Hierarchical intent architecture construction method, intent updating method and display device

Abstract

The application discloses a layered intention architecture construction method, an intention updating method and a display device. When the layered intention architecture is constructed, the intention similarity of each two secondary intentions is calculated based on example information and slot information, the secondary intentions are clustered to form primary intentions, and the layered intention architecture is constructed. When the intention is updated, the feature score of each initial intention corresponding to the current round of dialogue is calculated, and the corresponding associated intention and the feature score thereof are obtained based on the layered intention architecture. Then, the initial intention probability is updated in combination with the feature score of the historical intention, the feature score of the historical associated intention, the feature score of the historical user behavior and the feature score of the historical system behavior. It can be seen that, by constructing the layered intention architecture and fully considering the dialogue history and the NLU result under different intentions, the method and the display device can more accurately update the dialogue intention across intentions, accurately identify the target intention, and feed back the accurate interaction result to the user.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voice interaction, and in particular to a layered intent architecture construction method, an intent updating method, and a display device. BACKGROUND

[0002] With the rapid development of display devices, the functions of display devices will be increasingly rich, and the performance will be increasingly powerful. Currently, display devices include smart televisions, smart phones, refrigerators, wearable devices, vehicle navigation systems, and other products with intelligent display screens. To facilitate the free interaction of users with the above display devices, a human-computer dialogue system can be used to achieve flexible and fast human-computer voice interaction between users and display devices.

[0003] In the process of human-computer voice interaction, a user outputs a sentence, a human-computer dialogue system identifies the user's intent after semantic understanding, and then feeds back a corresponding reply to the user according to the intent. The user can ask follow-up questions in a short period of time through multiple rounds of interaction to obtain an accurate interaction result.

[0004] However, in this process, the user's multiple voice interaction requests can jump between different intents under the same theme, for example, the user searched for a game application in the previous round, and can search for a game video in the current round. If a reply is only given to the user according to the current round of voice interaction sentence, the current dialogue intent of the user can not be effectively and accurately identified, resulting in inaccurate interaction results of the human-computer dialogue system. SUMMARY

[0005] The present application provides a layered intent architecture construction method, an intent updating method, and a display device to solve the problem of inaccurate interaction results caused by the inability to effectively identify dialogue intents in multiple rounds of voice interaction.

[0006] In a first aspect, the present application provides a display device, comprising:

[0007] a display configured to present a user interface;

[0008] a controller connected to the display, the controller being configured to:

[0009] obtain an intent in a preset intent architecture as a secondary intent, and obtain example information and slot information of the secondary intent;

[0010] calculate an intent similarity between every two secondary intents based on the example information and the slot information;

[0011] cluster the secondary intents whose intent similarity meets a preset similarity condition to form a corresponding primary intent;

[0012] construct a hierarchical intent architecture based on the primary intent and the corresponding at least one secondary intent, the hierarchical intent architecture being used to provide associated intents similar to a specified secondary intent, probability information and slot information of each secondary intent corresponding to a specified primary intent in the hierarchical intent architecture being shared with each other, each of the secondary intents being affiliated to different primary intents.

[0013] In some embodiments of the present application, the controller, in performing the calculating of the intent similarity between each two secondary intents based on the example information and the slot information, is further configured to:

[0014] calculating a representation vector of a corresponding secondary intent based on the example information and the slot information;

[0015] calculating the intent similarity between each two secondary intents based on the representation vector of each secondary intent.

[0016] In some embodiments of the present application, the controller, in performing the calculating of the representation vector of a corresponding secondary intent based on the example information and the slot information, is further configured to:

[0017] calculating a sentence vector of a specified number of the example information corresponding to a specified secondary intent, calculating an average value of the specified number of the sentence vectors as the representation vector of the specified secondary intent based on the example information;

[0018] calculating a slot vector of a specified number of the slot information corresponding to the specified secondary intent, calculating an average value of the specified number of the slot vectors as the representation vector of the specified secondary intent based on the slot information;

[0019] concatenating the representation vector of the specified secondary intent based on the example information and the representation vector of the specified secondary intent based on the slot information to obtain the representation vector of the specified secondary intent.

[0020] In some embodiments of the present application, the controller, in performing the clustering of the secondary intents whose intent similarity meets a preset similarity condition to form a corresponding primary intent, is further configured to:

[0021] obtaining an intent similarity corresponding to a specified secondary intent, and selecting a secondary intent whose intent similarity exceeds a preset similarity threshold as an associated intent of the specified secondary intent;

[0022] determining whether each two of the associated intents belong to the same type, deleting the associated intents that do not belong to the same type, and clustering the associated intents after the deletion with the specified secondary intent to form a corresponding primary intent.

[0023] In a second aspect, the present application further provides a display device, comprising:

[0024] a display configured to present a user interface;

[0025] a controller connected with the display, the controller being configured to:

[0026] obtain an interaction sentence corresponding to a current round of dialogue in a multi-round dialogue, identify an initial intent corresponding to the interaction sentence, and calculate a feature score of the initial intent;

[0027] input the initial intent into a hierarchical intent architecture, obtain an associated intent similar to the initial intent, and calculate a feature score of the associated intent;

[0028] obtain a feature score of a historical intent, a feature score of a historical associated intent, a feature score of a historical user behavior, and a feature score of a historical system behavior corresponding to a historical round of dialogue in the multi-round dialogue;

[0029] based on the feature score of the initial intent, the feature score of the associated intent, the feature score of the historical intent, the feature score of the historical associated intent, the feature score of the historical user behavior, and the feature score of the historical system behavior, calculate an updated probability of each initial intent, the updated probability being used to determine a target intent corresponding to the current round of dialogue.

[0030] In some embodiments of the present application, the controller, in performing the calculation of the feature score of the initial intent, is further configured to:

[0031] obtain an initial probability of each initial intent, and sort each initial intent in descending order according to the initial probability;

[0032] based on a ranking number of each initial intent, calculate a ranking score of each initial intent;

[0033] based on the initial probability of the first initial intent, calculate a difference value between the initial probability of each subsequent initial intent and the initial probability of the first initial intent, and use the initial probability, the ranking score, and the difference value as the feature score of the corresponding initial intent.

[0034] In some embodiments of the present application, the controller, in performing the obtaining of the feature score of the historical intent corresponding to the historical round of dialogue, is further configured to:

[0035] obtain a frequency of occurrence of each historical intent corresponding to the historical round of dialogue in the multi-round dialogue, and calculate a frequency of occurrence of each historical intent;

[0036] obtain an intent probability of each historical intent, calculate a total value, an average value, and a maximum value of the intent probabilities of all historical intents corresponding to the historical round of dialogue;

[0037] calculate a ranking score of each of the historical intents, and calculate a total value, an average value and a maximum value of the ranking scores of all the historical intents corresponding to the historical round of dialogues;

[0038] take the occurrence frequency, the occurrence frequency, the total value, the average value and the maximum value of the intent probability, the total value, the average value and the maximum value of the ranking score, and the maximum value of the occurrence frequency of each of the historical intents as the feature score of the historical intents corresponding to the historical round of dialogues in the multi-round dialogue.

[0039] In some embodiments of the present application, the controller, when performing the obtaining of the feature score of the historical user behavior corresponding to the historical round of dialogues in the multi-round dialogue, is further configured to:

[0040] obtain the occurrence frequency of each of the historical user behaviors corresponding to the historical round of dialogues in the multi-round dialogue;

[0041] based on the occurrence frequency, calculate the occurrence frequency of each of the historical user behaviors, and take the occurrence frequency and the occurrence frequency of each of the historical user behaviors as the feature score of the historical user behavior corresponding to the historical round of dialogues in the multi-round dialogue.

[0042] In some embodiments of the present application, the controller, when performing the obtaining of the feature score of the historical system behavior corresponding to the historical round of dialogues in the multi-round dialogue, is further configured to:

[0043] obtain the occurrence frequency of each of the historical system behaviors corresponding to the historical round of dialogues in the multi-round dialogue;

[0044] based on the occurrence frequency, calculate the occurrence frequency of each of the historical system behaviors, and take the occurrence frequency and the occurrence frequency of each of the historical system behaviors as the feature score of the historical system behavior corresponding to the historical round of dialogues in the multi-round dialogue.

[0045] In a third aspect, the present application also provides a hierarchical intent architecture construction method, which comprises:

[0046] obtaining an intent in a preset intent architecture as a secondary intent, and obtaining example information and slot information of the secondary intent;

[0047] based on the example information and the slot information, calculating the intent similarity between each two of the secondary intents;

[0048] clustering the secondary intents whose intent similarity meets a preset similarity condition to form a corresponding primary intent;

[0049] construct a hierarchical intent architecture based on the primary intent and the corresponding at least one secondary intent, the hierarchical intent architecture being used to provide associated intents similar to the specified secondary intent, probability information and slot information of each secondary intent corresponding to the specified primary intent in the hierarchical intent architecture being shared with each other, each of the secondary intents being affiliated to different primary intents.

[0050] In a fourth aspect, the present application further provides an intent updating method based on the hierarchical intent architecture, the method comprising:

[0051] obtaining an interactive sentence corresponding to a current round of conversation in a multi-round conversation, identifying an initial intent corresponding to the interactive sentence, and calculating a feature score of the initial intent;

[0052] inputting the initial intent into the hierarchical intent architecture, obtaining an associated intent similar to the initial intent, and calculating a feature score of the associated intent;

[0053] obtaining a feature score of a historical intent, a feature score of a historical associated intent, a feature score of a historical user behavior, and a feature score of a historical system behavior corresponding to a historical round of conversation in the multi-round conversation;

[0054] calculating an updated probability of each of the initial intents based on the feature score of the initial intent, the feature score of the associated intent, the feature score of the historical intent, the feature score of the historical associated intent, the feature score of the historical user behavior, and the feature score of the historical system behavior, the updated probability being used to determine a target intent corresponding to the current round of conversation.

[0055] In a fifth aspect, the present application further provides a storage medium, which can store a program, the program being executable to implement some or all steps in each embodiment of the hierarchical intent architecture construction method and the intent updating method based on the hierarchical intent architecture provided by the present application.

[0056] From the above technical solutions, the layered intent architecture construction method, the intent updating method and the display device provided by the embodiments of the present application can obtain secondary intents in a preset intent architecture when constructing a layered intent architecture, calculate the intent similarity between each two secondary intents based on example information and slot information, cluster the secondary intents to form corresponding primary intents, and construct the layered intent architecture. When updating the intent, the feature scores of each initial intent corresponding to the current round of dialogue are calculated, and the corresponding associated intent in the layered intent architecture is obtained, and the feature score of the associated intent is calculated. Then, the feature scores of the historical intents, the feature scores of the historical associated intents, the feature scores of the historical user behaviors and the feature scores of the historical system behaviors corresponding to the historical round of dialogue in the multi-round dialogue are combined to update the initial intent probability, and the updated probability of each initial intent is calculated. It can be seen that the method and the display device describe the relationship between different intents by statistical method to construct the layered intent architecture. When updating the intent of the current round of voice interaction, the layered intent architecture is used and the dialogue history and the NLU result under different intents are fully considered, and then the dialogue intent updating across intents is more accurately performed to accurately identify the target intent and feed back the accurate interaction result to the user. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0058] Figure 1 A schematic diagram of an operating scenario between an intelligent control device and a control device according to some embodiments is shown;

[0059] Figure 2 A hardware configuration block diagram of the intelligent control device 200 according to some embodiments is shown;

[0060] Figure 3 A hardware configuration block diagram of the control device 100 according to some embodiments is shown;

[0061] Figure 4 A software configuration diagram in the intelligent control device 200 according to some embodiments is shown;

[0062] Figure 5 A flowchart of the layered intent architecture construction method according to some embodiments is shown;

[0063] Figure 6 A schematic diagram of the layered intent architecture according to some embodiments is shown;

[0064] Figure 7A flowchart illustrating an intent update method based on a hierarchical intent architecture according to some embodiments is shown.

[0065] Figure 8 A schematic diagram illustrating intent probability update according to some embodiments is shown. DETAILED DESCRIPTION

[0066] For the purpose of clarity and enabling a person of ordinary skill in the art to better understand and implement the present application, the present application will be described in detail by reference to the accompanying drawings, in which:

[0067] It should be noted that the brief description of the terms in the present application is only for the convenience of understanding the following described embodiments, and is not intended to limit the embodiments of the present application. Unless otherwise specified, these terms should be understood according to their ordinary and general meanings.

[0068] The terms "first", "second", "third", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar or identical objects or entities, and do not necessarily mean a specific order or sequence, unless otherwise noted. It should be understood that the terms used in this way can be interchanged under appropriate circumstances.

[0069] The terms "include" and "have" and any variations thereof are intended to cover but not exclusive inclusion, for example, a product or device including a series of components does not necessarily limit to all components clearly listed, but can include other components not clearly listed or inherent to these products or devices.

[0070] The term "module" refers to any known or later developed hardware, software, firmware, artificial intelligence, fuzzy logic, or combination of hardware or / and software code capable of performing a function associated with the element.

[0071] When the intelligent control device is a display device, the structure of the display device is introduced.

[0072] Figure 1 A use scenario of the intelligent control device according to some embodiments is shown. As shown, the intelligent control device 200 also communicates data with the server 400, and the user can operate the intelligent control device 200 through the smart device 300 or the control device 100. Figure 1

[0073] ​In some embodiments, the control device 100 can be a remote controller, and the communication between the remote controller and the smart control device includes at least one of infrared protocol communication or Bluetooth protocol communication, and other short-distance communication modes, to control the smart control device 200 in a wireless or wired manner. The user can input user instructions through at least one of a button on the remote controller, voice input, control panel input, etc., to control the smart control device 200.

[0074] In some embodiments, the smart device 300 can include any one of a mobile terminal, a tablet computer, a computer, a notebook computer, an AR / VR device, etc.

[0075] In some embodiments, the smart device 300 can also be used to control the smart control device 200. For example, the smart control device 200 is controlled using an application running on the smart device. In some embodiments, the smart device 300 and the smart control device can also be used for data communication.

[0076] In some embodiments, the smart control device 200 can also be controlled in ways other than the control device 100 and the smart device 300, for example, the user's voice instructions can be directly received by a voice instruction acquisition module configured inside the smart control device 200, or the user's voice instructions can be received by a voice control device configured outside the smart control device 200.

[0077] In some embodiments, the smart control device 200 also communicates data with the server 400. The smart control device 200 can be allowed to communicate through a local area network (LAN), a wireless local area network (WLAN), and other networks. The server 400 can provide various content and interactions to the smart control device 200. The server 400 can be a cluster or multiple clusters, and can include one or more types of servers.

[0078] In some embodiments, a software step performed by one step execution subject can be migrated to another step execution subject in data communication therewith as needed. For example, a software step performed by a server can be migrated to a smart control device in data communication therewith as needed, and vice versa.

[0079] Figure 2 A hardware configuration block diagram of the control device 100 according to some embodiments is shown. As shown, the control device 100 includes a controller 110, a communication interface 130, a user input / output interface 140, a memory, a power supply. The control device 100 can receive user input operation instructions, and convert the operation instructions into instructions that the smart control device 200 can recognize and respond to, acting as an intermediary between the user and the smart control device 200 in interaction. Figure 2 ​

[0080] In some embodiments, the communication interface 130 is configured to communicate with external devices, including at least one of a WIFI chip, a Bluetooth module, an NFC module, or an alternative module. In some embodiments, the user input / output interface 140 includes at least one of a microphone, a touchpad, a sensor, a button, or an alternative module.

[0081] Figure 3 A hardware configuration block diagram of the smart control device 200 is shown according to some embodiments. Referring to FIG. 2, the smart control device 200 includes a tuner and demodulator 210, a communicator 220, a detector 230, an external device interface 240, a controller 250, a display 260, an audio output interface 270, a memory, a power supply, at least one of a user interface, and the like. Figure 3 In some embodiments, the smart control device 200 includes at least one of a tuner and demodulator 210, a communicator 220, a detector 230, an external device interface 240, a controller 250, a display 260, an audio output interface 270, a memory, a power supply, a user interface, and the like. In some embodiments, the controller includes a central processing unit, a video processor, an audio processor, a graphics processor, a RAM, a ROM, a first interface to an n-th interface for input / output.

[0082] In some embodiments, the display 260 includes a display screen component for presenting a picture, and a driving component for driving the image display, a component for receiving an image signal from the controller output, and displaying video content, image content, and a menu control interface, and a user control UI interface, and the like. In some embodiments, the display 260 can be at least one of a liquid crystal display, an OLED display, and a projection display, and can also be a projection device and a projection screen.

[0083] In some embodiments, the controller 250 and the tuner and demodulator 210 can be located in different split devices, i.e., the tuner and demodulator 210 can also be in an external device of the main body device where the controller 250 is located, such as an external set-top box, and the like. In some embodiments, the controller 250 controls the operation of the smart control device and responds to user operations by storing various software control programs on the memory. The controller 250 controls the overall operation of the smart control device 200. For example, in response to receiving a user command for selecting a UI object to be displayed on the display 260, the controller 250 can perform an operation related to the object selected by the user command.

[0084] In some embodiments, the object can be any one of a selectable object, such as a hyperlink, an icon, or other operable control. The operation related to the selected object can be an operation of displaying a page connected to a hyperlink, a document, an image, or the like, or an operation of executing a program corresponding to the icon.

[0085] In some embodiments, the controller includes at least one of a Central Processing Unit (CPU), a video processor, an audio processor, a Graphics Processing Unit (GPU), a RAM (Random Access Memory), a ROM (Read-Only Memory), a first interface to an n-th interface for input / output, a communication bus, and the like.

[0086] In some embodiments, the user can input a user command through a Graphical User Interface (GUI) displayed on the display 260, and the user input interface receives the user input command through the Graphical User Interface (GUI). Alternatively, the user can input a user command by inputting a specific sound or gesture, and the user input interface receives the user input command by recognizing the sound or gesture through a sensor.

[0087] In some embodiments, the "user interface" is a medium interface for interaction and information exchange between an application or an operating system and a user, which realizes the conversion between the internal form of information and the form that the user can accept. The common form of the user interface is a Graphic User Interface (GUI), which refers to a user interface related to computer operation displayed in a graphical manner. It can be an icon, a window, a control element, and the like interface elements displayed in the display screen of an electronic device, wherein the control element can include at least one of an icon, a button, a menu, a tab, a text box, a dialog box, a status bar, a navigation bar, a Widget, and the like visible interface elements. In some embodiments, the user interface 280 is an interface (such as a physical button on the body of the smart control device, or the like) that can be used to receive control input.

[0088] Figure 4 A software configuration diagram in the smart control device 200 according to some embodiments is shown. Referring to Figure 4 In some embodiments, the system is divided into four layers from top to bottom, namely an Applications layer (referred to as "application layer" for short), an Application Framework layer (referred to as "framework layer" for short), an Android runtime and system library layer (referred to as "system runtime library layer" for short), and a kernel layer.

[0089] In some embodiments, at least one application program runs in the application layer, which can be a Window program, a system setting program, a clock program, etc. provided by the operating system, or an application program developed by a third party developer. In a specific implementation, the application programs in the application layer are not limited to the above examples.

[0090] The framework layer provides an application programming interface (API) and a programming framework for the application programs. The application framework layer includes some predefined functions. The application framework layer is equivalent to a processing center that decides which application program in the application layer to act. The application program can access the resources in the system and obtain the services of the system through the API interface in the execution.

[0091] As shown in Figure 4 In the embodiment of the application, the application framework layer includes managers, content providers, network management systems, etc., wherein the managers include at least one of the following modules: an activity manager for interacting with all activities running in the system; a location manager for providing access to the system location service for the system service or the application; a package manager for retrieving various information related to the application package currently installed on the device; a notification manager for controlling the display and clearing of the notification message; and a window manager for managing the icons, windows, toolbars, wallpapers and desktop components on the user interface.

[0092] In some embodiments, when the user interacts with the display device, in order to facilitate the user to freely talk with the display device to obtain a natural and intelligent interaction mode, the human-computer voice interaction system can be used to realize flexible and fast human-computer voice interaction between the user and the display device.

[0093] In addition to the voice recognition and text-to-speech modules, the human-computer voice interaction system also needs the following four key capabilities: natural language understanding (NLU), dialogue state tracking, dialogue policy learning, and natural language generation. Among them, the natural language understanding includes an analysis module of intent recognition and slot extraction and a data module such as question and answer retrieval; the dialogue state tracking takes the result of NLU as input and outputs the dialogue state containing the dialogue intent; the dialogue policy learning selects the corresponding dialogue action based on the obtained dialogue state; finally, the natural language generation converts the dialogue action into natural language to realize the interaction with the user.

[0094] In the process of human-computer voice interaction, the user outputs a sentence, the human-computer dialogue system identifies the user's intention after semantic understanding, and then feeds back the corresponding reply to the user according to the intention. The user can ask questions through multiple rounds of interaction in a short time, and the human-computer voice interaction system will feed back the reply to the user according to the user's output sentence in the current round to obtain an accurate interaction result.

[0095] However, in this process, the user's multiple voice interaction requests may jump between different intentions under the same theme, for example, the user searched for a game application in the last round, and may search for a game video in the current round. If the reply is only given to the user according to the current round of voice interaction sentence, the current dialogue intention of the user may not be effectively and accurately identified, resulting in inaccurate reply interaction results of the human-computer dialogue system.

[0096] To improve the ability of the human-computer voice interaction system to feed back more accurate interaction results in multiple rounds of voice interaction, the intention probability identified in the current round needs to be updated using the data generated in the historical rounds, for example, the probability of the game video intention identified in the current round needs to be updated based on the game application intention in the last round. Therefore, the display device provided in the present application can use a hierarchical dialogue intention system to process cross-intention intention understanding.

[0097] Based on this, the display device provided in the embodiments of the present application describes the relationship between different intentions by a statistical method, and constructs a hierarchical dialogue intention system to provide technical support for cross-intention dialogue intention updating. At the same time, when updating the intention probability in the current round, the hierarchical dialogue intention system constructed in the foregoing is used, and the upstream high-quality NLU result under different intentions is fully considered, so that the cross-intention dialogue intention updating is more accurate.

[0098] Figure 5 A flowchart of a hierarchical intention architecture construction method according to some embodiments is shown. The display device provided in the embodiments of the present application includes a display configured to present a user interface, and a controller connected to the display and configured to execute Figure 5 When the hierarchical intention architecture construction method shown in the figure is executed, the controller is configured to perform the following steps:

[0099] S11, obtain an intention in a preset intention architecture as a secondary intention, and obtain example information and slot information of the secondary intention.

[0100] In the construction of the hierarchical dialogue intention system, the relationship between different intentions is represented by a statistical method based on the current existing intention system, and the hierarchical dialogue intention system, i.e., the hierarchical intention architecture, is constructed based on this relationship, so as to improve the cross-intention understanding ability of the human-computer voice interaction system in the dialogue interaction process.

[0101] The preset intent architecture is a current existing domain intent system, which includes multiple intents. The intents in the preset intent architecture do not have a correlation relationship and exist independently.

[0102] To facilitate the establishment of a hierarchical dialogue intent system, the intents in the preset intent architecture need to be processed, and intents with similar characteristics are classified into a category, so that the independent intents in the preset intent architecture are clustered according to similar characteristics to form a hierarchical structure feature. Therefore, the intents in the preset intent architecture are taken as secondary intents, and the secondary intents are clustered to obtain primary intents.

[0103] The preset intent architecture includes at least 300 intents, such as movie search, music search, and animation search. Each intent has corresponding text and supported slots, such as the text "I want to watch the movie / TV series of Andy Lau this year" under the movie search, and the slots include actor, videoType, and year.

[0104] The text is example information, which is used to represent the example statement of the user's speech when the user has a voice dialogue with the display device. The slot is a label, which is used for intent understanding of the user's speech in the human-computer voice interaction system. A secondary intent corresponds to multiple example information, and an example information includes multiple slot information.

[0105] S12, based on the example information and the slot information, calculating the intent similarity between each two secondary intents.

[0106] Different example information may have the same slot information, so the intent similarity between each two secondary intents can be calculated based on the example information and the slot information. The intent similarity is used to represent the correlation or similarity between each two intents, to determine whether the two intents are the same class of intent, that is, whether they can be classified into the same class.

[0107] In some embodiments, the controller, in performing the step of calculating the intent similarity between each two secondary intents based on the example information and the slot information, is further configured to perform the following steps:

[0108] Step 121, based on the example information and the slot information, calculating the representation vector of the corresponding secondary intent.

[0109] Step 122, based on the representation vector of each secondary intent, calculating the intent similarity between each two secondary intents.

[0110] The representation vector method can be used when calculating the intent similarity, in which case the example information and the slot information corresponding to each secondary intent need to be represented by vectors respectively.

[0111] In some embodiments, the controller, in performing the calculation of the representation vector of the corresponding secondary intent based on the example information and the slot information, is further configured to perform the following steps:

[0112] Step 1211, calculating the sentence vectors of the number of example information corresponding to the specified secondary intent, calculating the average of the number of sentence vectors as the representation vector of the specified secondary intent based on the example information.

[0113] Step 1212, calculating the slot vectors of the number of slot information corresponding to the specified secondary intent, calculating the average of the number of slot vectors as the representation vector of the specified secondary intent based on the slot information.

[0114] Step 1213, splicing the representation vector of the specified secondary intent based on the example information and the representation vector of the specified secondary intent based on the slot information to obtain the representation vector of the specified secondary intent.

[0115] Each secondary intent under the preset intent architecture needs to calculate the vector of each example information and the vector of each slot information corresponding to the secondary intent to finally determine the representation vector of the corresponding secondary intent. That is, the calculation method of the representation vector of each secondary intent is the same, and here, only the calculation of the representation vector of the specified secondary intent is taken as an example.

[0116] In calculating the representation vector of the specified secondary intent, the sentence vectors of the number of example information corresponding to the specified secondary intent and the slot vectors of the number of slot information are calculated first. Since the sentence vector of each example information is different, and the slot vector of each slot information is also different. Therefore, in order to accurately calculate the representation vector of the specified secondary intent, the average of the sentence vectors of all example information under the intent is calculated as the representation vector of the specified secondary intent based on the example information; and the average of all slot vectors under the intent is calculated as the representation vector of the specified secondary intent based on the slot information.

[0117] Finally, since the specified secondary intent can be represented by example information and slot information, the representation vector of the example information and the representation vector of the specified secondary intent based on the slot information are spliced to obtain the representation vector of the specified secondary intent.

[0118] In some embodiments, in calculating the sentence vectors of the number of example information corresponding to the specified secondary intent, the Bert pre-training model is called, and each example information is input into the Bert pre-training model to obtain the corresponding sentence vector. For example, if the specified secondary intent corresponds to five example information, the five example information is input into the Bert pre-training model to obtain five corresponding sentence vectors. The average of the five sentence vectors can be calculated to obtain the representation vector of the specified secondary intent based on the example information.

[0119] In some embodiments, when calculating the slot vector corresponding to the specified secondary intent, a classification model subnetwork (tagEmbedding, which has been trained) is called, and each slot information is respectively input into the classification model subnetwork to obtain a corresponding slot vector. For example, if the specified secondary intent corresponds to five slot information, the five slot information is respectively input into the classification model subnetwork to obtain five corresponding slot vectors. The average of the five slot vectors is calculated to obtain the slot information-based representation vector of the specified secondary intent.

[0120] In some embodiments, if the representation vector of the specified secondary intent based on the example information is a 100-dimensional vector, and the representation vector of the specified secondary intent based on the slot information is a 200-dimensional vector, then the two are spliced to determine that the representation vector of the specified secondary intent is a 300-dimensional vector.

[0121] In some embodiments, each secondary intent in the preset intent architecture is calculated according to the above method to obtain a representation vector. If the preset intent architecture includes 300 secondary intents, 300 representation vectors can be generated. Then, based on the representation vector of each secondary intent, the intent similarity between each two secondary intents is calculated.

[0122] The representation vector of any secondary intent is calculated to obtain the intent similarity with the representation vectors of other secondary intents. The method for calculating the intent similarity can be: the similarity sim(a, b) of vectors a and b = cos(a, b) = (a, b) / ||a|| ||b||. (a, b) is the inner product of vectors a and b, ||a|| is the 2-norm of a, and ||b|| is the 2-norm of b.

[0123] When calculating the intent similarity between the first secondary intent and the other 299 secondary intents, a is the representation vector of the first secondary intent, b is the representation vector of the i-th secondary intent, and i is any one of 2 to 300. At this time, the number of intent similarities corresponding to the first secondary intent is 299.

[0124] Similarly, when calculating the intent similarity between the j-th secondary intent and the other 299 secondary intents, a is the representation vector of the j-th secondary intent, b is the representation vector of the i-th secondary intent, and i is any one of 1 to 300, but not including the j-th. At this time, the number of intent similarities corresponding to the j-th secondary intent is 299.

[0125] S13, clustering the secondary intents with intent similarity satisfying a preset similarity condition to form a corresponding primary intent.

[0126] In order to realize the clustering of the secondary intent, the secondary intents satisfying the preset similarity condition can be divided into a class, denoted as a primary intent.

[0127] In some embodiments, the controller, in performing the clustering of the secondary intents whose intent similarity satisfies the preset similarity condition to form the corresponding primary intent, is further configured to perform the following steps:

[0128] In step 131, the intent similarity corresponding to the specified secondary intent is obtained, and the secondary intent whose intent similarity exceeds the preset similarity threshold is selected as the associated intent of the specified secondary intent.

[0129] In step 132, it is determined whether each two associated intents belong to the same type, and the associated intents not belonging to the same type are deleted. The clustering of the specified secondary intent and the associated intents after deletion is performed to form the corresponding primary intent.

[0130] When the number of all secondary intents is 300, the number of intent similarity corresponding to the specified secondary intent is 299. When the preset similarity condition is set as the preset similarity threshold, for the specified secondary intent, the secondary intent whose corresponding intent similarity exceeds the preset similarity threshold is selected as the associated intent of the specified secondary intent.

[0131] When the preset similarity threshold is set as 6, the top 6 secondary intents of the 299 intent similarities are output as the associated intents of the specified secondary intent. The associated intent is the secondary intent having the association or similarity with the specified secondary intent, and it can be determined that the two intents are the same type of intent, that is, they can be classified into the same class of intent.

[0132] The associated intent is the secondary intent having the association or similarity with the specified secondary intent, and it can be divided into a class. However, if the number of example statements under a certain secondary intent is insufficient or the representation calculation method is problematic, the determination of the associated intent will not be accurate enough, and therefore, it is further necessary to determine whether the associated intent and the specified secondary intent belong to the same type, that is, whether each intent pair has a near-neighbor relationship. For example, although game search and media search have high similarity, game search belongs to the game field type and media search belongs to the media field type, and they do not have a near-neighbor relationship.

[0133] Therefore, in order to accurately cluster and divide each secondary intent, artificial cleaning can be used, that is, the expert knowledge is combined to determine whether each intent pair has a near-neighbor relationship, and the intent pairs not having a near-neighbor relationship are removed. The intent pair refers to the data group formed by any two intents in each associated intent and the specified secondary intent, including the intent pair of the associated intent and the associated intent, and the intent pair of the associated intent and the specified secondary intent.

[0134] The association intents not belonging to the same type are deleted, and the association intents after deletion are clustered based on the specified secondary intent to form corresponding primary intents. For example, when the specified secondary intent is a game video, if the association intents similar to the game video are cartoon video, movie video, and game search. However, from the type perspective, the cartoon video and the movie video are both in the media field type as the game video, and have a near-neighbor relationship, and can be classified into one category; and the game search is in the game field type, and obviously has a different type and does not have a near-neighbor relationship. Therefore, the game search needs to be deleted, and the remaining cartoon video, movie video, and game video are classified into the same category to form a primary intent.

[0135] The intents having a near-neighbor relationship between each other are clustered together to form a primary intent, and the primary intent is defined. The name of the primary intent can also be defined based on expert knowledge in combination with each secondary intent under the primary intent. For example, the primary intent corresponding to the cartoon video, movie video, and game video can be video, and the primary intent corresponding to the game video and game search can be game.

[0136] S14, constructing a hierarchical intent architecture based on the primary intent and the corresponding at least one secondary intent, the hierarchical intent architecture is used to provide the association intents similar to the specified secondary intent, the probability information and the slot information of each secondary intent corresponding to the specified primary intent in the hierarchical intent architecture are shared with each other, and each secondary intent can belong to different primary intents.

[0137] Based on each secondary intent in the foregoing preset intent architecture and the primary intent formed after clustering, the hierarchical intent architecture, that is, the hierarchical dialogue intent system, can be constructed.

[0138] Figure 6 A schematic diagram of a hierarchical intent architecture according to some embodiments is shown. Referring to FIG. 6, Figure 6 In the preset intent architecture, the secondary intents are cartoon video, movie video, game video, game search, and the like. After clustering each secondary intent by calculating the intent similarity as described above, the cartoon video, movie video, and game video can be clustered into one category, and the primary intent is determined to be video; and the game video and game search can be clustered into one category, and the primary intent is determined to be game.

[0139] Based on the structural characteristics of the hierarchical intent architecture, the probability information and the slot information of all the secondary intents corresponding to the specified primary intent in the hierarchical intent architecture are shared with each other, and each secondary intent can belong to different primary intents. The hierarchical intent architecture is used to provide associated intents similar to the specified secondary intent. For example, if the intent of the current sentence of the user is identified, the intent is input into the hierarchical intent architecture, and the associated intents similar to the intent can be obtained. The actual intent of the user dialogue can be effectively identified through the associated intent, the calculation accuracy of the probability of each intent can be improved, and then the man-machine voice interaction system can accurately feed back the interaction result to the user based on the selected target intent.

[0140] Therefore, the display device provided by the embodiment of the present application obtains the secondary intents in the preset intent architecture when constructing the hierarchical intent architecture, calculates the intent similarity between each two secondary intents based on the example information and the slot information, clusters the secondary intents whose intent similarity meets the preset similarity condition to form the corresponding primary intent, and constructs the hierarchical intent architecture based on the primary intent and the corresponding at least one secondary intent. It can be seen that the display device describes the relationship between different intents through a statistical method, provides the associated intents similar to the specified secondary intent through the constructed hierarchical intent architecture, updates the intent of the current round of voice interaction, accurately identifies the target intent, and feeds back the accurate interaction result to the user.

[0141] After the construction of the hierarchical intent architecture is completed based on the foregoing method, when the multi-round voice interaction is performed, the hierarchical intent architecture can be used to fully consider the upstream high-quality NLU result (i.e., the NLU result of the previous round of interaction) under different intents, and then the current round of cross-intent dialogue intent updating is more accurately performed to accurately identify the target intent and feed back the accurate interaction result to the user.

[0142] Figure 7 A flowchart of an intent updating method based on a hierarchical intent architecture according to some embodiments is shown. The embodiment of the present application provides a display device, which includes a display configured to present a user interface, and a controller connected with the display and configured to perform the following steps when the intent updating method based on the hierarchical intent architecture is performed as shown in Figure 7 The controller is configured to perform the following steps when the intent updating method based on the hierarchical intent architecture is performed as shown in

[0143] S21, obtaining an interaction sentence corresponding to a current round of dialogue in a multi-round dialogue, identifying an initial intent corresponding to the interaction sentence, and calculating a feature score of the initial intent.

[0144] When the user performs man-machine voice interaction with the display device, the user can only perform one round of interaction, or can perform multiple rounds of interaction in a short period of time. In the multi-round dialogue, the user can ask follow-up questions to the feedback made by the display device to form the dialogue of the next round.

[0145] Therefore, in identifying the target intent of the user output sentence in the current round of dialogue, in order to improve the recognition accuracy, the dialogue history and the NLU result can be explicitly represented based on the recognition results of the previous rounds of dialogue, i.e., fully considering the dialogue history and the NLU result, and then the probability update of the current round of dialogue intent can be completed by using a machine learning method. Through the updated intent probability, the target intent of the user output sentence in the current round of dialogue can be identified, and then the human-computer voice interaction system can accurately feed back the interaction result.

[0146] To this end, in order to realize the intent probability update, the interactive sentence corresponding to the current round of dialogue is first obtained, and the human-computer voice interaction system identifies the interactive sentence to obtain the corresponding initial intent. Here, the initial intent is updated in probability, and the initial intent recognized usually can be multiple, so the target intent can be determined according to the updated probability.

[0147] In the probability update, the update process includes two parts of feature extraction and probability update. The feature extraction part includes the feature scores of the initial intents recognized in the current round of dialogue, the feature scores of the associated intents, the feature scores of the historical intents recognized in the previous rounds of dialogue, the feature scores of the historical associated intents, the feature scores of the historical user behaviors, and the feature scores of the historical system behaviors, etc.

[0148] In some embodiments, the controller, in performing the calculation of the feature scores of the initial intents, is further configured to perform the following steps:

[0149] Step 211, obtaining the initial probability of each initial intent, and sorting each initial intent in descending order of the initial probability.

[0150] Step 212, calculating the ranking score of each initial intent based on the ranking number of each initial intent.

[0151] Step 213, calculating the initial probability difference between each subsequent initial intent and the initial probability of the first initial intent based on the initial probability of the first initial intent, and taking the initial probability, the ranking score, and the initial probability difference as the feature score of the corresponding initial intent.

[0152] In calculating the feature scores of the initial intents, the recognition result of the sentence corresponding to the current round of dialogue by NLU can be obtained, and multiple initial intents corresponding to the user output sentence and the initial probability of each initial intent can be obtained. The initial probability is the confidence score obtained after recognizing the sentence in the current round, at this time, the NLU recognition result of the previous rounds is not considered.

[0153] The initial intents are ranked according to the initial probabilities in descending order, and each initial intent corresponds to a rank. The rank score of the corresponding initial intent is calculated using the rank number. The calculation formula of the rank score can be 1 / r or 1 / log2(r+1). The two calculation methods can be selected.

[0154] The initial probability of the initial intent with the rank number 1 is obtained, and the initial probability difference between the initial intent with the rank number 1 and the initial intent with the rank number 2 is calculated. For example, if there are four initial intents, the initial probability difference between the initial intent with the rank number 1 and the initial intent with the rank number 2 is calculated, the initial probability difference between the initial intent with the rank number 1 and the initial intent with the rank number 3 is calculated, and the initial probability difference between the initial intent with the rank number 1 and the initial intent with the rank number 4 is calculated.

[0155] After the above calculations are completed, the initial probability, the rank score, and the initial probability difference are used as the feature score of the corresponding initial intent. For a specified initial intent, the feature score includes the specified initial probability, the specified rank score, and the specified initial probability difference. The specified initial probability difference includes n-1 values, and n is the total number of initial intents.

[0156] In step S22, the initial intent is input into the hierarchical intent architecture, and the associated intent similar to the initial intent is obtained. The feature score of the associated intent is calculated.

[0157] Since the initial intent is identified in the current round, but it is not accurate to determine which initial intent is the final target intent. Therefore, the associated intent similar to the initial intent can be used to assist in determination. At this time, the hierarchical intent architecture constructed in the foregoing embodiment is obtained, and each initial intent is input into the hierarchical intent architecture. A plurality of associated intents corresponding to each initial intent can be obtained, that is, one initial intent corresponds to at least one associated intent.

[0158] Similarly, the feature score of each associated intent is calculated. The feature score of the associated intent includes the associated probability, the rank score, and the associated probability difference. The calculation method of the feature score of the associated intent is the same as that of the feature score of the initial intent, and the corresponding steps can refer to the contents of steps 211 to 213. For example, the associated probability is obtained in the same way as the initial probability, the calculation process of the rank score is the same, and the calculation process of the associated probability difference is the same as that of the initial probability difference. Here, no longer be described.

[0159] Based on the calculation results of steps S21 and S22, the feature score of each initial intent and the feature score of each associated intent are used as the NLU result corresponding to the current round of dialogue.

[0160] S23, obtain the feature scores of the historical intents, the feature scores of the historical related intents, the feature scores of the historical user behaviors, and the feature scores of the historical system behaviors corresponding to the historical round of dialogues in the multi-round dialogue.

[0161] If there is a multi-round dialogue, the feature scores of the historical intents, the feature scores of the historical related intents, the feature scores of the historical user behaviors, and the feature scores of the historical system behaviors corresponding to the historical round of dialogues before the current round of dialogue are obtained. When identifying the intent of the current round of dialogue, the dialogue history data is fully considered, which can improve the identification accuracy of the target intent.

[0162] The historical intent refers to an intent identified in a historical round of dialogue. One historical round of dialogue can correspond to multiple historical intents, and there can be multiple historical rounds of dialogue. The historical related intent refers to an intent identified in a historical round of dialogue and obtained by inputting the intent into a hierarchical intent structure. The historical user behavior refers to an operation of the user on the interactive result of the display device, including operations such as affirmation, neglect, and negation. The historical system behavior refers to the response type of the display device to the interactive request of the user, including behaviors such as informing, confirming, and selecting.

[0163] In some embodiments, the controller, when performing the step of obtaining the feature scores of the historical intents corresponding to the historical round of dialogues in the multi-round dialogue, is further configured to perform the following steps:

[0164] Step 2311, obtain the occurrence frequency of each historical intent corresponding to the historical round of dialogues in the multi-round dialogue, and calculate the occurrence frequency of each historical intent.

[0165] Step 2312, obtain the intent probability of each historical intent, calculate the total value, the average value, and the maximum value of the intent probability of all historical intents corresponding to the historical round of dialogues.

[0166] Step 2313, calculate the ranking score of each historical intent, and calculate the total value, the average value, and the maximum value of the ranking score of all historical intents corresponding to the historical round of dialogues.

[0167] Step 2314, take the occurrence frequency, the occurrence frequency, the total value, the average value, and the maximum value of the intent probability, the total value, the average value, and the maximum value of the ranking score, and the maximum value of the occurrence frequency of each historical intent as the feature scores of the historical intents corresponding to the historical round of dialogues in the multi-round dialogue.

[0168] If there is a t-turn conversation, the historical turn conversation is the previous t-1 turn conversation. The occurrence frequency of each historical intent corresponding to the historical turn (previous t-1) conversation is obtained, and the occurrence frequency refers to the total number (sum) of candidate intents in which the historical intent appears in the previous t-1 turn conversation. For example, if the current turn t is 7 turns, the previous t-1 turns refer to the previous 6 turns, including the data of the historical 1st to 6th turns.

[0169] Then, the occurrence frequency p of each historical intent is calculated i = sum / min(T, k-1), where T is the maximum number of historical turns obtained, k is the current turn number, the first turn does not obtain this part of the feature, and i is the serial number of the historical intent. The occurrence frequency of the historical intent refers to the frequency of the i-th historical intent appearing in the k-th turn of the previous t-1 turn conversation, that is, k is one of the t-1 turns.

[0170] The NLU result corresponding to the historical turn conversation, that is, the intent probability of each historical intent, is obtained. The intent probability of the historical intent is the confidence score obtained after identifying the statement of the historical turn. According to the intent probability of each historical intent, the total value, average value and maximum value of the intent probability of all historical intents corresponding to the historical turn conversation are calculated.

[0171] The total value of the intent probability (confidence_sum) refers to the sum of the intent probabilities of all historical intents recognized in all historical turns. The average value of the intent probability refers to the average value of the sum of the intent probabilities of all historical intents recognized in all historical turns, which is confidence_sum / sum, and sum is the total number of all historical intents. The maximum value of the intent probability refers to the maximum value of the intent probabilities of all historical intents recognized in all historical turns.

[0172] According to the order of the intent probability of each historical intent from large to small, each historical intent is sorted, and at this time, each historical intent corresponds to a ranking, and the ranking score (rank score) of each historical intent is calculated. The calculation formula of the ranking score can be 1 / r, or 1 / log2(r+1). The two calculation methods can be used alternatively, and r is the ranking serial number.

[0173] Based on the ranking score of each historical intent, the total value, average value and maximum value of the ranking score of all historical intents corresponding to the historical turn conversation are calculated. The total value of the ranking score (rank_sum) refers to the sum of the ranking scores of all historical intents recognized in all historical turns. The average value of the ranking score refers to the average value of the sum of the ranking scores of all historical intents recognized in all historical turns, which is rank_sum / sum, and sum is the total number of all historical intents. The maximum value of the ranking score refers to the maximum value of the ranking scores of all historical intents recognized in all historical turns.

[0174] To improve the probability updating accuracy of the initial intent, the maximum number of occurrences of the historical intent with the most occurrences can also be obtained, i.e., the maximum value of the occurrence frequency. Finally, the occurrence frequency, the occurrence frequency, the total value, the average value and the maximum value of the intent probability, the total value, the average value and the maximum value of the ranking score, and the maximum value of the occurrence frequency of each historical intent are taken as the feature scores of the historical intent corresponding to the historical round of dialogue in the multi-round dialogue.

[0175] In some embodiments, in the historical round of dialogue, each identified intent can also be input into the hierarchical intent architecture to obtain the corresponding associated intent, thereby ensuring the accuracy of intent recognition in that round of dialogue. Therefore, the feature scores of the historical associated intent corresponding to the historical round of dialogue in the multi-round dialogue can also be obtained.

[0176] The feature scores of the historical associated intent include the occurrence frequency, the occurrence frequency, the total value, the average value and the maximum value of the intent probability, the total value, the average value and the maximum value of the ranking score, and the maximum value of the occurrence frequency of each historical associated intent. The calculation method of the feature scores of the historical associated intent can refer to the calculation method of the feature scores of the historical intent, and the related process can correspondingly refer to the content of the foregoing steps 2311 to 2314, which will not be described here.

[0177] In some embodiments, the controller, in executing the feature scores of the historical user behavior corresponding to the historical round of dialogue in the multi-round dialogue, is further configured to execute the following steps:

[0178] Step 2321, obtaining the occurrence frequency of each historical user behavior corresponding to the historical round of dialogue in the multi-round dialogue.

[0179] Step 2322, based on the occurrence frequency, calculating the occurrence frequency of each historical user behavior, and taking the occurrence frequency and the occurrence frequency of each historical user behavior as the feature scores of the historical user behavior corresponding to the historical round of dialogue in the multi-round dialogue.

[0180] The historical user behavior refers to the behavior of the user operating the interactive result of the feedback of the display device, including types such as affirmation, ignore, and negation. Therefore, in the multi-round dialogue, the types and corresponding numbers of behaviors made by the user based on the feedback interactive result of the display device in the previous historical round of dialogue can be obtained.

[0181] The occurrence frequency of the historical user behavior refers to the total number (sum) of the occurrence of the historical user behavior in the previous t-1 rounds of dialogue, for example, the total number (sum) of the affirmative behavior of the user.

[0182] Then, the occurrence frequency p of each historical user behavior is calculated i= sum / min(T, k-1), where T is the maximum number of historical rounds obtained, k is the current round number, the first round does not obtain this part of the feature, and i is the sequence number of the historical user behavior. The occurrence frequency of the historical user behavior refers to the frequency of the ith historical user behavior in the kth round of the t-1 rounds of dialogues, that is, k is one of the t-1 rounds.

[0183] Finally, the occurrence frequency and the occurrence frequency of each historical user behavior are taken as the feature scores of the historical user behaviors corresponding to the historical round dialogues in the multi-round dialogue.

[0184] In some embodiments, the controller, in executing the step of obtaining the feature scores of the historical system behaviors corresponding to the historical round dialogues in the multi-round dialogue, is further configured to execute the following steps:

[0185] Step 2331, obtaining the occurrence frequency of each historical system behavior corresponding to the historical round dialogue in the multi-round dialogue.

[0186] Step 2332, based on the occurrence frequency, calculating the occurrence frequency of each historical system behavior, and taking the occurrence frequency and the occurrence frequency of each historical system behavior as the feature scores of the historical system behaviors corresponding to the historical round dialogues in the multi-round dialogue.

[0187] The historical system behavior refers to the type of response of the display device to the user's interactive request, including behaviors such as informing, confirming, selecting, etc. Therefore, in the multi-round dialogue, the types and numbers of response behaviors made by the display device based on the user's voice interactive request in the historical round dialogues of the previous rounds can be obtained.

[0188] The occurrence frequency of the historical system behavior refers to the total number (sum) of the historical system behavior appearing in the display device based on the user's request to make the corresponding behavior in the t-1 rounds of dialogues, for example, the total number (sum) of the confirmation behavior referring to the display device confirming the intent and making the interactive result feedback.

[0189] Then, the occurrence frequency p of each historical system behavior is calculated i = sum / min(T, k-1), where T is the maximum number of historical rounds obtained, k is the current round number, the first round does not obtain this part of the feature, and i is the sequence number of the historical user behavior. The occurrence frequency of the historical user behavior refers to the frequency of the ith historical user behavior in the kth round of the t-1 rounds of dialogues, that is, k is one of the t-1 rounds.

[0190] Finally, the occurrence frequency and the occurrence frequency of each historical system behavior are taken as the feature scores of the historical system behaviors corresponding to the historical round dialogues in the multi-round dialogue.

[0191] S24, based on the initial intention feature score, the associated intention feature score, the historical intention feature score, the historical associated intention feature score, the historical user behavior feature score and the historical system behavior feature score, the updated probability of each initial intention is calculated, and the updated probability is used to determine the target intention corresponding to the current round of dialogue.

[0192] Based on the content of the foregoing steps S21 to S23, the feature extraction part of the intention probability updating is completed, and then the probability updating part can be performed.

[0193] When the probability is updated, the ranking model or the DNN network can be called to score each initial intention based on the foregoing feature scores, i.e., to update the initial probability of each initial intention. That is, the feature scores of each initial intention, the feature scores of each associated intention, the feature scores of each historical intention, the feature scores of each historical associated intention, the feature scores of each historical user behavior and the feature scores of each historical system behavior are respectively input into the ranking model or the DNN network for scoring, and finally, the updated intention probability is output by the softmax, i.e., the updated probability of the initial intention, which is used to determine the target intention corresponding to the current round of dialogue.

[0194] Figure 8 A schematic diagram of intention probability updating is shown according to some embodiments. Referring to Figure 8 In the adjacent two rounds of dialogue, the user's intention can change. For example, in the previous round (t-1) of interaction, the user's interactive sentence is "I want to watch the light head strong", and the corresponding initial intention is recognized as cartoon video or video, and the initial probability of cartoon video is 0.4 and the initial probability of video is 0.5.

[0195] In the next round of interaction (the tth round), the user's interactive sentence is "the movie last year", and the initial probability of each initial intention can be updated based on the interactive data of the current round and the interactive data of the historical round. At this time, the updated initial probability of cartoon video is 0.1 and the initial probability of video is 0.9. Therefore, the video can be taken as the target intention of the user's multi-round dialogue. The human-computer voice interaction system can feed back the corresponding interactive result based on the target intention, improving the accuracy.

[0196] In some embodiments, based on the NUL result of the historical round dialogue and the NUL result of the current round dialogue, not only the intention probability can be updated, but also the slot completion can be implemented. In the previous round (t-1) interaction, the slot is only the role name (roleName) and the action (actionWatch); and in the next round of interaction, based on the current interaction sentence, the completed slot includes the role name (roleName), the action (actionWatch), the year (titleYear), the media type (videoKey), etc.

[0197] Therefore, the display device provided by the embodiment of the present application can calculate the feature scores of each initial intention corresponding to the current round dialogue when updating the intention based on the hierarchical intention architecture, input each initial intention into the hierarchical intention architecture to obtain the corresponding associated intention, and calculate the feature scores of the associated intention; then acquire the feature scores of the historical intention, the feature scores of the historical associated intention, the feature scores of the historical user behavior, and the feature scores of the historical system behavior corresponding to the historical round dialogue in the multi-round dialogue; update the probability based on the feature scores of the above types, calculate the updated probability of each initial intention, and determine the target intention corresponding to the current round dialogue. It can be seen that the display device is based on the hierarchical intention architecture constructed as described above, and fully considers the dialogue history and the NLU result under different intentions, so as to more accurately update the dialogue intention across intentions, accurately identify the target intention, and feed back the accurate interaction result to the user.

[0198] Figure 5 A flowchart of a hierarchical intention architecture construction method according to some embodiments is shown. Referring to Figure 5 The embodiment of the present application provides a hierarchical intention architecture construction method, which is executed by the controller of the display device provided by the foregoing embodiment, and the method comprises the following steps:

[0199] S11, acquiring an intention in a preset intention architecture as a secondary intention, and acquiring example information and slot information of the secondary intention;

[0200] S12, calculating the intention similarity between each two secondary intentions based on the example information and the slot information;

[0201] S13, clustering the secondary intentions whose intention similarity meets a preset similarity condition to form a corresponding primary intention;

[0202] S14, constructing a hierarchical intention architecture based on the primary intention and at least one corresponding secondary intention, the hierarchical intention architecture is used to provide an associated intention similar to a specified secondary intention, the probability information and the slot information of each secondary intention corresponding to a specified primary intention in the hierarchical intention architecture are shared with each other, and each secondary intention can belong to different primary intentions.

[0203] Figure 7 A flowchart of an intent updating method based on a hierarchical intent architecture according to some embodiments is shown. Referring to Figure 7 The embodiment of the present application provides a kind of intent updating method based on hierarchical intent architecture, the controller of display equipment provided by the foregoing embodiment is executed, the method comprises:

[0204] S21, the interactive sentence corresponding to the initial intent corresponding to the interactive sentence of the current round conversation in multi-round conversation is obtained, and the feature score of the initial intent is calculated;

[0205] S22, the initial intent is input into hierarchical intent architecture, and the associated intent similar to the initial intent is obtained, and the feature score of the associated intent is calculated;

[0206] S23, the feature score of the historical intent corresponding to the historical round conversation in multi-round conversation, the feature score of the historical associated intent, the feature score of the historical user behavior and the feature score of the historical system behavior are obtained;

[0207] S24, based on the feature score of the initial intent, the feature score of the associated intent, the feature score of the historical intent, the feature score of the historical associated intent, the feature score of the historical user behavior and the feature score of the historical system behavior, the updated probability of each initial intent is calculated, and the updated probability is used to determine the target intent corresponding to the current round conversation.

[0208] From the above technical solution, the hierarchical intent architecture construction method, the intent updating method and the display device provided by the embodiment of the present application are known, when constructing hierarchical intent architecture, the secondary intent in preset intent architecture is obtained, the intent similarity between every two secondary intents is calculated based on example information and slot information, the secondary intent is clustered to form corresponding primary intent, to construct hierarchical intent architecture. When updating intent, the feature score of each initial intent corresponding to the current round conversation is calculated, and the corresponding associated intent is obtained based on hierarchical intent architecture, and the feature score of the associated intent is calculated; Then, the feature score of the historical intent corresponding to the historical round conversation in multi-round conversation, the feature score of the historical associated intent, the feature score of the historical user behavior and the feature score of the historical system behavior are combined to update the initial intent probability, and the updated probability of each initial intent is calculated. It can be seen that the method and the display device describe the relationship between different intents by statistical method, construct hierarchical intent architecture; When updating the intent of current round voice interaction, based on hierarchical intent architecture and fully considering the conversation history and NLU result under different intents, the cross-intent conversation intent updating is more accurately carried out, to accurately identify target intent, and accurately feedback interactive result to user.

[0209] In a particular implementation, the application further provides a storage medium, wherein the storage medium can store a program, and the program can include some or all steps of the embodiments of the method for constructing a hierarchical intent architecture and the method for updating an intent based on the hierarchical intent architecture when executed. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0210] Those skilled in the art can clearly understand that the technology in the embodiments of the application can be realized by means of software and necessary general hardware platforms. Based on such understanding, the technical solutions in the embodiments of the application can be embodied in the form of a software product, which can be stored in a storage medium, such as a ROM / RAM, a disk, an optical disk, etc.

[0211] The same or similar parts among the various embodiments in the specification can be referred to each other. Especially, for the method for constructing a hierarchical intent architecture and the method for updating an intent based on the hierarchical intent architecture, since they are basically similar to the display device embodiments, the description is relatively simple, and the relevant parts can be referred to the description in the display device embodiments.

[0212] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application.

[0213] In order to facilitate explanation, the above description has been made in combination with specific embodiments. However, the above exemplary discussion is not intended to exhaust or limit the embodiments to the specific forms disclosed above. According to the above teachings, various modifications and variations can be obtained. The selection and description of the above embodiments are to better explain the principles and practical applications, so that those skilled in the art can better use the embodiments and various different modified embodiments suitable for specific use considerations.

Claims

1. A display device, characterized by comprising: The application comprises: a display configured to present a user interface; a controller connected with the display, the controller being configured to: obtain an intent in a preset intent schema as a secondary intent, and obtain example information and slot information of the secondary intent, wherein there is no correlation between each intent in the preset intent schema; calculate an intent similarity between each two secondary intents based on the example information and the slot information; obtain an intent similarity corresponding to a specified secondary intent, and select a secondary intent whose intent similarity exceeds a preset similarity threshold as a correlated intent of the specified secondary intent; determine whether each two correlated intents belong to the same type, delete correlated intents that do not belong to the same type, cluster the deleted correlated intents with the specified secondary intent to form a corresponding primary intent; construct a hierarchical intent schema based on the primary intent and at least one corresponding secondary intent, the hierarchical intent schema being used to provide correlated intents similar to the specified secondary intent, and the probability information and slot information of each secondary intent corresponding to a specified primary intent in the hierarchical intent schema being shared with each other, and each secondary intent being able to belong to different primary intents; calculate an updated probability of an initial intent of a user interaction sentence based on the correlated intents provided by the hierarchical intent schema, and determine a target intent of the user interaction sentence based on the updated probability.

2. The display device of claim 1, wherein, The controller, in the execution of the calculation of the intent similarity between each two secondary intents based on the example information and the slot information, is further configured to: calculate a representation vector of a corresponding secondary intent based on the example information and the slot information; calculate the intent similarity between each two secondary intents based on the representation vector of each secondary intent.

3. The display device of claim 2, wherein, The controller, in the execution of the calculation of the representation vector of a corresponding secondary intent based on the example information and the slot information, is further configured to: calculate a sentence vector of a plurality of example information corresponding to a specified secondary intent, calculate an average value of a plurality of the sentence vectors as the representation vector of the specified secondary intent based on the example information; calculate a slot vector of a plurality of slot information corresponding to the specified secondary intent, calculate an average value of a plurality of the slot vectors as the representation vector of the specified secondary intent based on the slot information; splice the representation vector of the specified secondary intent based on the example information and the representation vector of the specified secondary intent based on the slot information to obtain the representation vector of the specified secondary intent.

4. A display device, characterized by The application comprises: a display configured to present a user interface; a controller connected with the display, the controller being configured to: The controller is further configured to: obtain an interaction sentence corresponding to a current round of dialogue in a multi-round dialogue, identify an initial intent corresponding to the interaction sentence, and calculate a feature score of the initial intent, wherein an initial probability of each initial intent is obtained, and each initial intent is sorted in descending order of the initial probability; based on a ranking number of each initial intent, calculate a ranking score of each initial intent; based on an initial probability of a first initial intent, calculate an initial probability difference between each subsequent initial intent and the first initial intent, and use the initial probability, the ranking score, and the initial probability difference as the feature score of the corresponding initial intent; 5. The display device of claim 4, wherein, input the initial intent into a hierarchical intent architecture to obtain an associated intent similar to the initial intent, and calculate a feature score of the associated intent, wherein the hierarchical intent architecture is constructed based on a primary intent and at least one secondary intent corresponding to the primary intent, the intent similarity between the at least one secondary intent corresponding to the primary intent exceeds a preset similarity threshold, and the at least one secondary intent belongs to the same type, the probability information and the slot information of each secondary intent corresponding to a specified primary intent in the hierarchical intent architecture are shared with each other, and each secondary intent can belong to different primary intents; obtain a feature score of a historical intent corresponding to a historical round of dialogue in a multi-round dialogue, a feature score of a historical associated intent, a feature score of a historical user behavior, and a feature score of a historical system behavior; based on the feature scores of the initial intent, the associated intent, the historical intent, the historical associated intent, the historical user behavior, and the historical system behavior, calculate an updated probability of each initial intent, and use the updated probability to determine a target intent corresponding to the current round of dialogue. The controller is further configured to: obtain a frequency of occurrence of each historical intent corresponding to a historical round of dialogue in a multi-round dialogue, and calculate a frequency of occurrence of each historical intent; 6. The display device of claim 4, wherein, obtain an intent probability of each historical intent, calculate a total value, an average value, and a maximum value of the intent probabilities of all historical intents corresponding to the historical round of dialogue; calculate a ranking score of each historical intent, and calculate a total value, an average value, and a maximum value of the ranking scores of all historical intents corresponding to the historical round of dialogue; use the frequency of occurrence, the frequency of occurrence, the total value, the average value, and the maximum value of the intent probability, the total value, the average value, and the maximum value of the ranking score, and the maximum value of the frequency of occurrence as the feature score of the historical intent corresponding to the historical round of dialogue in the multi-round dialogue. The controller is further configured to: obtain a frequency of occurrence of each historical user behavior corresponding to a historical round of dialogue in a multi-round dialogue; based on the frequency of occurrence, calculate a frequency of occurrence of each historical user behavior, and use the frequency of occurrence and the frequency of occurrence of each historical user behavior as the feature score of the historical user behavior corresponding to the historical round of dialogue in the multi-round dialogue.

7. The display device of claim 4, wherein, The controller is further configured to: acquire the frequency of occurrence of each historical system behavior corresponding to the historical round of dialogue in the multi-round dialogue; based on the frequency of occurrence, calculate the frequency of occurrence of each historical system behavior, and take the frequency of occurrence and the frequency of occurrence of each historical system behavior as the feature score of the historical system behavior corresponding to the historical round of dialogue in the multi-round dialogue.

8. A layered intent architecture construction method, characterized by, The method comprises: acquiring an intent in a preset intent architecture as a secondary intent, and acquiring example information and slot information of the secondary intent, wherein there is no correlation between each intent in the preset intent architecture; based on the example information and slot information, calculating the intent similarity between each two secondary intents; acquiring the intent similarity corresponding to a specified secondary intent, and selecting a secondary intent with an intent similarity exceeding a preset similarity threshold as a correlated intent of the specified secondary intent; determining whether each two correlated intents belong to the same type, deleting correlated intents that do not belong to the same type, clustering the deleted correlated intents with a specified secondary intent to form a corresponding primary intent; based on the primary intent and at least one corresponding secondary intent, constructing a hierarchical intent architecture, the hierarchical intent architecture being used to provide correlated intents similar to the specified secondary intent, the probability information and slot information of each secondary intent corresponding to a specified primary intent being shared with each other in the hierarchical intent architecture, and each secondary intent being able to belong to different primary intents, wherein the correlated intents provided by the hierarchical intent architecture are used to calculate the updated probability of the initial intent of the user interaction sentence, and the updated probability is used to determine the target intent of the user interaction sentence.

9. A method for intent update based on a hierarchical intent architecture, the method comprising: The method comprises: acquiring an interaction sentence corresponding to a current round of dialogue in a multi-round dialogue, identifying an initial intent corresponding to the interaction sentence, and calculating a feature score of the initial intent, wherein an initial probability of each initial intent is acquired, each initial intent is sorted in descending order of the initial probability, a ranking score of each initial intent is calculated based on the ranking number of each initial intent, and an initial probability difference between each subsequent initial intent and the first initial intent is calculated based on the initial probability of the first initial intent, and the initial probability, the ranking score, and the initial probability difference are taken as the feature score of the corresponding initial intent; inputting the initial intent into a hierarchical intent architecture to obtain correlated intents similar to the initial intent, and calculating a feature score of the correlated intents, wherein the hierarchical intent architecture is constructed based on a primary intent and at least one corresponding secondary intent, the intent similarity between the at least one secondary intent corresponding to the primary intent exceeds a preset similarity threshold and belongs to the same type, the probability information and slot information of each secondary intent corresponding to a specified primary intent are shared with each other in the hierarchical intent architecture, and each secondary intent can belong to different primary intents. obtaining a feature score of a historical intent corresponding to a historical round of dialogue, a feature score of a historical associated intent, a feature score of a historical user behavior, and a feature score of a historical system behavior in a multi-round dialogue; based on the feature score of the initial intent, the feature score of the associated intent, the feature score of the historical intent, the feature score of the historical associated intent, the feature score of the historical user behavior, and the feature score of the historical system behavior, calculating an updated probability of each of the initial intents, the updated probability being used to determine a target intent corresponding to a current round of dialogue.

Images