An interactive method, device and medium based on AIGC robot

Abstract

The present invention discloses an interactive method, device and medium based on AIGC robot, and relates to the field of data processing technology. The method includes: logging into the platform and creating an AIGC robot, entering the robot design page, creating a persona reply logic; collecting a user dialogue data set, optimizing the persona reply logic based on the data set, and obtaining the optimized logic; building a retrieval knowledge base, configuring the workflow, building a multi-round dialogue process and page navigation logic; configuring and publishing the AIGC robot based on the optimized reply logic and multi-round dialogue process; receiving user instructions, the AIGC robot interacts with the user according to the instructions and the knowledge base, so as to solve the technical problems of context understanding, knowledge update lag and insufficient accuracy of generated content in the prior art in human-computer interaction, and achieve the technical effect of improving the response accuracy and real-time performance of the AIGC robot in human-computer interaction.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to an interaction method, device and medium based on an AIGC robot. Background Art

[0002] With the rapid development of artificial intelligence technology, especially the rise of large language models and generative artificial intelligence (AIGC), more and more companies and platforms have begun to apply these technologies to various scenarios to improve user experience and work efficiency. As an intelligent assistant based on natural language processing (NLP) and deep learning technology, AIGC robots have been widely used in many fields such as intelligent customer service, virtual assistants, and dialogue systems. However, existing AIGC robots still face many challenges, including insufficient depth of contextual understanding, lagging knowledge updates, differences in the accuracy of generated content, and inability to effectively handle complex multi-round conversations. These problems have led to a decline in user experience, especially in multi-round conversations and complex interaction scenarios, where AIGC robots often fail to provide accurate and coherent services. Summary of the invention

[0003] The present application provides an interaction method, device and medium based on an AIGC robot, which solves the technical problems of context understanding, knowledge update lag and insufficient accuracy of generated content in the prior art in human-computer interaction. By configuring multi-round dialogue processes and page navigation logic, and configuring and publishing the AIGC robot based on optimized reply logic, the technical effect of improving the response accuracy and real-time performance of the AIGC robot in human-computer interaction is achieved, thereby improving user experience and interaction efficiency.

[0004] In view of the above problems, the present application provides an interaction method, device and medium based on an AIGC robot.

[0005] In a first aspect of the present application, an interaction method based on an AIGC robot is provided, the method comprising: logging into a platform and creating an AIGC robot, obtaining a robot design page, creating a persona response logic for the AIGC robot through the robot design page, the persona response logic comprising a model role, a language style and a model answer range; acquiring a user dialogue data set, optimizing the persona response logic based on the user dialogue data set, and obtaining a persona optimized response logic; constructing a retrieval knowledge base, and configuring and arranging a target workflow at the same time, and constructing a multi-round dialogue process and a page navigation logic according to the target workflow; configuring and publishing the AIGC robot based on the persona optimized response logic and the multi-round dialogue process and the page navigation logic to obtain a target AIGC robot; obtaining a user input instruction, and the target AIGC robot interacts with a target user based on the user input instruction and the retrieval knowledge base.

[0006] According to a second aspect of the present application, there is provided an electronic device comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is used to execute an interaction method based on an AIGC robot provided in the present application.

[0007] In the third aspect of the present application, the present application provides a computer-readable storage medium storing a computer program, and the computer program is used to execute an interaction method based on an AIGC robot provided by the present application.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] First, log in to the platform and create an AIGC robot, obtain the robot design page, and create a personality response logic for the AIGC robot through the robot design page. The personality response logic includes model roles, language styles, and model answer ranges. Subsequently, the user dialogue data set is collected and obtained, and the personality response logic is optimized based on the user dialogue data set to obtain the personality optimization response logic. Further, a retrieval knowledge base is constructed, and the target workflow is configured and arranged at the same time. According to the target workflow, a multi-round dialogue process and page navigation logic are constructed. Afterwards, the AIGC robot is configured and published based on the personality optimization response logic and the multi-round dialogue process and page navigation logic to obtain the target AIGC robot. Finally, the user input instruction is obtained, and the target AIGC robot interacts with the target user based on the user input instruction and the retrieval knowledge base. It solves the technical problems of existing technologies in human-computer interaction, such as context understanding, delayed knowledge updating, and insufficient accuracy of generated content. By configuring multi-round dialogue processes and page navigation logic, and configuring and releasing the AIGC robot based on optimized reply logic, it achieves the technical effect of improving the response accuracy and real-time performance of the AIGC robot in human-computer interaction, thereby enhancing user experience and interaction efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 A flowchart of an interaction method based on an AIGC robot is provided in an embodiment of the present application.

[0012] Figure 2A flowchart of obtaining a personality optimization reply logic based on an AIGC robot interaction method provided in an embodiment of the present application.

[0013] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0014] Description of the reference numerals: processor 11 , memory 12 , input device 13 , output device 14 . DETAILED DESCRIPTION

[0015] The present application solves the technical problems of context understanding, knowledge update lag and insufficient accuracy of generated content in the prior art in human-computer interaction by providing an interaction method, device and medium based on an AIGC robot. By configuring a multi-round dialogue process and page navigation logic, and configuring and publishing the AIGC robot based on optimized reply logic, the technical effect of improving the response accuracy and real-time performance of the AIGC robot in human-computer interaction is achieved, thereby improving user experience and interaction efficiency.

[0016] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0017] It should be noted that the terms "including" and "having" are intended to cover non-exclusive inclusions. For example, a process, method, device, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules that are not explicitly listed or are inherent to these processes, methods, products or devices.

[0018] Embodiment 1, as Figure 1 As shown, the present application provides an interaction method based on an AIGC robot, wherein the method comprises:

[0019] Log in to the platform and create an AIGC robot to obtain a robot design page. Use the robot design page to create a personality response logic for the AIGC robot. The personality response logic includes a model role, a language style, and a model answer range.

[0020] In one embodiment, after a user logs in to the platform and creates a new AIGC robot, the user will obtain the robot design page of the new AIGC robot. The user can set specific character settings and response logic for the robot according to the account permissions, which constitutes the character response logic of the AIGC robot. This character response logic refers to the rules for designing the interaction method for the AIGC robot, including model roles, language styles, and model answer ranges. The model role specifies a virtual identity for the AIGC robot, such as a fleet manager, vehicle monitoring assistant, driving behavior analysis consultant, etc. The role setting will affect the robot's tone, behavior style, and the context of interaction with the user. The language style The language style is determined by the role setting of the AIGC robot, such as formal, serious, etc. This setting can help the AIGC robot adapt to different user needs in different application scenarios. For example, a serious tone is required when dealing with abnormal driving alarms (such as smoking alarms, distracted driving alarms, etc.) and vehicle offline. The model answer range defines the range of questions that the AIGC robot can answer, including the services or knowledge areas it can provide. For example, the AIGC robot can be limited to vehicle operation status monitoring, driving behavior analysis, vehicle offline status prompts, etc., to ensure that the AIGC robot can accurately handle specific problems or abnormal situations in fleet information. Through these settings, the AIGC robot can interact with users in a more personalized and accurate manner according to the set personality response logic, especially in fleet information management, and can efficiently respond to user needs and provide accurate suggestions or alarms.

[0021] A user conversation data set is collected and acquired, and the persona response logic is optimized based on the user conversation data set to obtain a persona optimized response logic.

[0022] In one embodiment, a large amount of user conversation data is collected through continuous interaction with users to form a user conversation dataset. This user conversation dataset includes user input, robot replies, and various contextual information in the interaction. Based on these conversation records, the user's intention is extracted to better understand the user's behavior and language patterns. Subsequently, based on the extracted user intention, the existing persona reply logic is optimized, and the optimization process includes reply defect identification, optimization strategy analysis, etc. After optimization, the persona optimized reply logic can be obtained. The AIGC robot can understand the user's intention more accurately through this persona optimized reply logic, and provide responses that are more in line with user needs in subsequent interactions. The optimized persona reply logic enables the robot to flexibly respond to more diverse user questions and improve the naturalness and effectiveness of the interaction.

[0023] Furthermore, if Figure 2 As shown, the logic of obtaining the personality optimization reply includes:

[0024] Perform text segmentation on the user conversation data set to obtain a user conversation unit set; use TF-IDF to perform text processing on the user conversation unit set to obtain a conversation unit text vector set; perform user intent recognition and classification on the conversation unit text vector set to obtain a user intent classification data set; optimize and verify the persona response logic based on the user intent classification data set to obtain a persona optimized response logic.

[0025] Preferably, first, the collected user dialogue data set is segmented, and the dialogue content is split into sentences, that is, each question is split into a separate user dialogue unit, and each answer is also split into a separate user dialogue unit. In this way, the user dialogue data set can be converted into a user dialogue unit set, and each dialogue unit represents a specific interaction segment, which may be a user's question and an AIGC robot's answer. For example, the user inputs "the number of offline vehicles in a certain period of time", which can be used as a dialogue unit, and the specific number of offline vehicles answered by the AIGC robot is also another dialogue unit. After text segmentation, TF-IDF (TermFrequency-Inverse Document The TF-IDF algorithm processes each dialogue unit. The TF-IDF algorithm can quantify the importance of each word in the entire data set. TF stands for word frequency, which refers to the frequency of occurrence of a word in a single document. IDF stands for inverse document frequency, which measures the rarity of a word in the entire corpus. By calculating the product of TF and IDF, the weight of each word can be obtained, which is then converted into a dialogue unit text vector, that is, each dialogue unit is converted into a high-dimensional vector to represent its position in the semantic space, thereby obtaining a dialogue unit text vector set. Subsequently, the generated dialogue unit text vector set is further processed and a deep learning model is used for user intent recognition and classification. The model infers the user's actual intent based on the content of the dialogue unit and outputs a user intent classification dataset, in which each dialogue unit is labeled as a specific intent category, such as query, alarm, operation guidance, etc. Subsequently, the user intent classification dataset is used to optimize the previously defined persona response logic. By comparing dialogue samples of different intent categories with the current response logic, the deficiencies or defects are identified and optimized. The optimized persona response logic will better meet the needs of different users and ensure more accurate and effective responses in actual interactions. Through the above steps, the AIGC robot can continuously learn and adjust based on the large amount of user conversation data collected, improve the quality of interaction with users, and enable it to more accurately identify user intentions and make appropriate responses.

[0026] Furthermore, obtaining the dialogue unit text vector set includes:

[0027] Perform word segmentation processing on each dialogue unit in the user dialogue unit set to obtain a dialogue unit text segmentation set; obtain a text cleaning program, which includes stop word removal, special symbol removal, punctuation mark removal and text format unification; perform text cleaning on the dialogue unit text segmentation set based on the text cleaning program to obtain an available dialogue unit text segmentation set; use TF-IDF to calculate TF-IDF value and perform text vector conversion on the available dialogue unit text segmentation set to obtain the dialogue unit text vector set.

[0028] Optionally, after obtaining the user dialogue unit set, word segmentation is performed on each user dialogue unit. Word segmentation is a very important step in natural language processing. The purpose is to divide continuous text into smaller units, usually words or subwords. When performing word segmentation, the text information within each user dialogue unit is judged to determine the language used in the text of each user dialogue unit. If it is Chinese text, Chinese word segmentation tools such as jieba, THULAC, HanLP, etc. are called. If it is English text, English word segmentation tools such as NLTK, spaCy, etc. are called. These word segmentation tools break down continuous text into individual words through a combination of internal language models, dictionary matching, and rules. For example, if a user's dialogue unit is "Please generate the offline quantity of the first fleet", the result after word segmentation is: ["Please", "generate", "first", "fleet", "offline", "quantity"]. After completing the word segmentation of all dialogue units, all the segmented words are sorted by dialogue unit and added to the dialogue unit text word segmentation set together; after obtaining the dialogue unit text word segmentation set, the text cleaning program is obtained and activated. This text cleaning program includes at least four subroutines: stop word removal, special symbol removal, punctuation removal, and text format unification; by comparing the dialogue unit text word segmentation set with the stop word library through the stop word removal subroutine, all stop words in the dialogue unit text word segmentation set are removed. Stop words refer to words that do not carry useful information in text processing. For example, "de", "le", "zai", etc. Stop words do not provide substantial help for text analysis, so they need to be removed; by comparing the dialogue unit text word segmentation set with the special symbol library through the special symbol removal subroutine, special characters in the text are removed, such as meaningless symbols like @, #, $, etc. They usually affect the accuracy of subsequent analysis; by comparing the dialogue unit text word segmentation set with the punctuation library through the punctuation removal subroutine, punctuation marks in the text are removed, such as full stops, commas, question marks, etc. because these symbols are not helpful for analyzing the actual semantics of the text. The text format unification subroutine will compare the dialogue unit text word segmentation set with the standard word segmentation format and convert the segmented words that do not meet the requirements of the standard word segmentation format. For example, unified numerical formats (Arabic numerals), date formats (xxxx-xx-xx), etc. can all be completed in this step; after completing the text cleaning, the dialogue unit text word segmentation set is converted into an available dialogue unit text word segmentation set. This available dialogue unit text word segmentation set contains the effective vocabulary after removing stop words, special symbols, and punctuation marks and conforms to the standard text format;Subsequently, TF-IDF is used to process the available dialogue unit text segmentation set. TF-ID is a text feature extraction method that can quantify the importance of each word in the text. For each segmentation in the available dialogue unit text segmentation set, the TF value of the segmentation is calculated. The TF value represents the frequency of occurrence of the segmentation in the dialogue unit. For example, if the vehicle appears 3 times in a dialogue unit and the total number of words in the text is 10, the TF value of the segmentation is 0.3. The IDF value of the segmentation is obtained by taking the logarithm of the ratio of the total number of dialogue units to the number of dialogue units containing the segmentation. The IDF value represents the rarity of the word in the entire available dialogue unit text segmentation set. If a segmentation appears in all dialogue units, its IDF value will be low, indicating that the word does not have the ability to distinguish. Then, the IDF value of each segmentation is obtained by multiplying the TF value and the IDF value. TF-IDF value, and then represent the TF-IDF value of each word segment as a vector. For each dialogue unit, the vector calculated by TF-IDF will represent the position of the dialogue unit in the semantic space. For example, after cleaning, a dialogue unit contains the following words, ["vehicle", "fault", "repair"]. After applying TF-IDF calculation to these words, a vector [0.35, 0.45, 0.6] will be obtained. This vector is the feature representation of the dialogue unit. Finally, all dialogue units will be converted into similar vectors, which will form a dialogue unit text vector set. Through the above steps, the original user dialogue can be converted into a feature vector that can be processed by the deep learning model, providing a basis for subsequent intent recognition. These vectors not only contain the core information of the text, but also help the model understand the importance of each word in the context. ;

[0029] Furthermore, obtaining the user intention classification dataset includes:

[0030] A historical user conversation data set is collected and acquired, and the historical user conversation data set is preprocessed to obtain a historical conversation text vector set; an intention category label library is constructed, and the historical conversation text vector set is classified and labeled based on the intention category label library to obtain a historical conversation text intention sample set; a support vector machine is used to perform classification training on the historical conversation text intention sample set to obtain a conversation text intention classifier; user intention recognition and classification are performed on the conversation unit text vector set based on the conversation text intention classifier to obtain the user intention classification data set.

[0031] Optionally, in order to obtain the user's intention, first, collect conversation data from the interaction record log to construct a historical user conversation data set, and then use the aforementioned text segmentation and text processing steps to preprocess the historical user conversation data set, and convert the historical user conversation data set into a historical conversation text vector set so that it can be processed by the deep learning model; then, construct an intention category label library according to the actual application scenario, and this label library defines different intention categories, such as vehicle records, driver status, fault handling, abnormal offline, offline time, etc. Each intention category corresponds to a user's demand or behavior pattern. According to the intention category label library, assign a corresponding intention label to each vector in the historical conversation text vector set. For example, "I need the driving behavior of all drivers in the first fleet" is labeled as driver status, and "I need vehicles in the first fleet that are offline for no more than 10 minutes" is labeled as offline time. This labeling process can be completed by manual labeling or automatic labeling tools. Through this classification labeling step, each historical conversation text vector is converted into a sample containing a text vector and a corresponding intention label to form a historical conversation text intention sample set, each sample includes a vector representation and a label, and the label represents the user's intention;Afterwards, the annotated historical conversation text intent sample set is used as training data and input into the support vector machine (SVM) classification model for classification training. Specifically, the historical conversation text intent sample set is divided into a training set and a test set. Usually 80% of the data is used as the training set and 20% of the data is used as the test set. Then, the support vector machine (SVM) model is selected and configured, including selecting a suitable kernel function, such as a linear kernel, a radial basis function (RBF) kernel, etc., and adjusting the hyperparameters of the SVM, such as the C value (penalty parameter) and the Gamma value (for the RBF kernel function). These kernel functions and hyperparameters determine how the model maps data points in high-dimensional space. Then, the training set is input into the constructed SVM model to start the training process. By inputting the text vector and target label of each sample, the SVM model searches for an optimal decision hyperplane by maximizing the classification interval between different categories to ensure that the model can accurately distinguish different intent categories. During the training process, the SVM determines the position of the hyperplane by selecting a support vector and uses an optimization algorithm (such as ladder The parameters are adjusted to ensure that the model can correctly classify. During training, SVM optimizes model performance by adjusting hyperparameters such as C value and Gamma value to ensure that the model has good generalization ability when facing new data. After training, the test set is used to evaluate the performance of the model. By comparing the prediction results and the actual labels, the model's accuracy, precision, recall rate and F1 value are calculated. If the performance of the model meets expectations, the SVM classifier can be deployed as the final user intent classifier for real-time recognition of user intent. If the model performance does not meet the standards, it is necessary to adjust the hyperparameters (such as C value, Gamma value) or use other kernel functions for retraining until the accuracy of the model reaches the expected target. After the trained intent classifier is obtained, the dialog unit text vector set is input into the already trained dialog text intent classifier. The classifier will output the intent category corresponding to each dialog unit to form a user intent classification data set, which helps to accurately identify the user's needs and behavioral intentions during the conversation process and provides a basis for subsequent conversation management and personalized services. ;

[0032] Furthermore, the step of obtaining the character optimization reply logic includes:

[0033] Based on the user intention classification data set, the defect of the persona reply logic is compared and identified to obtain the reply logic defect feature information; the robot reply effect target is defined, and the optimization strategy analysis of the reply logic defect feature information is performed based on the robot reply effect target to determine the reply logic optimization strategy; based on the reply logic optimization strategy and the user intention classification data set, the persona reply logic is optimized to obtain the initial optimized reply logic; the initial optimized reply logic is verified, evaluated and iteratively tuned to obtain the persona optimized reply logic.

[0034] Optionally, in order to obtain the persona-optimized reply logic, first, a comprehensive comparison and analysis of the existing persona reply logic is performed based on the user intent classification dataset. In this process, by analyzing historical conversation samples, especially comparing the actual conversation content in the user intent classification dataset with the answers generated by the robot, check whether the robot's reply accurately reflects the user's intention. The goal of this step is to analyze whether the robot correctly understands the user's needs and generates appropriate replies based on the user's intentions, so as to find out the defects of the current reply logic. This comparison process includes comparison of content accuracy, comparison of context loss, comparison of semantic understanding, comparison of answer format, etc. The comparison of content accuracy is to check whether the AIGC robot has given the correct content. For example, when a user asks about a vehicle failure, the answer may be irrelevant to the failure, resulting in inaccurate reply content. The comparison of context loss is whether the AIGC robot can remember and refer to the previous conversation content in multiple rounds of conversations. If the user mentioned the first The vehicle of the first fleet broke down, but the AIGC robot did not mention the first fleet in its subsequent reply, which means that the reply logic has defects in maintaining the context. The comparison of semantic understanding is whether the AIGC robot accurately understands the actual needs of the user. For example, if the user asks about the offline status of the vehicle, if the answer given by the robot is not related to the offline status, but introduces other information that is not related to the vehicle status, it is a semantic understanding defect. The comparison of the answer format is that even if the answer of the AIGC robot is correct in content, its expression (such as tone, length, clarity, etc.) may not meet the user's expectations. For example, the AIGC robot may provide an overly brief or vague answer, and the user cannot get the required information from it. Through the comparison of these aspects, the possible reply defect features of the existing human reply logic can be identified, such as irrelevance, incompleteness, oversimplification, semantic errors, etc., thereby forming reply logic defect feature information. These defect feature information provides key input for subsequent optimization strategy analysis;Subsequently, based on business needs, the tasks or service goals that the robot needs to complete are clarified, and the robot response effect goals, that is, the ideal state of the AIGC robot's response, are defined, which include multiple dimensions such as accuracy, coherence, relevance, and timeliness. After determining the robot response effect goals, the response logic defect feature information is analyzed based on these goals. This means that the answers of the AIGC robot in actual conversations need to be reviewed to analyze its defects in accuracy, coherence, relevance, and timeliness. Defects may come from inaccurate content, lack of context, overly vague or brief answers, delayed timeliness, etc. Based on these defects, corresponding optimization strategies will be defined to make these defects meet the robot response effect target requirements. For example, if the answers under the fault handling intention are always too brief, it can be determined that the optimization strategy is to introduce more domain knowledge of such problems and increase the detail and practicality of the answers. If there is no appropriate response timeliness under the offline time intention, the AIGC robot can interact with the real-time data source through the API interface to ensure the provision of the latest vehicle status, offline time and other information, to ensure that the AIGC robot can make a timely response in combination with real-time data. By sorting out these optimization strategies, the response can be determined. Logic optimization strategy: After the reply logic optimization strategy is determined, the existing persona reply logic is optimized based on the reply logic optimization strategy and the user intent classification data set. This process will use the target reply generator to integrate and optimize the persona reply logic of the AIGC robot, so that the AIGC robot can generate more expected answers under similar intents. After completing this process, the AIGC robot will generate an initial optimized reply logic. The optimized initial reply logic will be verified and evaluated to ensure that it can effectively improve the overall performance of the AIGC robot. During the verification and evaluation, the optimized reply logic will be applied to historical data to observe its performance under various intents and compared with the original reply logic to evaluate the degree of improvement in its accuracy, coherence, relevance, and timeliness. If the improvement meets expectations, the current initial optimized reply logic will be used as the persona optimized reply logic. Otherwise, the data generated by the verification process will be used for user intent identification and classification, and the same operation will be performed on the identified intent classification data set. This process will be iterated until the expected effect is achieved, so that the AIGC robot can provide accurate, coherent, and efficient answers under a variety of user intents, improving its understanding of user intent and the quality of its responses. ;

[0035] Furthermore, the obtaining of the initial optimization response logic includes:

[0036] Based on the reply logic optimization strategy, reply data extraction training is performed on the user intention classification dataset to obtain a reply discriminator; based on the reply logic optimization strategy, random noise generation training is performed on the user intention classification dataset to obtain a reply generator; a network balancing mechanism is constructed, and based on the network balancing mechanism, the reply discriminator and the reply generator are alternately iterated and trained to generate a target reply generator; based on the target reply generator, the personality reply logic is fused and optimized to obtain the initial optimized reply logic.

[0037] Optionally, after obtaining the reply logic optimization strategy, the reply logic optimization strategy is parsed to identify the intents for which the reply logic optimization strategy is optimized, and the reply data of these intents are extracted from the user intent classification data set to form the training data of the reply discriminator. For example, if one of the reply logic optimization strategies is to improve the problem of oversimplification under the fault handling intent, all the reply data about the fault handling intent are extracted from the user intent classification data set; then, a reply discriminator is constructed based on a deep learning model (such as a convolutional neural network, a recurrent neural network or a Transformer model, etc.), and the input of the reply discriminator is the reply data. The output is a binary classification label to determine whether the reply meets expectations. The reply data and the corresponding label are input into the reply discriminator, the predicted label is calculated through forward propagation, and the cross-entropy loss function is used to measure the error between the predicted label and the actual label. The back-propagation algorithm and the optimization algorithm (such as Adam or SGD) are used to optimize the parameters of the discriminator so that it can effectively distinguish between correct and incorrect replies. Through training, the reply discriminator will gradually adjust the weights until it can accurately judge the quality of the robot's reply. In addition, based on the reply logic optimization strategy and the user intent classification dataset, random noise is generated, that is, micro-noise is added to the data extracted from the user intent classification dataset. Small perturbations or random selection of some elements from certain rules (such as word replacement, grammatical transformation, etc.) are used to generate some noise data that is slightly different from the original conversation but still maintains semantic coherence. These noise data are then merged with the extracted data to form the training data for the reply generator. After that, a generator is built based on a deep learning model (such as a recurrent neural network, a long short-term memory network, a Transformer model, etc.). The structure of the generator usually includes an encoder and a decoder. The encoder is responsible for converting user input into a feature vector, and the decoder generates the final reply text. The training data generated by noise is used to train the generator, and each input feature is associated with a correct reply. Pairing, during the training process, the generator generates appropriate text replies by learning the relationship between user input and replies. The generator is optimized by maximizing the log likelihood, with the goal of making the generated replies as close to the target replies as possible. During training, the generator will update its weight parameters based on the difference between the input and output texts, making the generated text more accurate and natural. The generator's loss function is usually the cross-entropy loss, which measures the difference between the generated replies and the actual target replies. During the training process, the generator's goal is to minimize the loss function, thereby gradually improving the quality of the generated replies. After multiple rounds of training, the reply generator will be able to generate diverse and accurate replies based on user input;Then, a network balancing mechanism is constructed. The purpose of the network balancing mechanism is to ensure that the training between the reply discriminator and the reply generator can be balanced during the training process to avoid inconsistency between the two during training. Common strategies include adjusting the training frequency and loss function weights of the two, so that the reply discriminator and the reply generator can learn effectively at the same time. The reply data is generated by using the current reply generator and input into the reply discriminator for evaluation. The reply discriminator provides feedback based on the difference between the generated reply and the target reply, and updates the weight of the reply discriminator. The reply generator adjusts its generation strategy based on the feedback from the reply discriminator to generate reply data that better meets the expectations. The purpose of alternating iterative training is to allow the two to continuously optimize in the process of mutual game. The reply generator learns how to generate answers that better meet the needs of real users, while the reply discriminator learns how to accurately evaluate the quality of these replies. After alternating training, the reply generator will be able to generate more diverse and accurate replies, and the reply discriminator can effectively evaluate the quality of each generated reply. Finally, at the end of training, the reply generator will be optimized to the target reply generator, which can generate more diverse and accurate replies based on the input. The user intends to generate high-quality replies that meet the user's expectations; after obtaining the target reply generator, the existing persona reply logic is optimized. The goal of the optimization is to combine the diverse and accurate replies generated by the target reply generator with the existing persona reply logic to improve the overall reply quality. Specifically, the target reply generator provides diverse reply data, which are usually more diverse and natural, but may not fully conform to the specific persona reply logic. The persona reply logic can ensure that the reply style and tone are consistent with the user's needs and expectations. The optimization process is to combine the reply data generated by the target reply generator with the tone, style, knowledge base, etc. defined by the persona to adjust the role adaptability, language style, and answer range. The optimized reply logic will not only combine the diversity of the generator, but also ensure that the generated answers meet the standards of accuracy, contextual coherence, and timeliness. By integrating the capabilities of the target reply generator and the rules of the persona reply logic, the initial optimized reply logic is finally obtained. This optimized logic will provide a more intelligent, personalized, and effective conversation experience on the basis of ensuring that the user's intention is accurately understood. ;

[0038] Furthermore, the method further comprises:

[0039] Performing context feature analysis on the user intention classification data set to obtain a conversation intention context feature set; and performing context optimization modification on the initial optimization reply logic based on the conversation intention context feature set.

[0040] Optionally, based on the conversation information in the user intention classification dataset, extract the user input of each round and the reply data of the AIGC robot, and determine whether the user input of each round is related to the previous conversation, that is, determine whether the intent labels of adjacent rounds of conversations are the same. If they are the same, these rounds have a contextual relationship. At this time, extract the user input of these rounds, the reply data and intent labels of the AIGC robot, and use TF-IDF to construct a conversation intention context feature set; then, analyze the conversation intention context feature set to identify historical information, user intent changes and conversation rounds in the current conversation, and then modify the initial optimized reply logic based on these context features. The modification process includes integrating historical conversation content to ensure that the reply is consistent with the previous content, identifying changes in user intent, adjusting the answer content to meet current needs, and adjusting the depth and form of the reply according to the conversation round, so that the reply is more coherent, relevant and in line with the context of the conversation. Finally, the modified reply logic can generate more accurate, natural and user-expected answers based on the user input and conversation background.

[0041] Build a retrieval knowledge base, configure and orchestrate to obtain a target workflow, and build a multi-round dialogue process and page navigation logic based on the target workflow.

[0042] In one embodiment, a retrieval knowledge base is constructed by collecting text data related to the business, including user common questions, data content involved, etc. This retrieval knowledge base is to enable the AIGC robot to efficiently extract relevant content from a large amount of information and perform intelligent query and reply; at the same time, a target workflow is also configured. The workflow configuration is to define the interaction process between the AIGC robot and the user, ensuring that each link is carried out in a predetermined order. Through the arrangement of the workflow, different operation steps, user input requirements and robot response methods can be clarified. The workflow usually includes multiple nodes, each node represents an operation, such as obtaining a user account, finding a specified menu, generating MQTT message carrier content, etc.; based on the arrangement of the workflow, a multi-round dialogue process and page navigation logic are constructed. When constructing a multi-round dialogue process, the order and nodes of each round of dialogue will be designed. When constructing the page navigation logic, the workflow will define when the user needs to jump to different pages. By constructing a retrieval knowledge base, configuring workflows and page navigation logic, the AIGC robot can efficiently handle multi-round dialogues, provide real-time information query and task support, thereby improving the user experience and ensuring the smoothness and efficiency of the interaction process.

[0043] Furthermore, the target workflow usage interaction process includes: obtaining a user account, finding a specified menu, generating MQTT message carrier content, and executing an MQTT sending instruction.

[0044] Preferably, the target workflow's interactive process includes obtaining a user account, finding a specified menu, generating MQTT message carrier content, and executing MQTT sending instructions, wherein the step of obtaining a user account ensures that only authorized users can perform subsequent operations by verifying the user's identity information. This process usually involves obtaining the user's login credentials to ensure the legitimacy of the user's identity. The step of finding the specified menu locates the corresponding functional module or menu item according to the user's needs. At this stage, the workflow will determine the menu path corresponding to the operation requested by the user by searching the knowledge base or existing data structure. The step of generating MQTT message carrier content packages the user's requested data and commands into a specific format by constructing a message carrier that complies with the MQTT protocol. This message carrier will contain the information, instructions or status data that needs to be transmitted to ensure that it can be transmitted through the MQTT protocol. The step of executing MQTT sending instructions sends the previously generated message carrier through the MQTT protocol, and the instruction will be sent to a specific device for processing and response. Through this step, real-time interaction with other hardware devices or services can be carried out to complete the user's request.

[0045] The AIGC robot is configured and published based on the personality optimization reply logic, the multi-round dialogue process and the page navigation logic to obtain a target AIGC robot; a user input instruction is obtained, and the target AIGC robot interacts with the target user based on the user input instruction and the retrieval knowledge base.

[0046] In one embodiment, the personality optimization reply logic is combined with the multi-round dialogue process and page navigation logic to ensure that the AIGC robot can provide expected personalized replies based on the user's input, and ensure that the AIGC robot can smoothly guide the user to complete the operation during the entire interaction process. The personality optimization reply logic includes setting the robot role, tone style and answer range to ensure that the robot's performance in the dialogue meets the established design standards. The multi-round dialogue process and page navigation logic ensure that the robot can smoothly jump between different dialogue stages and menus, thereby providing a coherent user experience. After the configuration is completed, the target AIGC robot can be released according to these configurations and prepare to interact with the user. At this time, the target AIGC robot has the ability to provide information according to user needs and can execute instructions and navigate according to the established process; subsequently, the user input instruction is obtained, and the target AIGC robot will analyze and match the corresponding operation or query request according to the user's input instruction. The AIGC robot uses the previously constructed retrieval knowledge base to query relevant information and provide accurate replies, or guides the user to enter a specific function page according to the user's needs. In this way, the target AIGC robot completes information interaction with the target user and provides the service or solution required by the user.

[0047] In summary, the embodiments of the present application have at least the following technical effects:

[0048] First, log in to the platform and create an AIGC robot, obtain the robot design page, and create a personality response logic for the AIGC robot through the robot design page. The personality response logic includes model roles, language styles, and model answer ranges. Subsequently, the user dialogue data set is collected and obtained, and the personality response logic is optimized based on the user dialogue data set to obtain the personality optimization response logic. Further, a retrieval knowledge base is constructed, and the target workflow is configured and arranged at the same time. According to the target workflow, a multi-round dialogue process and page navigation logic are constructed. Afterwards, the AIGC robot is configured and published based on the personality optimization response logic and the multi-round dialogue process and page navigation logic to obtain the target AIGC robot. Finally, the user input instruction is obtained, and the target AIGC robot interacts with the target user based on the user input instruction and the retrieval knowledge base. It solves the technical problems of existing technologies in human-computer interaction, such as context understanding, delayed knowledge updating, and insufficient accuracy of generated content. By configuring multi-round dialogue processes and page navigation logic, and configuring and releasing the AIGC robot based on optimized reply logic, it achieves the technical effect of improving the response accuracy and real-time performance of the AIGC robot in human-computer interaction, thereby enhancing user experience and interaction efficiency.

[0049] Embodiment 2, based on the same inventive concept as the interactive method based on the AIGC robot in the above embodiment, this embodiment provides an electronic device, Figure 3 The schematic structural diagram of the electronic device provided in the second embodiment of the present invention shows a block diagram of an exemplary electronic device suitable for implementing the implementation mode of the present invention. Figure 3 The electronic device shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention. Figure 3 As shown, the electronic device includes a processor 11, a memory 12, an input device 13 and an output device 14; the number of processors 11 in the electronic device can be one or more. Figure 3 Taking a processor 11 as an example, the processor 11, memory 12, input device 13 and output device 14 in the electronic device can be connected through a bus or other means. Figure 3 The example of connecting through bus is taken in the following.

[0050] Embodiment 3, based on the same inventive concept as the interactive method based on AIGC robot in the above embodiment, this embodiment provides a computer-readable storage medium, which can be used to store software programs, computer executable programs and modules, such as program instructions / modules corresponding to the interactive method based on AIGC robot in the embodiment of the present application. The processor executes various functional applications and data processing of the computer device by running the software programs, instructions and modules stored in the memory, that is, realizing the above-mentioned interaction based on AIGC robot.

[0051] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0052] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0053] This specification and drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. An interactive method based on an AIGC robot, characterized in that: The method comprises: Log in to the platform and create an AIGC robot, obtain a robot design page, and create a personality response logic for the AIGC robot through the robot design page. The personality response logic includes a model role, a language style, and a model answer range; Collecting and acquiring a user conversation data set, optimizing the persona response logic based on the user conversation data set, and obtaining a persona optimized response logic; Build a search knowledge base, configure and arrange to obtain the target workflow, and build a multi-round dialogue process and page navigation logic based on the target workflow; Based on the personality optimization reply logic, the multi-round dialogue process and the page navigation logic, the AIGC robot is configured and published to obtain a target AIGC robot; Obtaining a user input instruction, the target AIGC robot interacting with the target user based on the user input instruction and the search knowledge base; The step of obtaining the character optimization reply logic includes: Performing text segmentation on the user dialogue data set to obtain a user dialogue unit set; Performing text processing on the user dialogue unit set using TF-IDF to obtain a dialogue unit text vector set; Performing user intent recognition and classification on the dialogue unit text vector set to obtain a user intent classification data set; Optimizing and verifying the persona response logic based on the user intent classification data set to obtain a persona optimized response logic; The step of obtaining the character optimization reply logic includes: Based on the user intention classification data set, the character response logic is compared and identified for defects, to obtain response logic defect feature information; Defining a robot reply effect target, performing an optimization strategy analysis on the reply logic defect feature information based on the robot reply effect target, and determining a reply logic optimization strategy; Based on the reply logic optimization strategy and the user intent classification data set, the persona reply logic is optimized to obtain an initial optimized reply logic; Verifying, evaluating and iteratively tuning the initial optimized response logic to obtain the personality optimized response logic; The obtaining of the initial optimization response logic includes: Based on the reply logic optimization strategy, reply data extraction training is performed on the user intention classification data set to obtain a reply discriminator; Based on the reply logic optimization strategy, random noise generation training is performed on the user intention classification data set to obtain a reply generator; Constructing a network balancing mechanism, and performing alternating iterative training on the reply discriminator and the reply generator based on the network balancing mechanism to generate a target reply generator; Based on the target reply generator, the character reply logic is integrated and optimized to obtain the initial optimized reply logic; The method comprises: Performing context feature analysis on the user intention classification data set to obtain a conversation intention context feature set; The initial optimized reply logic is contextually optimized and modified based on the conversation intention context feature set.

2. The interactive method based on the AIGC robot according to claim 1, characterized in that: The obtaining of the dialogue unit text vector set includes: Performing word segmentation processing on each dialogue unit in the user dialogue unit set to obtain a dialogue unit text word segmentation set; Obtaining a text cleaning program, wherein the text cleaning program includes stop word removal, special symbol removal, punctuation mark removal, and text format unification; Performing text cleaning on the dialogue unit text segmentation set based on the text cleaning program to obtain a usable dialogue unit text segmentation set; The TF-IDF value is calculated and the text vector is converted to the available dialogue unit text segmentation set using TF-IDF to obtain the dialogue unit text vector set.

3. The interactive method based on AIGC robot as claimed in claim 1, characterized in that: The obtaining of the user intention classification dataset includes: Acquire a historical user conversation data set, and preprocess the historical user conversation data set to obtain a historical conversation text vector set; Constructing an intent category label library, classifying and labeling the historical conversation text vector set based on the intent category label library, and obtaining a historical conversation text intent sample set; Using a support vector machine to perform classification training on the historical conversation text intent sample set to obtain a conversation text intent classifier; The user intent classification data set is obtained by performing user intent recognition and classification on the dialogue unit text vector set based on the dialogue text intent classifier.

4. The interactive method based on AIGC robot according to claim 1, characterized in that: The target workflow usage interaction process includes: obtaining a user account, finding a specified menu, generating MQTT message carrier content, and executing an MQTT sending instruction.

5. An electronic device, characterized in that: The electronic device comprises: a processor; and a memory for storing instructions executable by the processor; wherein the processor is used to execute an interaction method based on an AIGC robot as described in any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, so the computer program is used to execute the interaction method based on the AIGC robot as described in any one of claims 1 to 4.

Images