An AI-based customer interaction service system

Abstract

This invention discloses an artificial intelligence-based customer interaction service system, including an interaction intent recognition module, an interaction target path planning and prediction module, and an intelligent interaction control module. The interaction robot recognizes intent based on intelligent question-and-answer from the client user, builds a model using known training corpus, and classifies unknown interaction text based on the constructed model. Before the interaction begins, the interaction robot pre-plans and generates a corresponding dialogue target path for the user based on preference information in the user profile and the corresponding personal knowledge base. It also performs emotion recognition and detection by combining user image data collected during the interaction, detects the user's emotional state in the collected images, and constructs a knowledge base that formulates corresponding voice interaction strategies based on different expression categories. The text knowledge base is converted into voice text and output to the interaction robot terminal for command control. This invention improves the accuracy of interaction robot interaction recognition.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a customer interaction service system based on artificial intelligence. Background Technology

[0002] Existing interactive robots in service venues primarily serve to display and introduce items, providing guidance services similar to those used in museums, shopping malls, or hospitals to explain basic information or guide users to relevant locations. These robots enhance the service capabilities of these venues, reduce the human and financial costs associated with traditional service methods, and increase the attractiveness of their services. Generally, these interactive robots need to provide navigation for specific items or stores, interacting with users to explain products, items, related knowledge, or materials, and guiding customers to their destinations. This partially replaces the functions of a large number of service personnel. While interactive robots utilize natural language processing technology, they still face challenges in semantic understanding, referential resolution, and contextual understanding. Furthermore, in prolonged and complex dialogue scenarios, their ability to grasp and remember context is limited, affecting the coherence of the conversation and the accuracy of their responses to user needs. Therefore, designing an AI-based customer interaction service system that improves the readiness of interaction recognition and enhances the user experience of interacting with the robot is essential. Summary of the Invention

[0003] The purpose of this invention is to provide an artificial intelligence-based customer interaction service system to solve the problems mentioned in the background art.

[0004] To address the aforementioned technical problems, the present invention provides the following technical solution: an artificial intelligence-based customer interaction service system, comprising an artificial intelligence-based customer interaction service method, the method comprising:

[0005] The interactive robot recognizes intent based on intelligent question-and-answer from client users, builds a model using known training corpus, and classifies unknown interactive text based on the built model.

[0006] Based on dialogue interaction, the interactive robot pre-plans and generates a corresponding dialogue target path for the interactive user based on the user profile's preference information and the corresponding personal knowledge base before the interaction begins.

[0007] Emotion recognition and detection are performed by combining user image data collected during the interaction process, and the emotional state of the user in the image is detected by using the collected images.

[0008] A knowledge base is constructed to formulate corresponding voice interaction strategies based on different expression categories. The text knowledge base is then converted into voice text and output to the interactive robot terminal for command control.

[0009] According to the above technical solution, the intention recognition is performed as follows: when the user starts to interact with the interactive robot by voice, the interactive robot is awakened by issuing a wake-up command. The interactive robot collects the voice signal of the interaction, performs signal feature processing based on the given filtered data, and constructs an interactive intention prediction model.

[0010] The training corpus is in the format (x, y), where x is the training data and y is the desired output category or intent. Using the prepared training data, training is performed according to a given filtering format. The required data features are extracted to form feature vectors. These extracted feature vectors, along with the corresponding desired output, are input into a deep learning algorithm to train the prediction model. The same feature extraction is performed on updated data to obtain feature vectors for prediction. The trained prediction model is then used to predict the updated data after feature processing, and the prediction result is returned. The algorithm determines whether the data is a wake-up word. If it is not a wake-up word for interactive intent recognition, it remains silent. If it is an interactive wake-up word, it enters the interactive wake-up state. Further, the network status of the interactive communication is determined, and different voice interaction methods are selected based on different network statuses.

[0011] According to the above technical solution, the interactive user pre-plans and generates a corresponding dialogue target path, which includes: using the question-and-answer dialogue data in the intelligent question-and-answer dialogue as a dataset, training an end-to-end sequence generation model, and returning the output according to the user's input.

[0012] According to the above technical solution, the pre-planning of a corresponding dialogue target path for the interactive user also includes, when a user's personal profile exists, using the user's personal profile and corresponding personal knowledge base, planning dialogue targets for the user before the dialogue begins, generating an initial target sequence containing n dialogue targets, and ensuring that there are no contradictory or repetitive dialogue questions between adjacent dialogue targets; when the user profile is empty, filling the user profile of the target sequence with multiple questions to model the user; and when the user profile is filled, the initial dialogue target sequence includes the points of interest in the profile.

[0013] According to the above technical solution, the pre-planning and generation of a corresponding dialogue target path by the interactive user also includes: data annotation of the dialogue target based on the user's current intent; the interactive robot and the user complete the dialogue target in x rounds of interaction; when the user actively starts new dialogue interaction content, the current dialogue target ends; the dialogue target sequence contains multiple target nodes, each node represents a dialogue target; a complete round of interactive dialogue corresponds to a target sequence; the nodes in the sequence have a dependency and sequential relationship; during the dialogue, the machine determines which dialogue target should be attributed to based on the user's output text, and at the same time determines whether the current dialogue target has been completed; if completed, the next dialogue target should be predicted.

[0014] According to the above technical solution, the detection of the user's emotional state in the image includes: using user image data collected during the interaction, transmitting the data frames collected by the camera in a frame-by-frame manner to detect the user's emotions, detecting and locating the user's face in each image frame, using bilateral filtering to reduce noise interference in the cropped facial image, performing scale transformation based on the detected facial region, performing coarse and fine registration on the feature points of expression recognition, matching the generated registered facial expression feature point descriptor with the descriptor of the next frame image to generate matched feature points, using the feature points as the initial facial expression for fitting the next frame image, obtaining the model features and appearance model features of the fitted facial region, extracting features from the appearance model features, and then combining them with the facial region model features to form expression features, classifying the expression features using an SVM classifier, and outputting the expression category.

[0015] According to the above technical solution, the detection of the user's emotional state in the image includes: using user image data collected during the interaction, transmitting the data frames collected by the camera in a frame-by-frame manner to detect the user's emotions, detecting and locating the user's face in each image frame, using bilateral filtering to reduce noise interference in the cropped facial image, performing scale transformation based on the detected facial region, performing coarse and fine registration on the feature points of expression recognition, matching the generated registered facial expression feature point descriptor with the descriptor of the next frame image to generate matched feature points, using the feature points as the initial facial expression for fitting the next frame image, obtaining the model features and appearance model features of the fitted facial region, extracting features from the appearance model features, and then combining them with the facial region model features to form expression features, classifying the expression features using an SVM classifier, and outputting the expression category.

[0016] According to the above technical solution, the construction of a knowledge base for formulating corresponding voice interaction strategies based on different expression categories includes: after the intent recognition node recognizes the interactive text, the next interaction instruction is divided into interactive question-and-answer instructions and control instructions. The question-and-answer instructions are further divided into offline and online states. When the question-and-answer instructions are in an offline state, intent recognition is performed by searching the database. If the recognition is successful, the interactive text is output. If the recognition fails, the question-and-answer instructions are searched through the network. When the intent recognition node recognizes that it is a control instruction, resource search is performed through the knowledge base of the corresponding voice interaction strategy.

[0017] According to the above technical solution, the control of the interactive robot terminal by instructions includes: in the intent recognition node, embedding the user interaction content and the local database through word vectors, obtaining semantic vectors using extracted text features, converting the similarity of the interactive texts into a pre-selected similarity between the two semantic vectors, sorting by cosine values, outputting the similar question with the highest confidence as the interactive text returned by the interaction, and publishing the interactive content. At this time, the interactive robot mobile node receives and integrates the interactive content through message subscription, and the interactive robot calls the received interactive text through the program to transmit the audio playback node for audio playback.

[0018] According to the above technical solution, the artificial intelligence-based customer interaction service system includes:

[0019] The interaction intent recognition module is used by the interactive robot to recognize the intent of the client user based on intelligent question and answer;

[0020] The interactive target path planning and prediction module is used to pre-plan and generate a corresponding dialogue target path for interactive users;

[0021] The intelligent interactive control module is used to convert the text knowledge base into voice text and output it to the interactive robot terminal for command control.

[0022] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention, by setting up an interactive intent recognition module, an interactive target path planning and prediction module, and an intelligent interactive control module, enables the interactive robot to recognize intent based on intelligent question-and-answer from the client user, construct a model using known training corpus, classify unknown interactive text based on the constructed model, and pre-plan and generate a corresponding dialogue target path for the interactive user based on the user profile's preference information and corresponding personal knowledge base before the interaction begins. It also combines user image data collected during the interaction to perform emotion recognition and detection, uses the collected images to detect the user's emotional state in the images, constructs a knowledge base to formulate corresponding voice interaction strategies based on different expression categories, and converts the text knowledge base into voice text for output to the interactive robot terminal for command control. This invention improves the accuracy of interactive robot interaction recognition and enhances the user's interactive experience with the interactive robot. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0024] Figure 1 This is a schematic diagram of the module composition of a customer interaction service system based on artificial intelligence, provided in Embodiment 2 of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: This example can be applied to scenarios involving interactive robots in service-oriented venues. The method can be executed by an AI-based customer interaction service system provided in this example. The method specifically includes the following steps:

[0027] The interactive robot recognizes intent based on intelligent question-and-answer from client users, builds a model using known training corpus, and classifies unknown interactive text based on the built model.

[0028] Based on dialogue interaction, the interactive robot pre-plans and generates a corresponding dialogue target path for the interactive user based on the user profile's preference information and the corresponding personal knowledge base before the interaction begins.

[0029] Emotion recognition and detection are performed by combining user image data collected during the interaction process, and the emotional state of the user in the image is detected by using the collected images.

[0030] A knowledge base is constructed to formulate corresponding voice interaction strategies based on different expression categories. The text knowledge base is then converted into voice text and output to the interactive robot terminal for command control.

[0031] In this embodiment of the invention, when a user begins voice interaction with the interactive robot, a wake-up command is issued to wake up the interactive robot's interaction intent recognition. The interactive robot collects the voice signal of the interaction, performs signal feature processing based on given filtered data, and constructs an interaction intent prediction model. The training corpus format is (x, y), where x is the training corpus and y is the expected output category or intent. Using the prepared training data, training is performed according to the given filtering format, and the required data features are extracted to form a feature vector. The extracted feature vector and the corresponding expected output are input into a deep learning algorithm to train the prediction model. The same feature extraction is performed on the updated data to obtain the feature vector used for prediction. The trained prediction model is used to predict the updated data after feature processing, and the recognition prediction result is returned. It is judged whether it is a wake-up word. If it is not judged to be a wake-up word for interaction intent recognition, it is left as is. If it is judged to be an interaction wake-up word, it enters the interaction wake-up state. The network status of the interaction communication is further judged, and different voice interaction methods are selected with the user according to different network status forms.

[0032] For example, in the absence of a network, interactive words are recognized based on the content of user interaction. AIML is used to intelligently match the data information in the offline intelligent question-and-answer library. When the matching result is a control command that exists in the command library, the interactive robot performs command mapping action matching according to the corresponding control command. When the matching result is not a control command, the interactive robot performs speech synthesis and output the information queried from the interactive question-and-answer library.

[0033] For example, in a network-connected environment, the interactive robot engages in online voice interaction with the user. After the user's interactive commands are processed by speech recognition, the open platform analyzes the user's intent, identifies the third-party application, converts the user's intent into structured information, and sends it to the application. The application processes the user's intent and returns the processed information as structured information to the open platform. Simultaneously, it determines whether the command is a control command. If it is a control command, the robot performs the interactive output according to the command. The open platform converts the text information of the interaction, returns the converted text, synthesizes it, and plays the synthesized interactive voice output.

[0034] In this embodiment of the invention, the question-and-answer dialogue corpus from intelligent question-and-answer dialogues is used as a dataset to train an end-to-end sequence generation model, which returns the output based on the user's input.

[0035] For example, when a user's personal profile exists, the user's personal profile and corresponding personal knowledge base are used to plan the dialogue goals for the user before the dialogue begins, generating an initial target sequence containing n dialogue goals, and there are no contradictory or repetitive dialogue questions between adjacent dialogue goals. When the user profile is empty, the user profile of the target sequence is filled by asking multiple questions to model the user. When the user profile is filled, the initial dialogue target sequence includes the points of interest in the profile.

[0036] For example, the data is labeled with the dialogue goal based on the user's current intent. The interactive robot and the user complete the dialogue goal in x rounds of interaction. When the user actively starts new dialogue interaction content, the current dialogue goal ends. The dialogue goal sequence contains multiple goal nodes, each node represents a dialogue goal. A complete round of interactive dialogue corresponds to a goal sequence. The nodes in the sequence have a dependency and sequence relationship. When the dialogue is in progress, the machine determines which dialogue goal should be attributed to based on the user's output text, and at the same time determines whether the current dialogue goal has been completed. If it has been completed, the next dialogue goal should be predicted.

[0037] In this embodiment of the invention, user image data collected during the interaction process is transmitted in frames using a camera to detect the user's emotions. The facial features for emotion recognition in each frame are detected and located. Bilateral filtering is used to reduce noise interference in the cropped facial images. Scale transformation is performed on the detected facial regions. Coarse and fine registration are performed on the feature points for expression recognition. The generated facial expression feature point descriptors are matched with the descriptors of the next frame image to generate matched feature points. These feature points are used as the initial facial expression for fitting the next frame image. The model features and appearance model features of the fitted facial regions are obtained. Feature extraction is performed on the appearance model features, and then combined with the facial region model features to form expression features. An SVM classifier is used to classify the expression features and output the expression category.

[0038] For example, emotion detection involves facial expression emotion recognition. The data output from emotion detection includes two parts: the first part is the image detected and identified from the original image, and the second part is the result of emotion recognition and classification. The current output result is compared with the previous detection output result to determine if there has been a change. If there is a change in emotion, the current emotion state is modified. For the first detection, since there is no result from the previous emotion recognition, the default initial emotion state is set to the natural state. The output expression category data is stored in the local database to build a knowledge base that formulates corresponding voice interaction strategies based on different expression categories.

[0039] In this embodiment of the invention, while the interactive robot obtains the user's interactive content, it performs feature fusion based on the user's facial expressions and interactive voice data to construct multimodal emotion recognition. The interactive robot interacts based on the tags in the local database and returns content information. When the user's interactive content and facial expression data obtained by the interactive robot cannot be found in the local data, the interactive robot crawls the user's interaction information from the Internet and uses an open-source chat interaction question and answer interface to obtain the interaction-related information.

[0040] For example, the text knowledge base is converted into speech text and output to the interactive robot terminal for command control. After the intent recognition node recognizes the interactive text, the next interactive command is divided into interactive question-and-answer commands and control commands. The question-and-answer commands are further divided into offline and online states. When the question-and-answer command is in the offline state, the intent is recognized by searching the database. If the recognition is successful, the interactive text is output. If the recognition fails, the question-and-answer command is searched through the network. When the intent recognition node recognizes that it is a control command, the resource is searched through the knowledge base of the corresponding voice interaction strategy. In the intent recognition node, the user interaction content and the local database are embedded through word vectors. The semantic vector is obtained by using the extracted text features. The similarity of the interactive text is converted into the pre-selected similarity of the two semantic vectors. By sorting the cosine values, the similar question with the highest confidence is output as the interactive text returned by the interaction, and the interactive content is published. At this time, the interactive robot mobile node receives the integrated interactive content through message subscription. The interactive robot calls the program to pass the received interactive text to the voice playback node for voice playback.

[0041] Example 2: Example 2 of the present invention provides a customer interaction service system based on artificial intelligence. Figure 1 This is a schematic diagram of the module composition of a customer interaction service system based on artificial intelligence provided in Embodiment 2 of the present invention, as shown below. Figure 1 As shown, the system includes:

[0042] The interaction intent recognition module is used by the interactive robot to recognize the intent of the client user based on intelligent question and answer;

[0043] The interactive target path planning and prediction module is used to pre-plan and generate a corresponding dialogue target path for interactive users;

[0044] The intelligent interactive control module is used to convert the text knowledge base into voice text and output it to the interactive robot terminal for command control.

[0045] In some embodiments of the present invention, the interaction intent recognition module includes:

[0046] The interaction intent prediction module is used to construct an interaction intent prediction model by performing signal feature processing based on given filtered data.

[0047] The interactive communication network status module is used to determine the network status of interactive communication and select different voice interaction methods with the user based on different network status forms.

[0048] The voice interaction output module is used by the interactive robot to synthesize and output the information retrieved from the interactive question and answer database.

[0049] In some embodiments of the present invention, the interactive target path planning and prediction module includes:

[0050] The sequence generation model module is used to train an end-to-end sequence generation model using the question-and-answer dialogue data from intelligent question-and-answer dialogues as a dataset.

[0051] The user profiling module is used to populate the user profile of the target sequence by asking multiple questions.

[0052] The target sequence generation module is used to generate an initial target sequence based on the user's personal profile and corresponding knowledge base based on the interaction.

[0053] The dialogue target prediction module is used to label and predict the dialogue targets of interactive users.

[0054] In some embodiments of the present invention, the intelligent interactive control module includes:

[0055] The interactive user emotion detection module is used to detect the emotions of interactive users through user image data collected during the interaction process.

[0056] The facial expression category classification module is used to classify facial expression features using an SVM classifier and output the facial expression category.

[0057] The interaction strategy knowledge base generation module is used to build a knowledge base that formulates corresponding voice interaction strategies based on different expression categories;

[0058] The interactive recognition and control module is used to convert the text knowledge base into speech text and output it to the interactive robot terminal for command control.

[0059] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0060] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-based customer interaction service system, executing an artificial intelligence-based customer interaction service method, characterized in that: The method includes: S1. The interactive robot recognizes intent based on intelligent question-and-answer from client users, builds a model using known training corpus, and classifies unknown interactive text based on the built model; The intention recognition process is as follows: when a user begins to interact with the interactive robot via voice, the interactive robot is activated by issuing a wake-up command. The interactive robot collects the voice signals of the interaction and performs signal feature processing based on given filtered data to construct an interactive intention prediction model. The training corpus is in the format (x, y), where x is the training corpus and y is the expected output category or intent. Using the prepared training data, training is performed according to a given filtering format. The required data features are extracted to form feature vectors. These extracted feature vectors, along with the corresponding expected output, are input into a deep learning algorithm to train the prediction model. The same feature extraction is performed on updated data to obtain feature vectors for prediction. The trained prediction model is then used to predict the updated data after feature processing, and the prediction result is returned. The system determines whether the text is a wake-up word. If it is not a wake-up word for interactive intent recognition, it remains silent. If it is an interactive wake-up word, it enters the interactive wake-up state. Further, the network status of the interactive communication is determined, and different voice interaction methods are selected based on different network statuses. In the absence of network, the interactive robot recognizes interactive words based on user interaction content and intelligently matches them with data information in the offline intelligent question-and-answer library through AIML. When the matching result is a control command that exists in the command library, the interactive robot performs command mapping action matching according to the corresponding control command. When the matching result is not a control command, the interactive robot synthesizes and outputs the information from the query interactive question-and-answer library. When connected to the internet, the interactive robot engages in online voice interaction with the user. After the user's interactive commands are recognized by speech recognition, the open platform analyzes the user's intent, identifies the third-party application, converts the user's intent into structured information, and sends it to the application. The application processes the user's intent and returns the processed information as structured information to the open platform. Simultaneously, it determines whether the command is a control command. If it is a control command, the robot performs the interactive output according to the command. The open platform converts the interactive text information, returns the converted text, synthesizes it, and plays the synthesized interactive voice output. S2. Based on dialogue interaction, the interactive robot pre-plans and generates a corresponding dialogue target path for the interactive user based on the user profile's preference information and the corresponding personal knowledge base before the interaction begins. The step of pre-planning and generating a corresponding dialogue target path for interactive users includes: using the question-and-answer dialogue data in intelligent question-and-answer dialogue as a dataset, training an end-to-end sequence generation model, and returning the output based on the user's input; When a user profile exists, the user profile and corresponding personal knowledge base are used to plan the dialogue goals for the user before the dialogue begins, generating an initial target sequence containing n dialogue goals. There are no contradictory or repetitive dialogue questions between adjacent dialogue goals. When the user profile is empty, the user profile of the target sequence is filled by asking multiple questions to model the user. When the user profile is filled, the initial dialogue target sequence includes the points of interest in the profile. The pre-planning of a corresponding dialogue target path by the interactive user also includes: data labeling of the dialogue target based on the user's current intent; the interactive robot and the user complete the dialogue target in x rounds of interaction; when the user actively starts new dialogue interaction content, the current dialogue target ends; the dialogue target sequence contains multiple target nodes, each node represents a dialogue target; a complete round of interactive dialogue corresponds to a target sequence; the nodes in the sequence have a dependency and sequential relationship; during the dialogue, the machine determines which dialogue target should be attributed to based on the user's output text, and at the same time determines whether the current dialogue target has been completed; if completed, the next dialogue target should be predicted. S3. Combine user image data collected during the interaction process to perform emotion recognition and detection, and use the collected images to detect the user's emotional state in the images; The process of detecting the user's emotional state in the image includes: using user image data collected during the interaction, transmitting the data frames collected by the camera in a frame-by-frame manner to detect the user's emotions, detecting and locating the user's face in each image frame, using bilateral filtering to process the cropped facial image to reduce noise interference, performing scale transformation based on the detected facial region, performing coarse and fine registration on the feature points of expression recognition, matching the generated registered facial expression feature point descriptor with the descriptor of the next frame image to generate matched feature points, using the feature points as the initial facial expression for fitting the next frame image, obtaining the model features and appearance model features of the fitted facial region, extracting features from the appearance model features, and then combining them with the facial region model features to form expression features, classifying the expression features using an SVM classifier, and outputting the expression category; In emotion detection, facial expression emotion recognition is performed. The data output from emotion detection includes two parts: the first part is the image detected and identified from the original image, and the second part is the output of an emotion recognition classification result. The current output result is compared with the previous detection output result to determine if there has been a change. If there is a change in emotion, the current emotion state is modified. For the first detection, since there is no previous emotion recognition result, the default initial emotion state is set to the natural state. The output expression category data is stored in the local database to build a knowledge base that formulates corresponding voice interaction strategies based on different expression categories. While the interactive robot obtains the user's interaction content, it performs feature fusion based on the user's facial expressions and voice data to construct multimodal emotion recognition. The interactive robot interacts based on the tags in the local database and returns content information. When the user's interaction content and facial expression data obtained by the interactive robot cannot be found in the local data, the interactive robot crawls the user's interaction information from the Internet and uses an open-source chat interaction question and answer interface to obtain the interaction-related information. S4. Construct a knowledge base that formulates corresponding voice interaction strategies based on different facial expression categories, and convert the text knowledge base into voice text output to the interactive robot terminal for command control; The construction of a knowledge base for formulating corresponding voice interaction strategies based on different expression categories includes: after the intent recognition node recognizes the interactive text, the next interaction instruction is divided into interactive question-and-answer instructions and control instructions. The question-and-answer instructions are further divided into offline and online states. When the question-and-answer instructions are in the offline state, intent recognition is performed by searching the database. If the recognition is successful, the interactive text is output. If the recognition fails, the question-and-answer instructions are searched through the network. When the intent recognition node recognizes that it is a control instruction, the resource is searched through the knowledge base of the corresponding voice interaction strategy. The control of the interactive robot terminal by instructions includes: in the intent recognition node, embedding user interaction content and local database through word vectors, obtaining semantic vectors using extracted text features, converting the similarity of interactive texts into a pre-selected similarity between two semantic vectors, sorting by cosine values, outputting the most confident similar question as the interactive text returned by the interaction, and publishing the interactive content. At this time, the interactive robot mobile node receives and integrates the interactive content through message subscription, and the interactive robot calls the received interactive text through the program to pass it to the voice playback node for voice playback.

Images