Speech robot call method and device using high generalization multi-task intent recognition

Abstract

The application relates to a speech robot call method and device using high generalization multi-task intention recognition. The method comprises the following steps: a speech robot and a user establish a speech call, real-time speech text data of the user is acquired during the call; the real-time speech text data is input into a multi-task intention recognition model to generate replacement text, an intention category and predicted text; the accuracy of the multi-task intention recognition model is determined according to the replacement text and the predicted text; when the accuracy is greater than a threshold value, communication tactics are updated based on the intention category; the speech robot continues the speech call with the user according to the updated communication tactics. The method disclosed by the application can make the intention model have stronger generalization ability, so that the user intention can be more accurately analyzed, and smoother speech communication with the user can be realized.

Description

Technical Field

[0001] This application relates to the field of computer information processing, and more specifically, to a voice robot communication method, apparatus, electronic device, and computer-readable medium employing highly generalized multi-task intent recognition. Background Technology

[0002] Intelligent voice robots, based on technologies such as speech recognition and synthesis, machine learning, and natural language understanding, automatically initiate outbound telephone calls according to business scenarios. They collect business results through voice dialogue interaction between humans and the robot, statistically process the data, and obtain user feedback. Intelligent voice robots are conversational intelligent robots for developers, enabling intelligent conversations based on natural language processing (NLP) across various messaging platforms, such as websites, apps, and physical robots. Users can configure their own unique knowledge base for intelligent question answering and integrate with third-party APIs through multi-turn dialogues to achieve self-service, such as order inquiries, logistics tracking, and self-service return robots. Intelligent voice robots can analyze dialogue content from recordings or text based on intelligent rules, uncovering potential problems and opportunities. This helps enterprises improve service quality, monitor public opinion risks, and optimize service strategies; typical application scenarios include intelligent customer service quality inspection and sales opportunity analysis.

[0003] Intent recognition is a crucial branch of natural language understanding and plays a vital role in robotics. Generally, the most common intent recognition methods currently available utilize pre-trained text models. That is, text and intent data are fine-tuned on a pre-trained model (which may be downloaded or retrained) to obtain the intent recognition model. However, this method has a significant drawback: during fine-tuning, changes in model parameters can destroy the original information in the pre-trained model. For these reasons, the accuracy of intent recognition models trained in current technologies for recognizing user intent is not high.

[0004] Therefore, there is a need for a new method, device, electronic device, and computer-readable medium for voice robot communication that employs highly generalized multi-task intent recognition.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] In view of this, this application provides a voice robot call method, device, electronic device and computer-readable medium that employs high generalization multi-task intent recognition. It can embed the multi-task training method into the training process of the intent model, prevent the model obtained in the pre-training stage from losing too much pre-training information when training the intent model, make the intent model more generalized, and thus be able to analyze user intent more accurately, thereby enabling smoother voice communication with users.

[0007] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0008] According to one aspect of this application, a voice robot call method employing high-generalization multi-task intent recognition is proposed. The method includes: establishing a voice call between a voice robot and a user, acquiring the user's real-time voice-text data during the call; inputting the real-time voice-text data into a multi-task intent recognition model to generate replacement text, intent category, and predicted text; determining the accuracy of the multi-task intent recognition model based on the replacement text and the predicted text; updating the communication script based on the intent category when the accuracy is greater than a threshold; and continuing the voice call with the user based on the updated communication script.

[0009] Optionally, the method further includes: acquiring historical speech-text data; setting intent category labels for the historical speech-text data; and training the improved BERT language model based on multi-task using the historical speech-text data with intent category labels to generate the multi-task intent recognition model.

[0010] Optionally, the multi-task intent recognition model is generated by training a BERT language model improved based on multi-task using historical speech-text data with intent category labels. This includes: training the BERT language model with historical speech-text data with intent category labels to generate a pre-trained model; replacing some characters in the historical speech-text data to generate input text data; inputting the input text data into the pre-trained model improved based on multi-task to obtain predicted text and predicted intent classification; and generating the multi-task intent recognition model when the loss functions corresponding to the predicted text and predicted intent classification satisfy a preset strategy.

[0011] Optionally, a pre-trained model is generated by training the BERT language model using historical speech and text data with intent category labels, including: converting all characters in the historical speech and text data into word vectors to generate word embedding tensors, sentence segmentation tensors, and position encoding tensors for each character; generating character vectors for each character based on word embedding tensors, sentence segmentation tensors, and position encoding tensors; and training the BERT language model using character word vectors with intent category labels to generate the pre-trained model.

[0012] Optionally, a partial character replacement operation is performed on historical speech text data to generate input text data, including: extracting a preset proportion of characters for replacement; storing the partial characters before the replacement operation; and generating the input text data using the character vectors of the unreplaced characters and the replaced characters.

[0013] Optionally, inputting the input text data into the pre-trained model improved based on multi-task to obtain predicted text and predicted intent classification includes: inputting the input text data into the pre-trained model; the pre-trained model performing intent prediction on the input text based on a bidirectional encoding mechanism; the pre-trained model recognizing replaced characters in the input text based on a multi-task mechanism; and the pre-trained model generating predicted text and predicted intent classification based on the calculation results.

[0014] Optionally, when the loss functions corresponding to the predicted text and the predicted intent classification satisfy a preset strategy, the multi-task intent recognition model is generated, including: comparing the similarity between the predicted text and the replaced characters to generate a first comparison result; comparing the predicted intent classification and the intent label to generate a second comparison result; and when the first comparison result is greater than a text threshold and the second comparison result is greater than an intent threshold, generating the multi-task intent recognition model based on the current parameters of the pre-trained model.

[0015] Optionally, the voice robot establishes a voice call with the user and acquires the user's real-time voice-text data during the call, including: determining a communication script based on user information; the voice robot conducting a voice call with the user based on the communication script; and converting the user's real-time voice data into voice-text data through speech recognition.

[0016] Optionally, determining the accuracy of the multi-task intent recognition model based on the replacement text and the predicted text includes: comparing the similarity between the replacement text and the predicted text; and determining the accuracy of the intent category based on the similarity comparison result.

[0017] Optionally, it also includes: when the accuracy rate is less than or equal to a threshold, the voice robot continues to conduct voice calls with the user based on the original communication script.

[0018] According to one aspect of this application, a voice robot communication device employing high-generalization multi-task intent recognition is proposed. The device includes: a text module for establishing a voice call between the voice robot and a user, acquiring real-time voice text data of the user during the call; a recognition module for inputting the real-time voice text data into a multi-task intent recognition model to generate replacement text, intent category, and predicted text; a judgment module for determining the accuracy of the multi-task intent recognition model based on the replacement text and the predicted text; an update module for updating the communication script based on the intent category when the accuracy is greater than a threshold; and a call module for the voice robot to continue the voice call with the user based on the updated communication script.

[0019] According to one aspect of this application, an electronic device is provided, comprising: one or more processors; a storage device for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the method as described above.

[0020] According to one aspect of this application, a computer-readable medium is provided having a computer program stored thereon that, when executed by a processor, implements the method described above.

[0021] According to the voice robot call method, apparatus, electronic device, and computer-readable medium employing high-generalization multi-task intent recognition of this application, a voice robot establishes a voice call with a user, and acquires the user's real-time voice text data during the call; the real-time voice text data is input into a multi-task intent recognition model to generate replacement text, intent category, and predicted text; the accuracy of the multi-task intent recognition model is determined based on the replacement text and the predicted text; when the accuracy is greater than a threshold, the communication script is updated based on the intent category; the voice robot continues the voice call with the user according to the updated communication script. This method can embed multi-task training methods into the training process of the intent model, preventing the model obtained in the pre-training stage from losing too much pre-training information when training the intent model, making the intent model more generalizable, and thus more accurately analyzing user intent, thereby enabling smoother voice communication with the user.

[0022] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application. Attached Figure Description

[0023] The above and other objects, features, and advantages of this application will become more apparent from the detailed description of exemplary embodiments with reference to the accompanying drawings. The drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0024] Figure 1 This is a flowchart illustrating a voice robot communication method employing highly generalized multi-task intent recognition, according to an exemplary embodiment.

[0025] Figure 2 This is a flowchart illustrating a voice robot communication method employing highly generalized multi-task intent recognition, according to another exemplary embodiment.

[0026] Figure 3 This is a schematic diagram illustrating a voice robot communication method employing highly generalized multi-task intent recognition, according to another exemplary embodiment.

[0027] Figure 4 This is a schematic diagram illustrating a voice robot communication method employing highly generalized multi-task intent recognition, according to another exemplary embodiment.

[0028] Figure 5 This is a schematic diagram illustrating a voice robot communication method employing highly generalized multi-task intent recognition, according to another exemplary embodiment.

[0029] Figure 6 This is a block diagram illustrating a voice robot communication device employing highly generalized multi-task intent recognition, according to an exemplary embodiment.

[0030] Figure 7 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Detailed Implementation

[0031] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that this application will be thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0032] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0033] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0034] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0035] It should be understood that although the terms first, second, third, etc., may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Therefore, the first component discussed below may be referred to as the second component without departing from the teachings of this application. As used herein, the term "and / or" includes all combinations of any one and more of the associated listed items.

[0036] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of exemplary embodiments, and the modules or processes in the drawings are not necessarily essential for implementing this application, and therefore cannot be used to limit the scope of protection of this application.

[0037] The technical terms used in this application are explained as follows:

[0038] Intent recognition models are models that predict intent from text (audio or other information), primarily composed of deep learning models. However, in actual production processes, intent recognition systems generally use a model + rule approach for prediction.

[0039] Multi-task learning: refers to having multiple objective functions for the same neural network model. The purpose may be to perform multiple tasks simultaneously, or to enhance the ability of a particular task.

[0040] Self-supervised training refers to a training method that uses data without target labels. For example, if there is only a large amount of text, the target labels can be obtained by modifying the original data. For example, for the text "Today is Wednesday", the data can be transformed using the masking method into: data "Today is # Wednesday", label "Starry Night".

[0041] Pre-trained models are models trained using a large amount of unlabeled data and self-supervised training methods. The resulting model reveals the internal features of individual data instances. When data is input into this model, abstract features representing that data are obtained, and these abstract features have extremely strong generalization ability. Examples include BERT for text data and Wav2vec2.0 for audio data. There are generally two ways to use pre-trained models: one is to input data into the pre-trained model to obtain abstract features, and then input these abstract features into other models for training; the other is to directly fine-tune the pre-trained model.

[0042] Figure 1 This is a flowchart illustrating a voice robot call method employing highly generalized multi-task intent recognition according to an exemplary embodiment. The voice robot call method 10 employing highly generalized multi-task intent recognition includes at least steps S102 to S110.

[0043] like Figure 1 As shown, in S102, the voice robot establishes a voice call with the user, and acquires the user's real-time voice-text data during the call. For example, a communication script can be determined based on user information; the voice robot conducts a voice call with the user based on the communication script; and the user's real-time voice data is converted into voice-text data through speech recognition.

[0044] In this embodiment, the user can be an individual user or an enterprise user. The user information may include basic information authorized by the user, such as business account information, the user's terminal device identification information, and the user's geographical location. The user information may also include behavioral information, such as the user's page operation data, the user's business access duration, and the user's business access frequency. The specific content of the user information can be determined according to the actual application scenario and is not limited here. More specifically, the current user's information can be obtained through webpage tracking based on user authorization. Remote information may be user data from other trading platforms or other business departments.

[0045] In one embodiment, if the user status information in the user information shows that the user is an unregistered user, then the communication script can be determined as a lead generation registration script, and the intelligent robot conducts a voice call with the user according to the call flow configured in the lead generation registration script.

[0046] In S104, the real-time speech text data is input into the multi-task intent recognition model to generate replacement text, intent category, and predicted text.

[0047] In one embodiment, the multi-task intent model of this application can be trained and computed for multiple tasks, and can generate multiple computation results in the model computation. More specifically, the computation results can be replacement text, intent category, and predicted text. The intent category can be a label that identifies the user's intent category, such as: "very interested", "ready to apply", "not interested", "under consideration", "among comparison products".

[0048] The replacement text and the predicted text include character vectors containing multiple characters. The replacement text and the predicted text can be displayed on the user's device in text form, or they can be stored in the background for use in subsequent steps.

[0049] In S106, the accuracy of the intent category in the current multi-task intent recognition model is determined based on the replacement text and the predicted text. The replacement text and the predicted text can be compared for similarity; the accuracy of the intent category in the current multi-task intent recognition model is determined based on the similarity comparison result.

[0050] The similarity comparison can be performed one by one between multiple character vectors in the replacement text and multiple character vectors in the prediction text, generating a similarity comparison result. The similarity comparison result displays the similarity between the replacement text and the prediction text.

[0051] The correlation between similarity and the accuracy of intent recognition model calculation can be established by calculating the results of multiple similarities, thereby determining the accuracy of intent recognition model based on similarity.

[0052] In S108, when the accuracy rate is greater than the threshold, the communication script is updated based on the intent category, and the voice robot continues to conduct voice calls with the user according to the updated communication script.

[0053] When the accuracy rate is greater than the threshold, new communication scripts can be generated based on the current user's intent, and communication can be carried out based on the new communication scripts.

[0054] In S110, when the accuracy rate is less than or equal to the threshold, the voice robot continues to conduct a voice call with the user based on the original communication script.

[0055] If the accuracy rate is less than or equal to the threshold, it means that the user's intention is still unclear, and the original communication script can continue to be used.

[0056] According to the voice robot call method employing high generalization multi-task intent recognition of this application, a voice robot establishes a voice call with a user, and acquires the user's real-time voice text data during the call; the real-time voice text data is input into a multi-task intent recognition model to generate replacement text, intent category, and predicted text; the accuracy of the multi-task intent recognition model is determined based on the replacement text and the predicted text; when the accuracy is greater than a threshold, the communication script is updated based on the intent category; the voice robot continues the voice call with the user according to the updated communication script. This method can embed multi-task training methods into the training process of the intent model, preventing the model obtained in the pre-training stage from losing too much pre-training information when training the intent model, making the intent model more generalizable, and thus more accurately analyzing user intent, thereby enabling smoother voice communication with the user.

[0057] It should be clearly understood that this application describes how specific examples are formed and used, but the principles of this application are not limited to any details of these examples. Rather, based on the teachings of the disclosure of this application, these principles can be applied to many other embodiments.

[0058] In one embodiment, the method further includes: acquiring historical speech-text data; setting intent category labels for the historical speech-text data; and training a multi-task-based improved BERT language model using the historical speech-text data with intent category labels to generate the multi-task intent recognition model.

[0059] Figure 2 This is a flowchart illustrating a voice robot communication method employing highly generalized multi-task intent recognition, according to another exemplary embodiment. Figure 2 The illustrated process 20 is a detailed description of "training the multi-task-based improved BERT language model using historical speech and text data with intent category labels to generate the multi-task intent recognition model".

[0060] like Figure 2 As shown, in S202, the BERT language model is trained using historical speech-text data with intent category labels to generate a pre-trained model. For example, all characters in the historical speech-text data are transformed into word vectors to generate word embedding tensors, sentence chunking tensors, and positional encoding tensors for each character; character vectors for each character are generated based on the word embedding tensors, sentence chunking tensors, and positional encoding tensors; and the BERT language model is trained using the character word vectors with intent category labels to generate the pre-trained model.

[0061] The specific process of training the BERT language model using historical speech and text data with intent category labels to generate a pre-trained model can be described as follows: Figure 3As shown: Texts T1…T6 are masked to replace T2 and T5. The masked text is then fed to BERT, which attempts to predict the masked T2 and T5. When the loss function during training reaches the preset strategy, the pre-trained model is generated.

[0062] In S204, a partial character replacement operation is performed on the historical speech-text data to generate input text data. For example, a preset proportion of characters can be extracted for replacement; the partial characters before the replacement operation can be stored; and the input text data can be generated using the character vectors of the unreplaced characters and the replaced characters.

[0063] Continuing with the mask operation described above, we replace some characters to generate the input text data. We can also further perform word vectorization on the characters in the text data to generate a character vector for each character.

[0064] In S206, the input text data is input into the pre-trained model improved based on multi-task methods to obtain predicted text and predicted intent classifications. For example, the input text data is input into the pre-trained model; the pre-trained model performs intent prediction on the input text based on a bidirectional encoding mechanism; the pre-trained model identifies replaced characters in the input text based on a multi-task mechanism; and the pre-trained model generates predicted text and predicted intent classifications based on the calculation results.

[0065] The application process of pre-trained models in existing technologies can be as follows: Figure 4 As shown: Starting with the BERT model obtained in the pre-training stage, the intent model is trained using the [text, intent] dataset.

[0066] In one embodiment of this application, the process of inputting the input text data into the pre-trained model improved based on multi-task to obtain predicted text and predicted intent classification can be as follows: Figure 5 As shown.

[0067] In the multi-task-based intention model training method of this application, in practical applications, such as... Figure 5 As shown, the texts T1…T6 are masked, and T3 and T7 are replaced with masks. The masked text is then fed to BERT, which attempts to predict the masked T3 and T7 and calculates the user's intent classification.

[0068] Since the text input to the original intention model is not masked during the training phase, while the text input during the pre-training phase is masked, the addition of a mask operation in this application can enhance the model's generalization ability and also match the pre-training phase.

[0069] Multi-task training has been added. When training the multi-task intent model in this application, not only is it required that the BERT model can accurately identify the intent, but it is also required that the BERT model can still complete the same tasks as in the pre-training stage. In this way, the BERT model can maintain the generalization ability of the pre-training stage when training the intent model.

[0070] In S208, the multi-task intent recognition model is generated when the loss functions corresponding to the predicted text and the predicted intent classification satisfy a preset strategy. For example, the predicted text and the replaced characters are compared for similarity to generate a first comparison result; the predicted intent classification and the intent label are compared to generate a second comparison result; when the first comparison result is greater than a text threshold and the second comparison result is greater than an intent threshold, the multi-task intent recognition model is generated based on the current parameters of the pre-trained model.

[0071] Specifically, for each sample data point, a temporary multi-task intent recognition model is constructed. The object information of each object in the object set is input into the temporary multi-task intent recognition model to obtain predicted labels. The predicted labels are compared with the corresponding real labels to determine if they match. The number of predicted labels matching the real labels is counted, and the percentage of such predicted labels in the total number of predicted labels is calculated. If the percentage is greater than or equal to a preset percentage, the temporary multi-task intent recognition model converges, resulting in a trained multi-task intent recognition model. If the percentage is less than the preset percentage, the parameters in the temporary multi-task intent recognition model are adjusted, and the predicted labels of each object are re-predicted using the adjusted temporary multi-task intent recognition model until the percentage is greater than or equal to the preset percentage. The method for adjusting the parameters in the temporary multi-task intent recognition model can employ stochastic gradient descent, gradient descent, or normal equations.

[0072] If the number of times the parameters of the temporary multi-task intent recognition model are adjusted exceeds a preset number, the model used to construct the temporary multi-task intent recognition model can be replaced to improve the model training efficiency.

[0073] Currently, the most common intent recognition methods in the industry are mainly as follows: Figure 4As shown, this method utilizes a pre-trained text model. Specifically, it fine-tunes the [text, intent] data on a pre-trained model (which may be downloaded or retrained) to obtain an intent recognition model. However, this method has a significant drawback: during fine-tuning, changes in model parameters can destroy the original information in the pre-trained model. This is because intent recognition and pre-training self-supervised tasks are fundamentally different. Although pre-training aims to generate BERT parameters with strong generalization capabilities, changes in parameters during intent model training will reduce this generalization ability. Therefore, this paper proposes a method using multi-task learning to enhance the generalization ability of intent models.

[0074] In this application, by embedding a multi-task training method (the added task is consistent with the task in the pre-training stage) into the training process of the intent model, the model obtained in the pre-training stage is prevented from losing too much pre-training information when training the intent model, thereby making the intent model more generalizable.

[0075] Those skilled in the art will understand that all or part of the steps of the above embodiments are implemented as a computer program executed by a CPU. When the computer program is executed by the CPU, it performs the functions defined by the method provided in this application. The program can be stored in a computer-readable storage medium, such as a read-only memory, a magnetic disk, or an optical disk.

[0076] Furthermore, it should be noted that the above figures are merely illustrative representations of the processes included in the method according to exemplary embodiments of this application, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0077] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0078] Figure 6 This is a block diagram illustrating a voice robot communication device employing highly generalized multi-task intent recognition, according to an exemplary embodiment. Figure 6 As shown, the voice robot communication device 60 with high generalization multi-task intent recognition includes: a text module 602, a recognition module 604, a judgment module 606, an update module 608, and a communication module 610.

[0079] The text module 602 is used to establish a voice call between the voice robot and the user, and to acquire the user's real-time voice and text data during the call.

[0080] The recognition module 604 is used to input the real-time speech text data into the multi-task intent recognition model to generate replacement text, intent category and predicted text;

[0081] The judgment module 606 is used to determine the accuracy of the current multi-task intent recognition model based on the replaced text and the predicted text;

[0082] The update module 608 is used to update the communication script based on the intent category when the accuracy is greater than the threshold.

[0083] The call module 610 is used by the voice robot to continue the voice call with the user according to the updated communication script.

[0084] According to the voice robot call device employing high generalization multi-task intent recognition of this application, a voice robot establishes a voice call with a user, and acquires the user's real-time voice text data during the call; the real-time voice text data is input into a multi-task intent recognition model to generate replacement text, intent category, and predicted text; the accuracy of the multi-task intent recognition model is determined based on the replacement text and the predicted text; when the accuracy is greater than a threshold, the communication script is updated based on the intent category; the voice robot continues the voice call with the user according to the updated communication script. This method can embed multi-task training methods into the training process of the intent model, preventing the model obtained in the pre-training stage from losing too much pre-training information when training the intent model, making the intent model more generalizable, and thus more accurately analyzing user intent, thereby enabling smoother voice communication with the user.

[0085] Figure 7 This is a block diagram of an electronic device according to an exemplary embodiment, including a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740.

[0086] Memory 730 is used to store computer programs;

[0087] When the processor 710 executes the program stored in the memory 730, it implements the data allocation permission adjustment method based on video facial expression actions according to any of the above embodiments.

[0088] The electronic device provided in this embodiment of the invention includes a processor 710 that executes a program stored in a memory 730 to obtain initial data allocation permissions and access information for a target; determines video text content using the access information; establishes a real-time video link with the target and displays the video text content based on the video link to generate video data; identifies the user's facial expressions and actions in the video data to determine their corresponding permission adjustment coefficients; and adjusts the user's data allocation permissions according to the initial permissions and the permission adjustment coefficients.

[0089] The communication bus 740 mentioned in the aforementioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 740 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus.

[0090] The communication interface 720 is used for communication between the aforementioned electronic device and other devices.

[0091] The memory 730 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device 730. Optionally, the memory 730 may also be at least one storage device located remotely from the aforementioned processor 710.

[0092] The processor 710 mentioned above can be a general-purpose processor 710, including a central processing unit (CPU) 710, a network processor (NP) 710, etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0093] This invention provides a computer-readable storage medium storing one or more programs, which can be executed by one or more processors 710 to implement the voice robot call method with high generalization multi-task intent recognition as described in any of the above embodiments. For example, the voice robot establishes a voice call with a user, acquiring the user's real-time voice text data during the call; inputting the real-time voice text data into a multi-task intent recognition model to generate replacement text, intent category, and predicted text; determining the accuracy of the intent category in the current multi-task intent recognition model based on the replacement text and the predicted text; updating the communication script based on the intent category when the accuracy is greater than a threshold; and continuing the voice call with the user based on the updated communication script.

[0094] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0095] Exemplary embodiments of this application have been specifically shown and described above. It should be understood that this application is not limited to the detailed structures, arrangements, or implementation methods described herein; rather, this application is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A voice robot communication method employing high-generalization multi-task intent recognition, characterized in that, include: The voice robot establishes a voice call with the user and obtains the user's real-time voice text data during the call; The real-time speech and text data is input into a multi-task intent recognition model. A preset proportion of characters are extracted from the real-time speech and text data and replaced to generate input text data containing the replaced characters. Intent prediction is performed on the input text data to generate intent categories. The replaced characters in the input text data are predicted to generate predicted text corresponding to the input text data. The similarity between the replaced character and the predicted text is compared. The accuracy of intent categories in this multi-task intent recognition model is determined based on the similarity comparison results. When the accuracy rate is greater than the threshold, the communication script is updated based on the intent category; The voice robot continues to conduct voice calls with the user based on the updated communication script.

2. The method as described in claim 1, characterized in that, Also includes: Obtain historical voice and text data; Set intent category labels for the historical voice and text data; The multi-task intent recognition model is generated by training a multi-task-based improved BERT language model with historical speech and text data labeled with intent categories.

3. The method as described in claim 2, characterized in that, The multi-task intent recognition model is generated by training a multi-task-based improved BERT language model on historical speech-text data with intent category labels, including: A pre-trained model is generated by training the BERT language model using historical speech and text data with intent category labels; Replace some characters in the historical speech-text data to generate the input text data; The input text data is fed into the pre-trained model based on multi-task improvement to obtain predicted text and predicted intent classification. The multi-task intent recognition model is generated when the loss functions corresponding to the predicted text and the predicted intent classification satisfy the preset strategy.

4. The method as described in claim 3, characterized in that, Pre-trained models are generated by training the BERT language model using historical speech-text data with intent category labels, including: All characters in the historical speech and text data are transformed into word vectors to generate word embedding tensors, sentence block tensors, and position encoding tensors for each character; Generate a character vector for each character based on word embedding tensor, sentence segmentation tensor, and position encoding tensor; The pre-trained model is generated by training the BERT language model with character word vectors labeled with intent categories.

5. The method as described in claim 3, characterized in that, The input text data is generated by replacing some characters in the historical speech-text data, including: Extract characters at a preset ratio and perform a replacement operation; Store the partial characters before the replacement operation; The input text data is generated using character vectors of the unreplaced characters and the replaced characters.

6. The method as described in claim 3, characterized in that, The input text data is fed into the pre-trained model improved based on multi-task methods to obtain predicted text and predicted intent classifications, including: The input text data is input into the pre-trained model; The pre-trained model predicts intent from the input text based on a bidirectional encoding mechanism; The pre-trained model identifies the replaced characters in the input text based on a multi-task mechanism; The pre-trained model generates predicted text and predicted intent classification based on the calculation results.

7. The method as described in claim 3, characterized in that, When the loss functions corresponding to the predicted text and the predicted intent classification satisfy a preset strategy, the multi-task intent recognition model is generated, including: The predicted text and the replaced characters are compared for similarity to generate the first comparison result; The predicted intent classification and intent label are compared to generate a second comparison result; When the first comparison result is greater than the text threshold and the second comparison result is greater than the intent threshold, the multi-task intent recognition model is generated based on the current parameters of the pre-trained model.

8. The method as described in claim 1, characterized in that, The voice robot establishes a voice call with the user, and during the call, it acquires the user's real-time voice and text data, including: Determine the communication script based on user information; The voice robot conducts voice calls with the user based on the communication script; Voice recognition converts users' real-time voice data into voice-text data.

9. The method as described in claim 1, characterized in that, Also includes: When the accuracy rate is less than or equal to the threshold, the voice robot continues to conduct voice calls with the user based on the original communication script.

10. A voice robot communication device employing high-generalization multi-task intent recognition, characterized in that, include: The text module is used to establish voice calls between the voice robot and the user, and to obtain the user's real-time voice and text data during the call; The recognition module is used to input the real-time speech and text data into the multi-task intent recognition model, extract a preset proportion of characters from the real-time speech and text data and replace them to generate input text data containing the replaced characters, perform intent prediction on the input text data to generate intent categories, and predict the replaced characters in the input text data to generate predicted text corresponding to the input text data. The judgment module is used to compare the similarity between the replaced character and the predicted text; The accuracy of intent categories in this multi-task intent recognition model is determined based on the similarity comparison results. The update module is used to update the communication script based on the intent category when the accuracy rate is greater than the threshold. The call module is used by the voice robot to continue the voice call with the user based on the updated communication script.

11. An electronic device, characterized in that, include: One or more processors; Storage device for storing 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 method as described in any one of claims 1 to 9.

12. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 9.

Images