] In this embodiment, the LSTM intent prediction step uses a five-layer structure: an input layer, an embedding layer, an LSTM layer, a fully connected layer, and an output layer.

[0082] The input data of the input layer includes historical dialogue sequences and associated slot features. The historical dialogue sequence is the user query text of the most recent preset round of dialogue, arranged in chronological order. The total length of the historical dialogue text is less than a preset number of characters, and the slot features are less than a preset number of key-value pairs.

[0083] The text embedding in the embedding layer uses a pre-trained Word2Vec model to convert historical dialogue text into a multi-dimensional vector representation; the slot embedding uses One-Hot encoding for the slot "key" and numerical mapping for the slot "value" to finally generate a multi-dimensional slot vector; the text vector and the slot vector are concatenated to form a multi-dimensional fusion feature vector; preferably, the embedding layer converts the text and slot into 128-dimensional and 64-dimensional vectors respectively and concatenates them into a 192-dimensional fusion feature.

[0084] The LSTM layer consists of two hidden layers. The output of the first layer is used as the input of the second layer. The initial bias of the forget gate is set (the initial bias is set to 0.8). A dropout layer is added between the two LSTM layers, and the dropout probability is set (probability 0.2).

[0085] The fully connected layer consists of 64 neurons, and the activation function is ReLU to normalize the input features.

[0086] The output dimension of the output layer is equal to the number of intent categories supported by the in-vehicle voice system. The activation function is softmax, which converts the output into a probability distribution.

[0087] In this embodiment, the LSTM intent prediction step involves model training based on an in-vehicle multi-turn dialogue dataset, and the specific training process includes:

[0088] The data was preprocessed by collecting multi-turn dialogues from the vehicle's voice logs, filtering out invalid dialogues, labeling slots for each turn of dialogue, and dividing the dataset into 80% training set, 10% validation set, and 10% test set.

[0089] The model is trained using the Adam optimizer, with settings for the learning rate and number of training epochs, and the batch size adjusted according to the performance of the training device.

[0090] The model is optimized by using grid search to optimize hyperparameters, employing an early stopping mechanism (if the validation set loss does not decrease for several consecutive rounds), and using quantization and pruning to compress the model.

[0091] The model is deployed locally as an edge model, using the TensorFlow framework, and its parameters are updated monthly via OTA.

[0092] S400. Determine the confidence level, calculate the maximum confidence level in the probability distribution, compare the maximum confidence level with a preset threshold, and if the maximum confidence level is greater than or equal to the preset threshold, directly output the prediction intention; if the maximum confidence level is less than the preset threshold, trigger the follow-up question interaction step.

[0093] In this embodiment, the preset threshold is determined based on ROC curve analysis of the validation set, and the specific determination method includes:

[0094] Input a validation set, which contains a preset number of rounds of dialogue with labeled intent;

[0095] Iterate through the candidate thresholds and set the range and step size for each threshold.

[0096] For each threshold, calculate the true positive rate (TPR) and the false positive rate (FPR). TPR = number of correctly predicted intents / number of actual intents, and FPR = number of incorrectly predicted intents / number of non-intents.

[0097] Plot the ROC curve with FPR on the horizontal axis and TPR on the vertical axis; calculate the area under the curve and select the threshold that maximizes the area under the curve as the optimal threshold.

[0098] In this embodiment, if the maximum confidence level is less than a preset threshold, a follow-up question interaction step is triggered. The follow-up question interaction step specifically includes: generating targeted slot follow-up questions for scenarios with insufficient confidence, obtaining the slot information supplemented by the user, and then returning to the slot extraction and association step to re-perform the association analysis.

[0099] In some embodiments, an in-vehicle voice intent prediction method based on dialogue history further includes fusing an attention mechanism to an LSTM model, specifically including:

[0100] To enhance the model structure, an attention layer is added between the LSTM layer and the fully connected layer. The attention layer adopts Multi-HeadAttention, and the input of the attention layer is the temporal features output by the LSTM layer and the weight vector of the historical slots.

[0101] The slot association is optimized by introducing "domain weight decay". The weight of historical slots across domains is 1 / (round difference + 2). Multi-slot association is supported. When the current query involves multiple slots, historical slots from multiple domains can be associated at the same time.

[0102] The confidence level assessment is optimized by adjusting the threshold based on the dialogue rounds, and supporting the simultaneous output of two high-confidence intents to trigger parallel execution.

[0103] To better understand the method disclosed in this embodiment, the following example of a multi-turn dialogue between a user and an air conditioner demonstrates the collaborative working principle of the method:

[0104] Round 1: User initiates query "Turn on the air conditioner"

[0105] Output the text "Turn on the air conditioner", triggering the GetHistorySlots interface (text length = 4 ≥ 2);

[0106] The slot extraction module extracts the initial slot (adjustment object = air conditioner, domain = vehicle control), the historical slots are empty, the slot integrity after association is 100%, and the InputFeature interface is triggered.

[0107] LSTM model input features, output probability distribution (air conditioner on: 0.95, others: 0.05), confidence level = 0.95 ≥ 0.7;

[0108] The confidence module triggers the ExecuteIntent interface, and the function execution module turns on the driver's side air conditioning and sends a message "Driver's side air conditioning has been turned on".

[0109] Round 2: User initiates query "adjust to 25 degrees"

[0110] The output text "Adjust to 25 degrees" triggers the GetHistorySlots interface to obtain the slot for cycle 1 (adjustment object = air conditioner, domain = vehicle control).

[0111] The slot extraction module extracts the initial slot (parameter type = temperature, parameter value = 25℃), associates it with the historical slot (supplementary adjustment object = air conditioner), sets the integrity to 100%, and triggers the InputFeature interface;

[0112] The LSTM model outputs the probability distribution (adjusting air conditioning temperature: 0.92, others: 0.08), with a confidence level of 0.92 ≥ 0.7.

[0113] The function execution module adjusts the temperature to 25℃ and provides feedback: "The driver's air conditioning temperature has been adjusted to 25℃".

[0114] Round 3: User initiates a query for "a little higher".

[0115] Output the text "A little higher", trigger the GetHistorySlots interface, and get the slots for rounds 1-2 (adjustment object = air conditioner, parameter type = temperature, parameter value = 25℃).

[0116] The slot extraction module extracts the initial slot (parameter value = higher, adjustment object = none), associates it with the historical slot (supplement the adjustment object = air conditioner, parameter type = temperature), completes the integrity = 100%, and triggers the InputFeature interface;

[0117] The LSTM model outputs the probability distribution (adjusting air conditioning temperature: 0.85, adjusting fan speed: 0.12, others: 0.03), with a confidence level of 0.85 ≥ 0.7.

[0118] The function execution module raises the temperature by 1°C (to 26°C) and reports "The driver's air conditioning temperature has been adjusted to 26°C".

[0119] Round 4: User initiates query for "foot blowjob".

[0120] Output the text "Blow feet", trigger the GetHistorySlots interface, and get the slots for rounds 1-3 (adjustment object = air conditioner, domain = vehicle control).

[0121] The slot extraction module extracts the initial slot (parameter type = blowing mode, parameter value = foot blowing, adjustment object = none), associates it with the historical slot (supplement adjustment object = air conditioner), completes the integrity = 100%, and triggers the InputFeature interface.

[0122] The LSTM model outputs the probability distribution (adjusting the air conditioner's fan speed: 0.90, others: 0.10), with a confidence level of 0.90 ≥ 0.7.

[0123] The function execution module switched the air blowing mode to foot blowing and provided the feedback "The driver's air conditioning blowing mode has been switched to foot blowing".

[0124] This embodiment discloses a method and system for predicting in-vehicle voice intent based on dialogue history. The method includes dialogue history storage, slot extraction and association, LSTM intent prediction, confidence judgment, and follow-up questioning steps: It stores slot and intent information from multiple rounds of dialogue locally, and intelligently associates historical and current slots based on the principles of "domain priority, time decay, and conflict coverage"; it constructs an optimized LSTM model to extract temporal features and outputs the intent probability distribution; it determines the optimal confidence threshold using ROC curves, and dynamically decides whether to directly output the intent or trigger targeted follow-up questions; it also provides two implementation variants, a lightweight version and an enhanced version, to adapt to in-vehicle terminals with different computing power. This invention improves the accuracy and reduces latency of multi-round dialogue intent prediction, is fully usable in weak network / offline scenarios, significantly improves the accuracy, real-time performance, and user experience of in-vehicle voice interaction, and is applicable to voice control systems of various intelligent vehicles.

[0125] Based on the same inventive concept, embodiments of the present invention also provide an in-vehicle voice intent prediction system based on dialogue history, employing the aforementioned intent prediction method, such as... Figure 2 It includes: a dialogue history storage module, a historical slot information extraction and association module, an LSTM model intent prediction module, and a confidence judgment module; among which:

[0126] The dialogue history storage module is used to store the dialogue history. It uses local caching to store the dialogue data of the most recent preset rounds. Each round of dialogue data includes the user query text, the initial dialogue intent, historical slot information and interaction timestamps, and sets the storage validity period.

[0127] The historical slot information extraction and association module is used to extract and associate historical slot information, obtain the current query text and initial slot information, combine the historical slot information in the dialogue history, and generate historical and current associated slot features through preset association rules;

[0128] The LSTM model intent prediction module is used to predict intent using an LSTM model. It takes associated slot features and historical dialogue sequences as input and outputs the probability distribution of the user's current intent through a pre-trained LSTM model.

[0129] The confidence level judgment module is used to judge the confidence level, calculate the maximum confidence level in the probability distribution, compare the maximum confidence level with a preset threshold, and directly output the prediction intention if the maximum confidence level is greater than or equal to the preset threshold; if the maximum confidence level is less than the preset threshold, the follow-up question interaction step is triggered.

[0130] The specific working methods of the dialogue history storage module, the historical slot information extraction and association module, the LSTM model intent prediction module, and the confidence judgment module have been described in detail in the above intent prediction method, and will not be repeated here.

[0131] Based on the same inventive concept, embodiments of the present invention also provide an electronic device. Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Figure 3 As shown, an embodiment of the present invention provides an electronic device including: one or more processors 101, a memory 102, and one or more I / O interfaces 103. The memory 102 stores one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement any of the intent prediction methods described in the above embodiments; the one or more I / O interfaces 103 are connected between the processors and the memory, configured to enable information interaction between the processors and the memory.

[0132] The processor 101 is a device with data processing capabilities, including but not limited to a central processing unit (CPU); the memory 102 is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read / write interface) 103 is connected between the processor 101 and the memory 102, and can realize information interaction between the processor 101 and the memory 102, including but not limited to a data bus (Bus).

[0133] In some embodiments, the processor 101, memory 102, and I / O interface 103 are interconnected via bus 104, and thus connected to other components of the computing device.

[0134] In some embodiments, the one or more processors 101 include a field-programmable gate array.

[0135] This invention also provides a computer-readable medium. The computer-readable medium stores a computer program, which, when executed by a processor, implements the steps of any of the intent prediction methods described in the above embodiments. The computer-readable storage medium may be volatile or non-volatile.

[0136] This invention also provides a computer program product, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in the processor of an electronic device, the processor in the electronic device executes the above-described intent prediction method.

[0137] Those skilled in the art will understand that all or some of the steps, systems, and apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software can be distributed on a computer-readable storage medium, which may include computer storage media (or non-transitory media) and communication media (or transient media).

[0138] As is known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable program instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), flash memory or other memory technologies, portable compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is known to those skilled in the art that communication media typically contain computer-readable program instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0139] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0140] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0141] The computer program product described herein can be implemented specifically through hardware, software, or a combination thereof. In one alternative embodiment, the computer program product is specifically embodied in a computer storage medium; in another alternative embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0142] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0143] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0144] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0145] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0146] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A method for predicting in-vehicle voice intent based on dialogue history, characterized in that, include: The dialogue history is stored using a local cache to store the dialogue data from the most recent preset rounds. Each round of dialogue data includes the user's query text, the initial intent of the dialogue, historical slot information, and interaction timestamps, with a set storage validity period. Extract and associate historical slot information to obtain the current query text and initial slot information. Combine historical slot information in the dialogue history and generate historical and current associated slot features through preset association rules. The intent is predicted by using an LSTM model, taking associated slot features and historical dialogue sequences as input, and outputting the probability distribution of the user's current intent through a pre-trained LSTM model. The confidence level is judged, the maximum confidence level in the probability distribution is calculated, and the maximum confidence level is compared with a preset threshold. If the maximum confidence level is greater than or equal to the preset threshold, the prediction intention is directly output. If the maximum confidence level is less than the preset threshold, a follow-up questioning interaction step is triggered.

2. The intention prediction method according to claim 1, characterized in that, In the process of extracting and associating slot information, the slot type includes adjustment object, parameter type, parameter value, domain, and time / location, and the slot extraction rules include: For the adjustment object, match the in-vehicle function keywords, which include air conditioning, volume, navigation, and windows; For parameter types, the matching function corresponds to parameter keywords, which include temperature, wind speed, volume, and destination; For parameter values, numerical parameter values ​​are extracted by both the number and the unit, and parameter values ​​of different types are matched with preset options. Regarding the domain, based on the domain to which the regulated object belongs, the domain includes vehicle control, navigation, multimedia, and information query; For time / location, extract time words and location words.

3. The intention prediction method according to claim 1, characterized in that, In the step of extracting and associating slot information, historical and currently associated slot features are generated through preset association rules. The specific method includes: The domain is determined as follows: if the domain field of the currently queried slot is clear, then the slots with the same domain in the history are filtered; if the domain field of the currently queried slot is ambiguous, then the weight of each historical domain is calculated, and the weight = the number of occurrences / the sum of the round differences, and the domain with the highest weight is selected as the current domain. Time decay is defined as follows: "round difference" is defined as "current dialogue round - historical dialogue round", and historical slot weight is defined as 1 / (round difference + 1). For slot conflict judgment and supplementation, if the current slot and the historical slot have the same "parameter type" but different "parameter value", it is judged as a conflict, and the current slot shall prevail; if the current slot is missing key information, the nearest non-conflicting slot is extracted from the historical slots in the same domain and supplemented to the current slot.

4. The intention prediction method according to claim 1, characterized in that, In the LSTM intent prediction step, the LSTM model adopts a five-layer structure: input layer, embedding layer, LSTM layer, fully connected layer, and output layer. The input data of the input layer includes historical dialogue sequences and associated slot features. The historical dialogue sequence is the user query text of the most recent preset round of dialogue, arranged in chronological order. The total length of the historical dialogue text is less than a preset number of characters, and the slot features are less than a preset number of key-value pairs. The text embedding in the embedding layer uses a pre-trained Word2Vec model to convert historical dialogue text into a multi-dimensional vector representation; the slot embedding uses One-Hot encoding for the slot "key" and numerical mapping for the slot "value" to finally generate a multi-dimensional slot vector; the text vector and the slot vector are concatenated to form a multi-dimensional fusion feature vector. The LSTM layer consists of two hidden layers. The output of the first layer is used as the input of the second layer to set the initial bias of the forget gate. A dropout layer is added between the two LSTM layers to set the dropout probability. The fully connected layer consists of 64 neurons, and the activation function is ReLU to normalize the input features. The output dimension of the output layer is equal to the number of intent categories supported by the in-vehicle voice system. The activation function is softmax, which converts the output into a probability distribution.

5. The intention prediction method according to claim 1, characterized in that, In the LSTM intent prediction step, the model training is based on an in-vehicle multi-turn dialogue dataset, and the specific training process includes: The data was preprocessed by collecting multi-turn dialogues from the vehicle's voice logs, filtering out invalid dialogues, labeling slots for each turn of dialogue, and dividing the dataset into 80% training set, 10% validation set, and 10% test set. The model is trained using the Adam optimizer, with settings for the learning rate and number of training epochs, and the batch size adjusted according to the performance of the training device. The model is optimized by using grid search to optimize hyperparameters, employing an early stopping mechanism (if the validation set loss does not decrease for several consecutive rounds), and using quantization and pruning to compress the model. The model is deployed locally as an edge model, using the TensorFlow framework, and its parameters are updated monthly via OTA.

6. The intention prediction method according to claim 1, characterized in that, The preset threshold is determined based on ROC curve analysis of the validation set, and the specific determination method includes: Input a validation set, which contains a preset number of rounds of dialogue with labeled intent; Iterate through the candidate thresholds and set the range and step size for each threshold. For each threshold, calculate the true positive rate (TPR) and the false positive rate (FPR). TPR = number of correctly predicted intents / number of actual intents, and FPR = number of incorrectly predicted intents / number of non-intents. Plot the ROC curve with FPR on the horizontal axis and TPR on the vertical axis; calculate the area under the curve and select the threshold that maximizes the area under the curve as the optimal threshold.

7. The intention prediction method according to claim 1, characterized in that, If the maximum confidence level is less than the preset threshold, a follow-up question interaction step is triggered. The follow-up question interaction step specifically includes: generating targeted slot follow-up questions for scenarios with insufficient confidence, obtaining the slot information supplemented by the user, and then returning to the slot extraction and association step to re-perform the association analysis.

8. The intention prediction method according to claim 1, characterized in that, It also includes an attention fusion mechanism for LSTM models, with specific fusion methods including: To enhance the model structure, an attention layer is added between the LSTM layer and the fully connected layer. The attention layer adopts Multi-HeadAttention, and the input of the attention layer is the temporal features output by the LSTM layer and the weight vector of the historical slots. The slot association is optimized by introducing "domain weight decay". The weight of historical slots across domains is 1 / (round difference + 2). Multi-slot association is supported. When the current query involves multiple slots, historical slots from multiple domains can be associated at the same time. The confidence level assessment is optimized by adjusting the threshold based on the dialogue rounds, and supporting the simultaneous output of two high-confidence intents to trigger parallel execution.

9. A vehicle-mounted voice intent prediction system based on dialogue history, employing the intent prediction method described in any one of claims 1-8, characterized in that, include: The module includes a dialogue history storage module, a historical slot information extraction and association module, an LSTM model intent prediction module, and a confidence level judgment module; among which: The dialogue history storage module is used to store the dialogue history. It uses local caching to store the dialogue data of the most recent preset rounds. Each round of dialogue data includes the user query text, the initial dialogue intent, historical slot information and interaction timestamps, and sets the storage validity period. The historical slot information extraction and association module is used to extract and associate historical slot information, obtain the current query text and initial slot information, combine the historical slot information in the dialogue history, and generate historical and current associated slot features through preset association rules; The LSTM model intent prediction module is used to predict intent using an LSTM model. It takes associated slot features and historical dialogue sequences as input and outputs the probability distribution of the user's current intent through a pre-trained LSTM model. The confidence level judgment module is used to judge the confidence level, calculate the maximum confidence level in the probability distribution, compare the maximum confidence level with a preset threshold, and directly output the prediction intention if the maximum confidence level is greater than or equal to the preset threshold; if the maximum confidence level is less than the preset threshold, the follow-up question interaction step is triggered.

10. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the intent prediction method as described in any one of claims 1 to 8.

Images