A voice interactive recommendation method, system, storage medium and vehicle-mounted device

Abstract

This specification discloses a voice interactive recommendation method, system, storage medium and vehicle-mounted device, the method comprising: collecting voice control commands issued by a user, and identifying a first behavioral intention that the user expects the current vehicle to perform based on the voice control command; inputting a behavior sequence prediction model based on the first behavioral intention for processing to obtain a number of recommendation intentions and their model prediction probabilities; calculating the recommendation probability of each of the several recommendation intentions based on the model prediction probabilities of each of the several recommendation intentions and the transition probabilities of each of the several recommendation intentions; the transition probabilities of each of the several recommendation intentions are global transition probabilities or individual transition probabilities; and determining a second behavioral intention for recommendation based on the recommendation probabilities of each of the several recommendation intentions.

Description

Technical Field

[0001] This specification relates to voice control technology, and in particular to a voice interactive recommendation method, system, storage medium, and vehicle-mounted device. Background Art

[0002] With the continuous development of vehicle intelligence, in-vehicle voice interaction systems have emerged to assist in vehicle control.

[0003] The in-vehicle voice interaction system primarily uses voice interaction with the driver to understand the driver's intentions, thereby providing a variety of services to the driver. For example, if the driver says "close the windows," the in-vehicle voice interaction system can recognize the driver's intention as "close the windows" and control the windows to close.

[0004] However, current in-vehicle voice interaction systems generally use algorithms to complete single-round voice conversation recognition, and do not actively predict the owner's subsequent needs. For example, after controlling the windows to close, it does not predict whether the user still needs to turn on the air conditioner or play music, resulting in the current in-vehicle voice interaction system being not very intelligent. Summary of the Invention

[0005] This specification provides a voice interactive recommendation method, system, storage medium and in-vehicle device. By combining model prediction probability and transition probability, it accurately explores the user's subsequent intentions after a single round of voice dialogue, realizes in-depth interaction with the user over multiple rounds of dialogue, and thus makes the in-vehicle device more intelligent, thereby solving or partially solving the current technical problem of low intelligence caused by only being able to achieve a single round of voice dialogue.

[0006] To solve the above technical problems, the first aspect of this specification discloses a voice interactive recommendation method, which includes:

[0007] collecting a voice control command issued by a user, and identifying a first behavioral intention that the user desires the current vehicle to perform based on the voice control command;

[0008] Inputting the behavior sequence prediction model into the first behavior intention for processing, and obtaining a number of recommended intentions and their model prediction probabilities;

[0009] Calculate the recommendation probabilities of the multiple recommendation intentions based on the model prediction probabilities of the multiple recommendation intentions and the transition probabilities of the multiple recommendation intentions; the transition probabilities of the multiple recommendation intentions are global transition probabilities or individual transition probabilities;

[0010] According to the recommendation probabilities of the plurality of recommendation intentions, a second behavior intention is determined for recommendation.

[0011] Preferably, after identifying the first behavioral intention that the user desires the vehicle to perform based on the voice control command, the method further includes:

[0012] Control the vehicle to perform an operation corresponding to the first behavioral intention.

[0013] Preferably, the several recommended intents include empty intents.

[0014] Preferably, the behavior sequence prediction model is trained in the following manner:

[0015] Collect real behaviors in the car;

[0016] Mining the association between intentions based on the real in-vehicle behaviors, and constructing an intention sequence based on the relevance of intentions;

[0017] Constructing a behavior intention tree according to the intention sequence;

[0018] The behavior sequence initial model is trained using the behavior intention tree to obtain the behavior sequence prediction model; wherein the behavior sequence initial model is built using a deep learning algorithm framework.

[0019] Preferably, before calculating the recommendation probabilities of the plurality of recommendation intentions based on the model prediction probabilities of the plurality of recommendation intentions and the transition probabilities of the plurality of recommendation intentions, the method includes:

[0020] For each recommended intention, determining whether the cumulative number of occurrences of the first behavioral intention is greater than or equal to a preset threshold;

[0021] If so, determining the transition probability of the recommendation intention as the individual transition probability;

[0022] If not, the transition probability of the recommendation intention is determined to be a global transition probability.

[0023] Preferably, the individual transition probability is obtained by:

[0024] Real-time statistics are collected on the number of times the current vehicle executes the first behavioral intention, and the number of times the current vehicle executes the recommended intention after executing the first behavioral intention;

[0025] The individual transition probability is obtained based on the number of times the current vehicle executes the recommended intention after executing the first behavioral intention and the number of times the current vehicle executes the first behavioral intention.

[0026] Preferably, the global transition probability is obtained by:

[0027] Real-time statistics are collected on the number of times all vehicles execute the first behavioral intention, and the number of times all vehicles execute the recommended intention after executing the first behavioral intention;

[0028] The global transition probability is obtained based on the number of times that all vehicles execute the recommended intention after executing the first behavioral intention and the number of times that all vehicles execute the first behavioral intention.

[0029] Preferably, the step of calculating the recommendation probabilities of the plurality of recommendation intentions based on the model prediction probabilities of the plurality of recommendation intentions and the transition probabilities of the plurality of recommendation intentions specifically includes:

[0030] For each recommendation intent, the model prediction probability and transition probability of the recommendation intent are processed using the UCB algorithm based on the confidence upper bound to calculate the recommendation probability of the recommendation intent.

[0031] Preferably, the method of processing the model prediction probability and transition probability of the recommendation intention using the confidence upper bound UCB algorithm to calculate the recommendation probability of the recommendation intention specifically includes:

[0032] Based on the model prediction probability of the recommendation intention, the transition probability of the recommendation intention and the respective hyperparameters, a weighted sum is performed to obtain a probability sum value;

[0033] Summing the hyperparameters of the model prediction probability and the respective hyperparameters of the transition probability to obtain a hyperparameter sum value;

[0034] The ratio of the probability sum value to the hyperparameter sum value is incorporated into the upper confidence bound of the UCB algorithm as the experimental mean to calculate the recommendation probability of the recommendation intent.

[0035] Preferably, determining the second behavioral intention for recommendation based on the recommendation probabilities of the plurality of recommendation intentions specifically includes:

[0036] According to the recommendation probabilities of the respective recommendation intentions, the second behavior intention with the highest probability is determined, and the recommendation is made after the first behavior intention is executed.

[0037] In a second aspect of this specification, a voice interactive recommendation system is disclosed, the system comprising:

[0038] a collection module, configured to collect a voice control command issued by a user and identify a first behavioral intention that the user desires the current vehicle to perform based on the voice control command;

[0039] an obtaining module, configured to obtain a plurality of recommendation intentions based on the first behavioral intention;

[0040] A prediction module, configured to process the behavior sequence prediction model based on the first behavior intention input to obtain a plurality of recommended intentions and their model prediction probabilities;

[0041] a calculation module, configured to calculate the recommendation probability of each of the plurality of recommendation intentions based on the model prediction probabilities of each of the plurality of recommendation intentions and the transition probabilities of each of the plurality of recommendation intentions; the transition probabilities of each of the plurality of recommendation intentions being global transition probabilities or individual transition probabilities;

[0042] The recommendation module is used to determine the second behavior intention for recommendation based on the recommendation probabilities of the plurality of recommendation intentions.

[0043] A third aspect of this specification discloses a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0044] In a fourth aspect of this specification, an in-vehicle device is disclosed, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0045] Through one or more technical solutions of this specification, this specification has the following beneficial effects or advantages:

[0046] This specification discloses a voice interactive recommendation method, which collects voice control commands issued by the user and identifies the first behavioral intention that the user expects the current vehicle to perform based on the voice control command. Then, based on the first behavioral intention, a behavior sequence prediction model is input for processing to obtain several recommended intentions and their model prediction probabilities; the recommendation probability of each of the several recommended intentions is calculated based on the model prediction probability of each of the several recommended intentions and the transition probability of each of the several recommended intentions; the transition probability of each of the several recommended intentions is a global transition probability or an individual transition probability; and based on the recommendation probability of each of the several recommended intentions, a second behavioral intention is determined for recommendation. The technical solution in this specification, based on determining several recommended intentions that are relevant to the first behavioral intention, accurately mines the user's subsequent intentions after a single round of voice dialogue by combining the model prediction probability and the transition probability, thereby achieving in-depth interaction with the user over multiple rounds of dialogue, thereby making the vehicle-mounted equipment more intelligent.

[0047] Furthermore, based on the UCB algorithm, a decision probability formula is constructed by combining the model prediction probability and the transition probability, and the recommendation probability of the recommendation intention is calculated using the decision probability formula. Since the transition probability is divided into individual transition probability and global transition probability, this manual combines the two parts and the model output for overall mining based on the UCB algorithm, which can meet the overall user needs and user personalized needs according to different situations, thereby outputting the recommendation intention more accurately.

[0048] The above description is only an overview of the technical solution of this specification. In order to more clearly understand the technical means of this specification, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of this specification more obvious and easy to understand, the specific implementation methods of this specification are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present description. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0050] Figure 1 A flowchart of a voice interactive recommendation method according to one embodiment of the present specification is shown;

[0051] Figure 2 A schematic diagram of the structure of a behavior intention tree according to an embodiment of this specification is shown;

[0052] Figure 3 FIG2 shows a schematic diagram of an execution logic for calculating a recommendation probability of a recommendation intention according to one embodiment of the present specification;

[0053] Figure 4 A logical diagram of voice interactive recommendation according to one embodiment of the present specification is shown;

[0054] Figure 5 A schematic diagram of a voice interactive recommendation system according to an embodiment of the present specification is shown. DETAILED DESCRIPTION

[0055] The following describes exemplary embodiments of the present specification in more detail with reference to the accompanying drawings. Although exemplary embodiments of the present specification are shown in the accompanying drawings, it should be understood that the present specification can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present specification and to fully convey the scope of the present specification to those skilled in the art.

[0056] seq2seq: Sequence-to-sequence generative model.

[0057] UCB: upper confidence bound, cold start algorithm.

[0058] vin: Vehicle frame number, used to distinguish vehicle ID.

[0059] Apriori: An association mining algorithm.

[0060] transformer: NLP (natural language processing) deep learning algorithm framework.

[0061] See Figure 1 The embodiment of this specification discloses a voice interactive recommendation method, which includes the following steps:

[0062] Step 101 : collecting a voice control command issued by a user, and identifying a first behavioral intention that the user expects the vehicle to perform based on the voice control command.

[0063] In this embodiment, the microphone in the vehicle-mounted device captures voice control commands and identifies the corresponding first behavioral intent. For example, if a user issues the voice control command "close the windows," the first behavioral intent can be identified as "close the windows." The vehicle is then controlled to perform the action corresponding to the first behavioral intent.

[0064] To further explore the user's subsequent intentions after the first behavioral intention is expected to be executed, after the first behavioral intention is identified, the VIN code of the vehicle to which it belongs is obtained and an intention recommendation request is generated. The intention recommendation request contains at least: the first behavioral intention and the VIN code.

[0065] Step 102: Input the behavior sequence prediction model based on the first behavior intention to perform processing to obtain a number of recommended intentions and their model prediction probabilities.

[0066] In this embodiment, the behavior sequence prediction model can be trained offline. The network structure of the initial behavior sequence model is built using the Transformer deep learning algorithm framework. Actual behavior intentions are derived from actual user travel records. Incorporating the seq2seq concept, various related behavior intentions are used as nodes for model training to obtain the behavior sequence prediction model. The behavior sequence prediction model takes behavior intention as input and outputs the subsequent intentions of that behavior intention and the probability of their association. The association probability represents the correlation between the input behavior intention and the output subsequent intention. The loss function uses softmax.

[0067] Specifically, collect the real behavior in the car. Based on the real behavior in the car, the association between intentions is mined. For example, the Apriori algorithm is used to mine the intention to the subsequent intention, and the association between the intentions can be mined. Based on this, the intentions related to the business scenario are extracted, and the intention sequence is constructed according to the intention correlation. The behavior intention tree is constructed according to the intention sequence; the behavior intention tree of this embodiment has a directionality, such as Figure 2 As shown in the figure, the structure diagram of the behavior intention tree is shown. Further, the cleaning step can be performed based on the user's real travel records, and then the behavior intention tree is used to train the behavior sequence initial model to obtain the behavior sequence prediction model. For example, the behavior intention tree has a sequence ( Figure 2 (not shown): behavior intention A → behavior intention B. The input of this embodiment is behavior intention A, which is used to predict behavior intention B. Of course, the input can also be a behavior intention sequence, such as a behavior intention tree with a sequence ( Figure 2 (not shown): behavior intention A → behavior intention B → behavior intention A → behavior intention C → behavior intention D. The input of this embodiment is behavior intention A → behavior intention B → behavior intention A, which is used to predict behavior intention C and behavior intention D. Furthermore, the behavior intention tree can be used to test and adjust the behavior sequence prediction model until it meets the requirements.

[0068] In this embodiment, the first behavioral intention and the VIN code are combined to obtain several recommended intentions. Specifically, all recommended intentions associated with the vehicle are obtained using the VIN code of the vehicle. The first behavioral intention is then input into the behavior sequence prediction model for processing to obtain several associated recommended intentions and their model-predicted probabilities.

[0069] Furthermore, an operation of filtering multiple recommended intents can be performed. Assume that the multiple recommended intents include all recommended intents associated with the first behavior intent. For example, in Table 1, recommended intent B, empty intent, recommended intent C, recommended intent E, recommended intent F, and so on are included.

[0070] Table 1

[0071] vin First behavioral intention Recommendation Intent Number of recommendations 11 A B 11 11 A C 10 11 A E 11 11 A F 10 11 A Empty Intent 10

[0072] Based on Table 1, several recommended intents are filtered according to the filtering mechanism, and recommended intent B, empty intent, and recommended intent E are obtained as the required recommended intents.

[0073] The filtering mechanism can be configured independently. If not, the default filtering mechanism will be used. The filtering mechanism can be: filtering the executed recommended intents, but this is not a limitation. Continuing with the above example, if the recommended intent C is to turn on the air conditioner, and the air conditioner in the car is already turned on, the recommended intent C can be filtered out. It is worth noting that, regardless of whether the recommended intents in this embodiment are filtered or not, there are always empty intents. The purpose of the empty intent is to end the mining of the user's subsequent intentions to avoid excessive interaction that disturbs the user and causes a poor user experience.

[0074] Step 103 : Calculate the recommendation probabilities of the recommendation intentions according to the model prediction probabilities of the recommendation intentions and the transition probabilities of the recommendation intentions.

[0075] In this embodiment, the transition probability of each of the plurality of recommendation intentions is a global transition probability or an individual transition probability.

[0076] The global transition probability is the probability that all vehicles will execute the recommended intention after executing the first behavioral intention. The global transition probability is obtained by: real-time counting the number of times all vehicles execute the first behavioral intention, and the number of times all vehicles execute the recommended intention after executing the first behavioral intention; based on the number of times all vehicles execute the recommended intention after executing the first behavioral intention and the number of times all vehicles execute the first behavioral intention, the global transition probability is obtained. For example, the number of times all vehicles execute the first behavioral intention A within a certain period of time is A=50 times, and the number of times they execute the recommended intention B after executing A is B|A=20, then P ucb (B|A)=20 / 50=0.4.

[0077] The individual transition probability is the probability that the current vehicle will execute the recommended intention after executing the first behavioral intention. The individual transition probability is calculated by counting in real time the number of times the current vehicle executes the first behavioral intention and the number of times the current vehicle executes the recommended intention after executing the first behavioral intention; and then calculating the individual transition probability based on the number of times the current vehicle executes the recommended intention after executing the first behavioral intention and the number of times the current vehicle executes the first behavioral intention. The individual transition probability is calculated in the same manner as described above, differing only in the statistical object, so it will not be detailed here.

[0078] For each recommended intention, determine whether the cumulative number of occurrences of the first behavioral intention is greater than or equal to a preset threshold; if so, determine the transition probability of the recommended intention as an individual transition probability; if not, determine the transition probability of the recommended intention as a global transition probability.

[0079] In this embodiment, the core technical difficulty of recommendation is to accurately locate the user's subsequent needs. For example, after satisfying the user's need to turn on the air conditioner, accurately recommend the subsequent needs that the user expects to perform. In this regard, considering that there will be uncontrollable probabilities when only using algorithmic decisions, this embodiment, based on the UCB (Upper Confidence Bound, based on the upper confidence bound) algorithm, combines the model prediction probability and the transition probability to construct a decision probability formula, and uses the decision probability formula to calculate the recommendation probability of the recommendation intention. Since the transition probability is divided into individual transition probability and global transition probability, this embodiment combines the two parts and the model output on the basis of the UCB algorithm for overall mining, which can meet the overall user needs and user personalized needs according to different situations, thereby outputting the recommendation intention more accurately. In addition, since there will be a probability of non-execution of the action in each next round of recommendations, it can avoid excessive interaction that disturbs the user and causes a poor user experience.

[0080] In the specific implementation process, a weighted sum is performed based on the model-predicted probability of the recommendation intent, the transition probability of the recommendation intent, and their respective hyperparameters to obtain a probability sum value; the hyperparameters of the model-predicted probability and the hyperparameters of the transition probability are summed to obtain a hyperparameter sum value; the ratio of the probability sum value to the hyperparameter sum value is incorporated into the upper confidence bound of the UCB algorithm as the experimental mean to calculate the recommendation probability of the recommendation intent. To further illustrate and explain this implementation process, the constructed decision probability formula is used for illustration below.

[0081] The decision probability formula is:

[0082]

[0083] The decision probability formula is derived by combining the UCB algorithm, the model prediction probability, and the transition probability. The decision probability formula is used to calculate several recommendation intents and obtain the recommendation probabilities for each of them.

[0084] Among them, P represents the recommendation probability of the recommendation intention, A represents the first behavior intention, B represents the recommendation intention, and P 转移 (B|A) represents the transition probability of recommendation intention, P 转移 (B|A)=P ucb (B|A) or P 转移 (B|A)=P ucbi (B|A), P ucb (B|A) represents the global transition probability, P ucbi (B|A) represents the individual transition probability, i represents the user id, P model (B|A) represents the model output probability, α A , β ARepresents a hyperparameter, optional, α A +β A =1, n represents the total number of recommendations, n B Indicates the number of recommendations for recommendation intent B.

[0085] To further illustrate and explain the present invention, please refer to the following Figure 3 , which is a schematic diagram of the execution logic for calculating the recommendation probability of a recommendation intent. In this example, the first behavior intent A and the recommended intent B are used as examples.

[0086] Step 301: Identify the first behavioral intention A.

[0087] Step 302: Using the first behavior intention A and the vin code, the behavior sequence prediction model is combined to obtain the recommended intention B and the model output probability P of the recommended intention B. model (B|A).

[0088] Step 303: Determine whether the cumulative number of occurrences of the first behavioral intention A is greater than or equal to a preset threshold. If so, execute step 304 and call the individual transition probability P ucbi (B|A). If not, execute step 305 and call the global transition probability P ucb (B|A). Among them, P 转移 =P ucbi (B|A) or P ucb (B|A).

[0089] Step 306: Get the total number of recommendations n and the number of recommendations n for recommendation intention B B .

[0090] Step 307: Execute the decision probability formula to calculate the recommendation probability of recommendation intention B. The decision probability formula is shown in the above formula (1).

[0091] If there are multiple recommendation intentions, they can be input into the decision probability formula respectively to calculate their respective recommendation probabilities.

[0092] In some optional embodiments, after determining the recommendation probabilities for each of several recommendation intents, business rules are used to determine whether to continue recommending. For example, a business rule may limit the number of recommendations to one per day; another example may limit the number of times a user rejects a recommendation to two within a preset time period. If any of these conditions are met, recommendations are discontinued. For example, if a recommendation has already been made once that day, recommendations are discontinued. Otherwise, recommendations can continue.

[0093] Step 104 : Determine the second behavior intention and make a recommendation based on the recommendation probabilities of the plurality of recommendation intentions.

[0094] In this embodiment, based on the recommendation probabilities of the respective recommendation intentions, the second behavioral intention with the highest probability is determined, and the recommendation is made after the first behavioral intention is executed.

[0095] Specifically, the second behavior intention with the highest probability is recommended by sorting the recommendation probability values. The recommendation method uses voice broadcast to ask the user whether the second behavior intention needs to be executed, and performs relevant operations based on the user's response.

[0096] Furthermore, after the recommendation is made, relevant information about the recommendation intent will be counted, such as the total number of recommendations, the number of recommendations for each recommendation intent, and the number of successful recommendations. Business rules will be updated based on the relevant information about the recommendation intent.

[0097] To further illustrate and explain this specification, see Figure 4 , is a logical diagram of voice interactive recommendation in this embodiment.

[0098] Step 401: Identify the first behavioral intention A.

[0099] In step 402, the first behavioral intention A and the VIN code are combined with the behavioral sequence prediction model to obtain several recommended intentions and their model-predicted probabilities. The recommended intentions are shown in Table 1. For example, recommended intention B, no intention, recommended intention C, recommended intention E, recommended intention F, and so on.

[0100] Step 403: Filter using the filtering mechanism. After filtering, recommended intent B, empty intent, and recommended intent E are obtained as several recommended intents.

[0101] Step 404: Execute a decision operation. Specifically, based on the UCB algorithm, a decision probability formula is obtained by combining the global transition probability, the individual transition probability, and the model prediction probability. The decision probability formula is then used to calculate the recommendation probability of each of the multiple recommendation intents.

[0102] Step 405: Filter using business rules. For example, the recommended intent B, the empty intent, and the recommended intent E are filtered using business rules, and the recommendation probability of the recommended intent B and the empty intent are retained.

[0103] Step 406: Perform a sorting operation. Specifically, sort the recommendation probabilities of the recommended intents to obtain the second behavioral intent with the highest probability. For example, sort the recommendation probabilities of recommended intent B and the empty intent. It is worth noting that the order in which the business rule filtering and sorting operations are performed is not restricted.

[0104] Step 407: Output the second behavioral intention. Specifically, the user may be asked in a voice announcement whether to perform the operation corresponding to the second behavioral intention, and the business rules are updated accordingly.

[0105] The above is an explanation of the voice interactive recommendation method of this manual. The solution in this manual combines model prediction probability and transition probability to accurately explore the user's subsequent intentions after a single round of voice conversation, achieve deep interaction with the user over multiple rounds of conversation, and thus make the in-vehicle equipment more intelligent.

[0106] Based on the same inventive concept as in the above embodiment, this specification also discloses a voice interactive recommendation system. Figure 5 , the system comprising:

[0107] The acquisition module 501 is configured to acquire a voice control command issued by a user and identify a first behavioral intention that the user desires the current vehicle to perform based on the voice control command;

[0108] Prediction module 502, configured to process the behavior sequence prediction model based on the first behavior intention input to obtain a plurality of recommended intentions and their model prediction probabilities;

[0109] A calculation module 503 is configured to calculate a recommendation probability for each of the plurality of recommendation intentions based on the model prediction probabilities of the plurality of recommendation intentions and the transition probabilities of the plurality of recommendation intentions; the transition probabilities of the plurality of recommendation intentions are global transition probabilities or individual transition probabilities;

[0110] The recommendation module 504 is configured to determine a second behavioral intention for recommendation based on the recommendation probabilities of the plurality of recommendation intentions.

[0111] Based on the same inventive concept as in the aforementioned embodiment, an embodiment of this specification further discloses a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the aforementioned methods when executed by a processor.

[0112] Based on the same inventive concept as in the aforementioned embodiments, an embodiment of this specification also discloses a vehicle-mounted device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the aforementioned methods are implemented.

[0113] Through one or more embodiments of this specification, this specification has the following beneficial effects or advantages:

[0114] This specification discloses a voice interactive recommendation method, which collects voice control commands issued by the user and identifies the first behavioral intention that the user expects the current vehicle to perform based on the voice control command. Then, based on the first behavioral intention, a behavior sequence prediction model is input for processing to obtain several recommended intentions and their model prediction probabilities; the recommendation probability of each of the several recommended intentions is calculated based on the model prediction probability of each of the several recommended intentions and the transition probability of each of the several recommended intentions; the transition probability of each of the several recommended intentions is a global transition probability or an individual transition probability; and based on the recommendation probability of each of the several recommended intentions, a second behavioral intention is determined for recommendation. The technical solution in this specification, based on determining several recommended intentions that are relevant to the first behavioral intention, accurately mines the user's subsequent intentions after a single round of voice dialogue by combining the model prediction probability and the transition probability, thereby achieving in-depth interaction with the user over multiple rounds of dialogue, thereby making the vehicle-mounted equipment more intelligent.

[0115] Furthermore, based on the UCB algorithm, a decision probability formula is constructed by combining the model prediction probability and the transition probability, and the recommendation probability of the recommendation intention is calculated using the decision probability formula. Since the transition probability is divided into individual transition probability and global transition probability, this manual combines the two parts and the model output for overall mining based on the UCB algorithm, which can meet the overall user needs and user personalized needs according to different situations, thereby outputting the recommendation intention more accurately.

[0116] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems may also be used in conjunction with the teachings herein. Based on the above description, the required structure for constructing such systems is readily apparent. Furthermore, this specification is not directed to any particular programming language. It should be understood that the contents of this specification described herein may be implemented using a variety of programming languages, and the above description of specific languages ​​is intended to disclose the best mode of implementation of this specification.

[0117] In the description provided herein, numerous specific details are described. However, it is understood that the embodiments of this description may be practiced without these specific details. In some instances, well-known methods, structures, and techniques are not shown in detail so as not to obscure the understanding of this description.

[0118] Similarly, it should be understood that in order to streamline this specification and aid understanding of one or more of the various inventive aspects, in the above description of the exemplary embodiments of this specification, various features of this specification are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting an intention that the claimed specification requires more features than are expressly recited in each claim. Rather, as reflected in the claims below, inventive aspects lie in less than all the features of the individual embodiments disclosed above. Accordingly, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of this specification.

[0119] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and in addition may be divided into multiple submodules or subunits or subcomponents. All features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed herein may be combined in any combination, except that at least some of such features and / or processes or units are mutually exclusive. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0120] Furthermore, those skilled in the art will appreciate that although some embodiments herein include certain features included in other embodiments but not other features, combinations of features from different embodiments are intended to be within the scope of this specification and to form different embodiments. For example, in the claims below, any of the claimed embodiments may be used in any combination.

[0121] The various component embodiments of this specification can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It will be appreciated by those skilled in the art that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all of the functions of some or all of the components in the gateway, proxy server, or system according to the embodiments of this specification. This specification can also be implemented as a device or apparatus program (e.g., computer program and computer program product) for executing part or all of the methods described herein. Such a program implementing this specification can be stored on a computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0122] It should be noted that the above embodiments illustrate rather than limit this specification, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between brackets should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This specification may be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

Claims

1. A voice interactive recommendation method, comprising: collecting a voice control command issued by a user, and identifying a first behavioral intention that the user desires the current vehicle to perform based on the voice control command; Inputting the behavior sequence prediction model into the first behavior intention for processing, and obtaining a number of recommended intentions and their model prediction probabilities; The recommendation probabilities of the several recommended intentions are calculated based on the model prediction probabilities of the several recommended intentions and the transition probabilities of the several recommended intentions. The transition probabilities of the several recommended intentions are global transition probabilities or individual transition probabilities. For each recommended intention, it is determined whether the cumulative number of occurrences of the first behavioral intention is greater than or equal to a preset threshold. If so, the transition probability of the recommended intention is determined to be the individual transition probability. If not, the transition probability of the recommended intention is determined to be the global transition probability. The global transition probability is the probability that all vehicles will execute the recommended intention after executing the first behavioral intention. The individual transition probability is the probability that the current vehicle will execute the recommended intention after executing the first behavioral intention. According to the recommendation probabilities of the plurality of recommendation intentions, a second behavior intention is determined for recommendation.

2. The method according to claim 1, further comprising: after identifying the first behavioral intention that the user desires the vehicle to perform based on the voice control command; Control the vehicle to perform an operation corresponding to the first behavioral intention.

3. In the method as claimed in claim 1, the plurality of recommended intents include empty intents.

4. The method according to claim 1, wherein the behavior sequence prediction model is trained in the following manner: Collect real behaviors in the car; Mining the association between intentions based on the real in-vehicle behaviors, and constructing an intention sequence based on the relevance of intentions; Constructing a behavior intention tree according to the intention sequence; The behavior sequence initial model is trained using the behavior intention tree to obtain the behavior sequence prediction model; wherein the behavior sequence initial model is built using a deep learning algorithm framework.

5. The method according to claim 1, wherein the individual transition probability is obtained by: Real-time statistics are collected on the number of times the current vehicle executes the first behavioral intention, and the number of times the current vehicle executes the recommended intention after executing the first behavioral intention; The individual transition probability is obtained based on the number of times the current vehicle executes the recommended intention after executing the first behavioral intention and the number of times the current vehicle executes the first behavioral intention.

6. The method according to claim 1, wherein the global transition probability is obtained by: Real-time statistics are collected on the number of times all vehicles execute the first behavioral intention, and the number of times all vehicles execute the recommended intention after executing the first behavioral intention; The global transition probability is obtained based on the number of times that all vehicles execute the recommended intention after executing the first behavioral intention and the number of times that all vehicles execute the first behavioral intention.

7. The method of claim 5, wherein calculating the recommendation probabilities of the plurality of recommendation intentions based on the model prediction probabilities of the plurality of recommendation intentions and the transition probabilities of the plurality of recommendation intentions comprises: For each recommendation intent, the model prediction probability and transition probability of the recommendation intent are processed using the UCB algorithm based on the confidence upper bound to calculate the recommendation probability of the recommendation intent.

8. The method of claim 7, wherein the method uses a confidence upper bound (UCB) algorithm to process the model prediction probability and transition probability of the recommendation intent to calculate the recommendation probability of the recommendation intent, specifically comprising: Based on the model prediction probability of the recommendation intention, the transition probability of the recommendation intention and the respective hyperparameters, a weighted sum is performed to obtain a probability sum value; Summing the hyperparameters of the model prediction probability and the respective hyperparameters of the transition probability to obtain a hyperparameter sum value; The ratio of the probability sum value to the hyperparameter sum value is incorporated into the upper confidence bound of the UCB algorithm as the experimental mean to calculate the recommendation probability of the recommendation intent.

9. The method of claim 1, wherein determining the second behavioral intention for recommendation based on the recommendation probabilities of the plurality of recommendation intentions specifically comprises: According to the recommendation probabilities of the respective recommendation intentions, the second behavior intention with the highest probability is determined, and the recommendation is made after the first behavior intention is executed.

10. A voice interactive recommendation system, comprising: a collection module, configured to collect a voice control command issued by a user and identify a first behavioral intention that the user desires the current vehicle to perform based on the voice control command; an obtaining module, configured to obtain a plurality of recommendation intentions based on the first behavioral intention; A prediction module, configured to process the behavior sequence prediction model based on the first behavior intention input to obtain a plurality of recommended intentions and their model prediction probabilities; a calculation module, configured to calculate the recommendation probability of each of the plurality of recommended intentions based on the model prediction probabilities of each of the plurality of recommended intentions and the transition probabilities of each of the plurality of recommended intentions; the transition probabilities of each of the plurality of recommended intentions are global transition probabilities or individual transition probabilities; wherein, for each recommended intention, it is determined whether the cumulative number of occurrences of the first behavioral intention is greater than or equal to a preset threshold; if so, the transition probability of the recommended intention is determined to be the individual transition probability; if not, the transition probability of the recommended intention is determined to be the global transition probability; the global transition probability is the probability that all vehicles will execute the recommended intention after executing the first behavioral intention; the individual transition probability is the probability that the current vehicle will execute the recommended intention after executing the first behavioral intention; The recommendation module is used to determine the second behavior intention for recommendation based on the recommendation probabilities of the plurality of recommendation intentions.

11. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

12. An in-vehicle device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 9 when executing the program.

Images