Remote video helper scheduling method and system in combination with user portrait

Abstract

The invention relates to the technical field of data processing, and provides a remote video helper scheduling method and system in combination with a user portrait. The method comprises the following steps: collecting target user information according to a preset portrait index to obtain a user portrait; performing prototype matching based on the user portrait and the to-be-processed item, and obtaining item prediction duration; obtaining a real-time video of each window and a queue to be processed from the video helper platform; traversing the video recognition progress, and predicting the entry duration of each window; and performing item scheduling in combination with the predicted duration, the queue information and the item settlement duration, and generating an updated to-be-processed queue. The technical problems of matter processing queue congestion and low processing efficiency caused by lack of accurate analysis and effective data processing based on user data difference in existing remote video helper scheduling are solved, the overall service efficiency of a remote video helper platform is improved, the user waiting time is shortened, and the user experience is improved. And the user experience is improved.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and specifically to a remote video assistance matter scheduling method and system combined with user portraits. Background Art

[0002] With the rapid development of digital government services, remote video assistance has gradually become an important way to improve the efficiency of government services and enhance the experience of the masses in handling affairs. However, the current remote video assistance scheduling methods are mostly based on manual allocation or simple polling mechanisms, and do not consider the personalized needs, handling preferences, and historical behaviors of users. There are problems such as low service resource matching efficiency, uneven utilization rate of assistance personnel, and insufficient user satisfaction. User portrait technology can build a precise personalized service model based on user historical data and behavior preferences, but in the existing technology, there are few deep integrations of user portraits with remote video assistance matter scheduling methods. Therefore, there is an urgent need for a remote video assistance matter scheduling method combined with user portraits to achieve precise resource matching, efficient task allocation, and improve the overall quality of remote video assistance services and user satisfaction. Summary of the Invention

[0003] This application provides a remote video assistance matter scheduling method and system combined with user portraits, aiming to solve the technical problems that the existing remote video assistance matter scheduling lacks precise analysis and effective data processing based on user data differences, resulting in congestion in the matter processing queue and low processing efficiency. It realizes precise scheduling based on both user portraits and matter progress, effectively optimizes the data processing process, improves the overall service efficiency of the remote video assistance platform, shortens the user waiting time, and improves the technical effect of the user experience.

[0004] In the first aspect disclosed in this application, a remote video assistance matter scheduling method combined with user portraits is provided. The method includes: collecting information on Q target users according to a preset portrait index set to obtain Q target user portrait index sets, where Q is a positive integer; performing matter processing prototype matching analysis based on the target user portrait index sets and the Q to-be-processed matters of the Q target users to obtain Q predicted processing durations of the to-be-processed matters; interacting with L processing windows of the remote video assistance platform to obtain L real-time processed matter video information and L window to-be-processed queues, where L is a positive integer; traversing the L real-time processed matter video information for progress identification to determine L real-time predicted completion durations; scheduling the Q to-be-processed matters based on the Q predicted processing durations of the to-be-processed matters, the L window to-be-processed queues, and the L real-time predicted completion durations to obtain L updated to-be-processed queues.

[0005] Another aspect disclosed in this application provides a remote video errand scheduling system combined with user portraits. The system includes: an information collection module: collecting information on Q target users according to a preset portrait index set to obtain Q target user portrait index sets, where Q is a positive integer; a matching and analysis module: performing prototype matching analysis of matter processing for matter processing based on the target user portrait index sets and Q pending matters of Q target users to obtain Q predicted processing durations for the pending matters; an errand platform interaction module: interacting with L processing windows of the remote video errand platform to obtain L real-time processed matter video information and L window pending queues, where L is a positive integer; a progress recognition module: traversing the L real-time processed matter video information for progress recognition to determine L real-time predicted completion durations; a matter scheduling module: scheduling the Q pending matters based on the Q predicted processing durations for the pending matters, the L window pending queues, and the L real-time predicted completion durations to obtain L updated pending queues after update.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: The above-mentioned remote video errand scheduling method combined with user portraits first collects information on multiple target users according to the set user portrait indexes to construct portrait data for each user; subsequently, combines the user portraits with their corresponding pending matters, and predicts the approximate processing time required for each matter by matching with a pre-established matter processing model; then, obtains the current video processing status and pending queue information of all service windows from the remote video errand platform; recognizes the progress of matter processing for each window by analyzing the voice text content in the video, and estimates the expected completion time of each window; finally, comprehensively considers the predicted processing duration of user matters, the current queuing situation of each window, and the expected completion time, and performs intelligent scheduling and allocation of all pending matters to optimize the task queue of each window and achieve a more efficient and reasonable service arrangement.

[0007] The above description is only an overview of the technical solutions of this application. In order to be able to understand the technical means of this application more clearly, it can be implemented according to the content of the description. And in order to make the above and other purposes, features, and advantages of this application more obvious and understandable, the following specifically enumerates the specific embodiments of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0009] Figure 1 It is a schematic flowchart of a remote video errand scheduling method combined with user portraits in an embodiment.

[0010] Figure 2 It is an architecture diagram of a remote video errand scheduling system combined with user portraits in an embodiment.

[0011] Explanation of reference numerals: Information collection module 11, matching and analysis module 12, errand platform interaction module 13, progress identification module 14, errand scheduling module 15. Detailed implementation manners

[0012] In the embodiments of the present application, by providing a remote video errand scheduling method and system combined with user portraits, the technical problem that the existing remote video errand scheduling lacks accurate analysis and effective data processing based on user data differences, resulting in congestion in the errand processing queue and low processing efficiency, is solved. The accurate scheduling is realized according to the dual bases of user portraits and errand progress, the data processing process is effectively optimized, the overall service efficiency of the remote video errand platform is improved, the user waiting time is shortened, and the technical effect of improving the user experience is achieved.

[0013] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0014] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products or devices.

[0015] Embodiment 1, as Figure 1 shown, the present application provides a remote video errand scheduling method combined with user portraits, and the method includes: Collect information on Q target users according to a preset portrait index set to obtain Q target user portrait index sets, where Q is a positive integer.

[0016] In the embodiments of the present application, a set of portrait metrics for describing user characteristics is predefined as a preset portrait metric set, including user attribute characteristics, user demand characteristics, and user historical behavior characteristics. Subsequently, information collection is carried out for Q target users respectively, and the metric information corresponding to each user is obtained through methods such as questionnaire surveys, data scraping, or user interaction records, forming a data set that can reflect the personalized characteristics of each user, that is, Q target user portrait metric sets, where Q represents the total number of target users and is a positive integer. Exemplarily, the portrait of user A: {Age: 35, Education level: Bachelor's degree, Business type: Tax consulting, Historical average consultation duration: 8 minutes}; the portrait of user B: {Age: 62, Education level: High school, Business type: Pension query, Historical average consultation duration: 15 minutes}. This step provides a data basis for subsequent intelligent scheduling. Through quantified user characteristics, service demands can be predicted more accurately.

[0017] Furthermore, the present application provides that the preset portrait metric set includes user attribute metrics, user demand characteristic metrics, and user historical behavior characteristic metrics.

[0018] Preferably, the preset portrait metric set includes user attribute metrics, user demand characteristic metrics, and user historical behavior characteristic metrics. Among them, the user basic attribute metrics are mainly used to describe the personal information and background of the user, helping the platform understand their basic characteristics, including age, gender, geographical location (the geographical location where the user is located), occupation type (the type of work of the user, such as enterprise or institution staff, self-employed, freelancer, etc.), education level (such as high school, junior college, bachelor's degree, postgraduate, etc.), etc.; the user service demand characteristic metrics help identify and describe the specific needs of the user for services, including demand type (the type of service required by the user, such as business consultation, complaint feedback, operation guidance, government affairs handling, etc.), urgency of the matter (the urgency of the matter to be processed by the user, usually divided into general, urgent, very urgent), complexity of the matter (divided into simple, medium, complex according to the time and resources required for handling the matter), matter category label (the clear classification to which the matter belongs, which may include categories such as medical insurance, social security, tax, household register, employment, transportation, etc.); the user historical behavior characteristic metrics help the platform predict the user's future demands and handling behaviors by analyzing the user's past behaviors, including the number of historical assistance times, historical average handling duration, historical satisfaction score, historical matter feedback records, special service demand preferences (such as preferences for communication languages, preferences for the gender of handling personnel), etc. These metrics help to accurately understand the characteristics of each user, thus providing an accurate basis for prediction and scheduling for subsequent matter handling.

[0019] Based on the target user portrait metric sets and the Q to-be-processed matters of the Q target users, matter processing prototype matching analysis is carried out to obtain the predicted durations of the Q to-be-processed matters.

[0020] In one embodiment, by using the set of target user portrait metrics collected previously and combining with the to-be-processed matters corresponding to each user, an analysis basis for the characteristics of matter processing is established. Subsequently, through a pre-constructed matter processing prototype library (i.e., a database that classifies and extracts characteristics from a large number of historical completed matters to form a database of typical matter processing models), according to the current portrait metrics and to-be-processed matter characteristics of each user, a similarity analysis and matching are performed with the existing models in the prototype library. For example, for multiple dimensions such as the user's age, region, demand category, urgency level, and historical behavior data, the matching similarity degree with the corresponding models in the prototype library is calculated, and then the best-matched prototype model is selected. Finally, according to the matched matter processing prototype, the Q target user portrait metrics and the Q to-be-processed matters are respectively input into the matter processing prototype to predict the processing duration, and the specific duration that each target user's current to-be-processed matter may take is determined, that is, the predicted durations of the Q to-be-processed matters. This prediction result will be used as an important basis for subsequent intelligent scheduling to ensure a better match between the processing capacity and user needs and improve the resource utilization efficiency.

[0021] Furthermore, the present application provides a matter processing prototype matching analysis based on the set of target user portrait metrics and the Q to-be-processed matters of Q target users to obtain the predicted durations of the Q to-be-processed matters, including: Pre-construct a matter processing prototype matching library; perform a similarity matching on the matter processing prototype matching library based on the Q to-be-processed matters and the set of target user portrait metrics, and obtain the predicted durations of the Q to-be-processed matters according to the matching results.

[0022] Preferably, through historical data collection and analysis, a matter processing prototype matching library is pre-constructed. This matter processing prototype matching library includes multiple matter processing prototypes, and each matter processing prototype corresponds to a type of agency service matter. All of them are obtained through training with the historical data of this type of agency service matter. All the prototypes in the matter processing prototype matching library can predict the time requirements during the processing of different matter types and different user characteristics, providing a tool for subsequent prediction. Subsequently, each matter to be processed and the target user profile are encoded, and the extracted various features are converted into a computable vector form. The features of the user profile and the matter to be processed can usually be converted into feature vectors through numerical processing (such as normalization, one-hot encoding). For example, age, gender, matter urgency, etc. can be mapped to different numerical values respectively, thus constituting Q target feature vectors. Then, the cosine similarity is used to calculate the cosine value of the angle between the Q target feature vectors and the calibration vectors of each prototype, measuring the similarity in the direction of the two vectors. The closer the value is to 1, the more similar it is, and the closer the value is to 0, the less similar it is. After the calculation is completed, the matter processing prototype with the highest similarity is selected for each matter to be processed, and then the processing duration of the matter to be processed is predicted according to the selected matter processing prototype. Each matter processing prototype quantifies the processing time of each matter to be processed according to the learned mapping relationship, that is, the predicted duration of the matter to be processed. This process ensures the accuracy of the predicted duration and can obtain a more reasonable time prediction according to the characteristics of each user and the specific situation of the matter, providing a scientific basis for subsequent scheduling decisions.

[0023] Furthermore, the present application provides a pre-constructed matter processing prototype matching library, including: Obtain a set of historical agency service matter logs; classify the set of historical agency service matter logs according to the similarity degree of matter types and the similarity degree of user profile indicators to obtain M classified sets of historical agency service matter logs, where M is a positive integer greater than or equal to Q; construct a matter processing prototype matching library based on the M classified sets of historical agency service matter logs.

[0024] Optionally, collect and organize a large number of historical assistant task logs, which record detailed information about user requests, including user categories, types of tasks to be done, task descriptions, processing durations, etc. The historical logs provide rich association information between users and tasks as sample data and serve as the basic data for constructing subsequent task processing prototypes. Subsequently, in the same way as described above, each set of data in the historical assistant task log set is converted into a sample feature vector. Then, by plotting the sum of squared errors (SSE) images corresponding to different K values and selecting the K value with a relatively small change in SSE, i.e., the K value at the elbow. After determining the K value, randomly select K sample feature vectors as the initial clustering center points (centroids), and assign each data point (sample feature vector) to the cluster with the nearest clustering center point through the Euclidean distance. Then, calculate the mean of all samples in each cluster and update the position of the clustering center point. Repeat the above process until the clustering center no longer changes or reaches the preset maximum number of iterations. After the clustering algorithm is executed, each historical task log will be assigned to a certain cluster, obtaining M classified historical assistant task log sets (each cluster represents a type of task). For example, all logs related to medical insurance matters can be grouped into one category, and all logs related to tax matters can be grouped into another category. Each classified historical assistant task log set contains logs with similar task types and similar user portraits. Here, M is a positive integer greater than or equal to Q, representing the total number of classified sets. After that, based on the classified historical assistant task log sets, further construct a task processing prototype matching library, that is, use each classified historical assistant task log set to train a task processing prototype. In this way, each task processing prototype can predict the processing duration of a category of assistant tasks. Through the above process, the most suitable processing prototype can be matched for each task to be processed, thereby predicting the processing duration of the task, optimizing scheduling and resource allocation, improving the overall service efficiency of the remote video assistant platform, and shortening the user waiting time.

[0025] Furthermore, this application includes: Construct a two-way classification function, where the two-way classification function is: ; where is the two-way classification similarity, is the task type of the j-th classified historical assistant task log in the i-th classified historical assistant task log set, is the set of user portrait index of the j-th classified historical assistant task log in the i-th classified historical assistant task log set, is the union of task types in the i-th classified historical assistant task log set, is the mean of the set of user portrait index in the i-th classified historical assistant task log set, To balance the weights of the similarity of matter types and the similarity of user profile metrics, is the total number of classified historical assistance matter logs in the i-th classified historical assistance matter log set, and M is the total number of classified historical assistance matter log sets; the M classified historical assistance matter log sets are classified and authenticated using the dual classification function. If the output dual classification similarity is greater than or equal to the preset similarity, the authentication passes.

[0026] Optionally, after obtaining the M classified historical assistance matter log sets, it is necessary to recognize these classified historical assistance matter log sets and determine whether the data in each set is highly consistent. In this process, a dual classification function will be constructed to measure the similarity between each historical assistance matter log in the classified historical assistance matter log set and other logs. The dual classification function is as follows: ; Among them, is the dual classification similarity, used to measure the similarity of different historical matter logs; is the matter type of the j-th classified historical assistance matter log in the i-th classified historical assistance matter log set; is the user profile metric set of the j-th classified historical assistance matter log in the i-th classified historical assistance matter log set; is the union of matter types in the i-th classified historical assistance matter log set; is the mean of the user profile metric sets in the i-th classified historical assistance matter log set; is a weight coefficient, used to balance the influence of matter type similarity and user profile similarity on the total similarity; Let \(N_i\) be the total number of classified historical assistance task logs in the \(i\)-th set of classified historical assistance task logs, and \(M\) be the total number of sets of classified historical assistance task logs. Through the above function, the two-way classification similarity of each historical assistance task log is calculated. The two-way classification similarity reflects the similarity between this task log and other logs, and is used to measure whether the data in the set of classified historical assistance task logs is consistent. For each set of classified historical assistance task logs, the two-way classification similarity between each group of data in the set and other data is calculated through the two-way classification function. If the two-way classification similarity of each group of data in a set is greater than or equal to the preset similarity, it means that the data in this set is highly consistent and belongs to the same category; if there is data in a set with a similarity less than the preset similarity, it means that there is misclassified data in this set. At this time, this data will be removed, and the remaining data in the set will be re-authenticated until all the data in the set passes the authentication. For the data removed from the set, a new set is combined and authenticated until all the data has a corresponding set. Through the design and calculation of the two-way classification function, combined with the similarity of the matter type and user portrait indicators, the above process enables each set of classified historical assistance task logs to accurately represent a category, thus laying a foundation for the construction of the subsequent matter processing prototype, effectively optimizing the data processing flow, and improving the overall service efficiency of the remote video assistance platform.

[0027] Further, this application includes: Taking the \(M\) sets of classified historical assistance task logs as training data respectively, constructing \(M\) matter processing prototypes; summarizing the \(M\) matter processing prototypes to obtain the matter processing prototype matching library.

[0028] Optionally, after the M classified historical to-do item log sets pass the authentication, each classified historical to-do item log set is partitioned to obtain M training data and validation data. Both types of data include historical user profile metrics, historical to-do items, and historical processing durations. Subsequently, use neural networks, random forests, decision trees, etc. to construct M initial item processing prototypes. Taking the multi-layer perceptron (MLP) in the neural network as an example, the initial item processing prototype constructed based on the MLP includes an input layer, multiple hidden layers, and an output layer. Then, use a random initialization method (such as Xavier initialization) to assign initial values to the weights of each layer in the prototype. These initialized weights will be used to start training the prototype. In this process, the bias terms also need to be initialized. After that, input the training set into the prototype for forward propagation, calculate the predicted result of the processing duration through passing through each layer, then use the mean squared error loss function to calculate the loss between the predicted result and the historical processing duration in the training data, calculate the gradient of the loss with respect to the weights of each layer layer by layer through the backpropagation algorithm, and use the Adam optimizer to optimize the prototype parameters, adjusting the weights to minimize the value of the loss function. Repeat the above process until the maximum number of iterations is reached or the loss no longer decreases significantly. After the training is completed, use the validation data to test the prototype and evaluate the accuracy of the prototype in the processing duration prediction task. If the performance of the prototype on the validation data meets the expectations, then output the current initial item processing prototype as the final item processing prototype; if the validation result is not ideal, then adjust the learning rate, hidden layer structure, or other hyperparameters to further improve the prototype performance and ensure that the final processing prototype can accurately predict the processing duration. After all prototypes are trained, summarize the processing prototypes trained based on the M classified historical to-do item log sets into an overall item processing prototype matching library, and use the item type with the highest frequency of occurrence and the mean value of the user profile metrics in the classified historical to-do item log set used by each prototype as the calibration data to calibrate the prototypes in the item processing prototype matching library, thereby constructing the final item processing prototype matching library. This prototype matching library contains M different item processing prototypes, and each prototype represents a typical item processing pattern in a classified historical to-do item log set, which can be used for subsequent prediction of the processing duration of to-be-processed items, thereby optimizing task scheduling and resource allocation.

[0029] Obtain L real-time processing item video information and L window to-be-processed queues from the L processing windows of the interactive remote video help platform, where L is a positive integer.

[0030] In one embodiment, L processing windows (where L is an integer) in the remote video assistance platform are interacted through a communication interface, and video information of processing matters and to-be-processed queue data from each window are received, forming L pieces of real-time processing matter video information and L window to-be-processed queues. Each processing window represents an independent task processing unit, and the situation of to-be-processed matters is displayed in real time. In the to-be-processed queue of each processing window, there is a set of matters to be processed, and each matter corresponds to a predicted duration, that is, the time required to process this matter is predicted based on historical data and user profiles. As new matters enter the queue, the processing duration of this matter is predicted and added to the to-be-processed queue of the corresponding window. In this way, the queue situation of each window can be updated in real time, and tasks can be reasonably allocated according to the predicted duration, ensuring that the workload of each window is balanced and the overall processing efficiency is improved.

[0031] Traverse the L pieces of real-time processing matter video information for progress recognition to determine L real-time predicted completion durations.

[0032] In one embodiment, after obtaining L pieces of real-time processing matter video information by interacting with the remote video assistance platform, these real-time processing matter video information will be traversed. For each piece of real-time processing matter video information traversed, by performing speech-to-text conversion and keyword extraction to analyze this real-time processing matter video information of each window, the progress status of each matter is recognized in real time, and the real-time predicted completion duration of each matter is calculated. This duration represents the remaining time required to complete the processing of this matter in the current window. Through such progress recognition and prediction, the processing queue can be dynamically adjusted, task scheduling can be optimized, and the overall work efficiency and user experience can be improved.

[0033] Furthermore, the present application provides traversing the L pieces of real-time processing matter video information for progress recognition to determine L real-time predicted completion durations, including: Perform speech-to-text conversion on the L pieces of real-time processing matter video information to obtain L pieces of real-time speech text; extract L sets of historical speech text of the L processing windows, and perform keyword iterative analysis in combination with the L pieces of real-time speech text to determine L iterative progress features; perform progress recognition based on the L iterative progress features to obtain the L real-time predicted completion durations.

[0034] Preferably, the real-time processing matter video information of L processing windows is subjected to speech-text conversion. This process analyzes the speech in the audio through speech recognition technologies (such as Google Speech-to-Text, Microsoft Azure Speech API, Kaldi, etc.), and recognizes each word and sentence in the speech through speech segmentation, noise suppression, and text conversion, so as to obtain L real-time speech texts, which contain the content of the communication between the user and the staff during the processing of the matter. Subsequently, the historical speech text sets of each processing window are extracted, and these historical texts record the video speech data of the matters processed by each window before. By combining the current real-time speech text with the historical speech text sets, keyword iterative analysis is carried out. The process of iterative analysis is to extract relevant keywords from the real-time speech text and compare them with the keywords in the historical speech text to find out the keyword patterns related to the current matter progress. These keywords can include the content of the user's questions, the answers of the staff, the description of the matter progress, etc. Based on these iteratively obtained progress features, the current processing progress of each matter can be identified, and these progress features help to judge the remaining processing time of the matter. By sequentially inputting the obtained L iterative progress features into the progress recognition prototype for analysis, the closing duration of each real-time matter can be accurately predicted, that is, the real-time predicted closing duration of each processing window. Among them, the progress recognition prototype is constructed by the same training method as described above according to the historical progress features and historical closing durations of the historical speech text sets. Finally, these predicted closing durations will be used for task scheduling and resource allocation to ensure a reasonable load for each window, improve the processing efficiency of the platform, and enhance the user experience.

[0035] Furthermore, the present application provides extracting L historical speech text sets of L processing windows, combining the L real-time speech texts for keyword iterative analysis, and determining L iterative progress features, including: Extract the L first historical speech texts from the L historical speech text sets in chronological order from the front to the back; perform keyword matching on the L first historical speech texts to construct L first memory vectors; extract the L second historical speech texts from the L historical speech text sets again in chronological order from the front to the back, and perform keyword similarity matching on the L second historical speech texts based on the L first memory vectors respectively to obtain L second historical matching progress keyword sets to update the L first memory vectors and obtain L second memory vectors; perform iterative matching on the remaining historical speech texts in the L historical speech text sets based on the L second memory vectors to obtain L historical memory vectors; perform keyword iterative analysis on the L real-time speech texts based on the L historical memory vectors to obtain L real-time memory vectors; use the progress feature recognition network layer to analyze the L real-time memory vectors to obtain the L iterative progress features.

[0036] Optionally, according to the time sequence and the preset time window, according to the timestamps in the L historical voice text sets, L first historical voice texts are extracted, and these texts represent the voice communication content of each processing window in the historical data at different time points. Subsequently, these first historical voice texts are matched with keywords through the preset progress keyword library to construct L first memory vectors, which contain important feature information of the corresponding historical voice texts and indicate the position of the historical text in the processing progress. Subsequently, the second historical voice text is extracted again in the time sequence, and this time the voice data following the first historical text is extracted. For each window, the first memory vector is used to perform keyword similarity matching with the preset progress keyword library in the second historical voice text, and the progress keywords in the second historical voice text are determined to obtain L second historical matching progress keyword sets. Afterwards, the first memory vector is updated based on these second historical matching progress keyword sets, that is, these keywords are spliced ​​to the back of the corresponding first memory vector to obtain L second memory vectors, which contain more context information than the first memory vector and can more accurately represent the current progress state. Then, the remaining historical voice texts are iteratively matched based on the second memory vector. This process is gradually performed on all the remaining texts in the historical voice text set to obtain L complete historical memory vectors. These historical memory vectors are used to represent the progress status of each historical voice text and help understand the progress changes in the entire matter processing process. Based on these historical memory vectors, the real-time voice text is iteratively analyzed for keywords, that is, the keywords recorded in these historical memory vectors are compared with the keywords in the L real-time voice texts, and the keywords with them are extracted from the L real-time voice texts and spliced ​​into L real-time memory vectors. These real-time memory vectors represent the status of the current matter progress. Finally, the generated real-time memory vectors are further analyzed using the progress feature recognition network layer. The progress feature recognition network layer uses a deep learning algorithm (such as a fully connected neural network, which is constructed in the same way as above and is also iteratively performed through steps such as forward propagation, loss calculation, back propagation, and parameter optimization) to identify the progress features of the task based on the information in the memory vector and obtain iterative progress features. These progress features provide a basis for predicting the completion time and task scheduling of the current matter. Through this series of steps, we can not only accurately capture the progress patterns in historical data, but also analyze and predict the progress of matters in each processing window in real time, thereby optimizing task scheduling and resource allocation, shortening user waiting time, and improving user experience.

[0037] Furthermore, the present application provides matching the L first historical voice texts according to a preset progress keyword library to obtain L first historical matching progress keyword sets, adding the preset progress keyword library and the L first historical matching progress keyword sets into an initially empty vector to obtain L first memory vectors.

[0038] Optionally, a progress keyword library is linked. This progress keyword library is pre-built and includes a set of keywords related to task progress. This library is usually composed of progress keywords defined by historical data analysis, domain knowledge or experts, and covers iconic words or phrases at different task processing stages, such as start, process, wait, complete, etc. These keywords play a guiding role in task progress identification. Subsequently, keywords are extracted from each historical voice text through natural language processing (NLP) technology. These keywords include specific task information, event status descriptions or indicative words of progress changes. The extracted keywords are then matched with keywords in the preset progress keyword library. The matching method can use text similarity calculation methods such as cosine similarity, TF-IDF, Word2Vec, etc. to find progress keywords related to the historical voice text. For example, if the historical voice text contains words such as start process and process, these words will match the preset start and process keywords. For each first historical voice text, a first matching progress keyword set is generated, which includes keywords related to progress in the text. Afterwards, an initially empty vector is created to store the relevant features of the keywords. By merging the keywords in the preset progress keyword library with the keywords in the L first historical matching progress keyword sets, each keyword occupies an independent dimension in the vector, and the matched keywords correspond to 1 and the unmatched keywords correspond to 0, thereby constructing L first memory vectors. These first memory vectors represent the progress feature information in the corresponding historical voice text, which can provide a data basis for subsequent progress identification and task scheduling, thereby optimizing the data processing flow.

[0039] The Q pending items are scheduled based on the predicted durations of the Q pending items, the L window pending queues and the L real-time predicted completion durations to obtain updated L updated pending queues.

[0040] In one embodiment, in combination with the predicted duration of each matter to be processed, the current situation of the to-be-processed queues of each processing window, and with reference to the real-time predicted completion duration of each window, the matters to be processed are reasonably allocated through a scheduling algorithm. Specifically, the priority of each matter to be processed is evaluated according to its predicted duration, and the matters to be processed are sorted according to the shortest task first strategy. Subsequently, the sorted matters to be processed are traversed in sequence, and the window with the earliest real-time predicted completion duration is selected for task allocation. Each matter to be processed will be allocated to a window capable of processing the matter, and it is ensured that the waiting time of the target user does not exceed the waiting time tolerance limit during allocation, so as to ensure that each task can be allocated to the most suitable window and maximize the overall processing efficiency. After determining the scheduling plan for the matters to be processed, the to-be-processed queues of the L windows are updated according to this scheduling plan, that is, the matters to be processed that need to be added to the queue are added, thereby forming L updated to-be-processed queues. Through this scheduling strategy based on the real-time predicted completion duration and the predicted duration of the matters to be processed, tasks can be efficiently allocated, avoiding the situation of overloading or idling of any window, thus optimizing the overall task processing process and improving work efficiency.

[0041] In summary, the embodiments of the present application at least have the following technical effects: The embodiments of the present application collect information on Q target users according to a preset portrait index set, obtaining Q target user portrait index sets, where Q is a positive integer; based on the target user portrait index sets and the Q matters to be processed of the Q target users, a matter processing prototype matching analysis is performed to obtain the predicted durations of the Q matters to be processed; L processing windows of the interactive remote video help platform are obtained, obtaining L real-time processing matter video information and L window to-be-processed queues, where L is a positive integer; the L real-time processing matter video information is traversed for progress recognition to determine the L real-time predicted completion durations; based on the predicted durations of the Q matters to be processed, the L window to-be-processed queues, and the L real-time predicted completion durations, the Q matters to be processed are scheduled to obtain L updated to-be-processed queues after update. These technical effects together solve the technical problem that the existing remote video help matter scheduling lacks accurate analysis and effective data processing based on user data differences, resulting in congestion in the matter processing queue and low processing efficiency, and achieve the technical effects of accurate scheduling according to the dual bases of user portraits and matter progress, effectively optimizing the data processing process, improving the overall service efficiency of the remote video help platform, shortening the user waiting duration, and improving the user experience.

[0042] Embodiment 2, based on the same inventive concept as the remote video help matter scheduling method combining user portraits in the foregoing embodiment, as Figure 2As shown in the figure, the present application provides a remote video assistance matter scheduling system combined with user portraits. The system includes: Information collection module 11: Collect information on Q target users according to a preset portrait index set, and obtain Q target user portrait index sets, where Q is a positive integer; Matching analysis module 12: Based on the target user portrait index sets and the Q pending matters of the Q target users, perform matter processing prototype matching analysis to obtain Q predicted durations of pending matters; Assistance platform interaction module 13; Interact with L processing windows of the remote video assistance platform to obtain L real-time processing matter video information and L window pending queues, where L is a positive integer; Progress identification module 14: Traverse the L real-time processing matter video information for progress identification to determine L real-time predicted completion durations; Matter scheduling module 15: Based on the Q predicted durations of pending matters, the L window pending queues, and the L real-time predicted completion durations, schedule the Q pending matters to obtain L updated pending queues after update.

[0043] Further, the information collection module 11 is further configured to execute the following method: The preset portrait index set includes user attribute indexes, user demand feature indexes, and user historical behavior feature indexes.

[0044] Further, the matching analysis module 12 is further configured to execute the following method: Pre-construct a matter processing prototype matching library; Based on the Q pending matters and the target user portrait index sets, perform similarity matching on the matter processing prototype matching library, and obtain the Q predicted durations of pending matters according to the matching results.

[0045] Further, the matching analysis module 12 is further configured to execute the following method: Obtain a set of historical assistance matter logs; Classify the set of historical assistance matter logs according to the similarity degree of matter types and the similarity degree of user portrait indexes to obtain M classified historical assistance matter log sets, where M is a positive integer greater than or equal to Q; Based on the M classified historical assistance matter log sets, construct a matter processing prototype matching library.

[0046] Further, the matching analysis module 12 is further configured to execute the following method: Construct a two-item classification function, where the two-item classification function is: ; Where is the two-item classification similarity, is the matter type of the jth classified historical assistance matter log in the ith classified historical assistance matter log set, is the user portrait index set of the j-th classified historical assistance item log in the i-th classified historical assistance item log set, is the union of the item types in the i-th classified historical assistance item log set, is the mean value of the user portrait index set in the i-th classified historical assistance item log set, is the weight for balancing the similarity of item types and the similarity of user portrait indexes, is the total number of classified historical assistance item logs in the i-th classified historical assistance item log set, and M is the total number of classified historical assistance item log sets; the M classified historical assistance item log sets are classified and authenticated using the dual-item classification function, and if the output dual-item classification similarity is greater than or equal to the preset similarity, the authentication passes.

[0047] Furthermore, the matching analysis module 12 is further configured to execute the following method: Respectively use the M classified historical assistance item log sets as training data to construct M item processing prototypes; summarize the M item processing prototypes to obtain the item processing prototype matching library.

[0048] Furthermore, the progress recognition module 14 is further configured to execute the following method: Perform speech-to-text conversion on L real-time processing item video information to obtain L real-time speech texts; extract L historical speech text sets of L processing windows, and perform keyword iterative analysis in combination with the L real-time speech texts to determine L iterative progress features; perform progress recognition based on the L iterative progress features to obtain the L real-time predicted completion durations.

[0049] Furthermore, the progress recognition module 14 is further configured to execute the following method: Extract L first historical speech texts from the L historical speech text sets in chronological order from front to back; perform keyword matching on the L first historical speech texts to construct L first memory vectors; extract L second historical speech texts from the L historical speech text sets again in chronological order from front to back, and perform keyword similarity matching on the L second historical speech texts based on the L first memory vectors respectively to obtain L second historical matching progress keyword sets to update the L first memory vectors to obtain L second memory vectors; perform iterative matching on the remaining historical speech texts in the L historical speech text sets based on the L second memory vectors in turn to obtain L historical memory vectors; perform keyword iterative analysis on the L real-time speech texts based on the L historical memory vectors to obtain L real-time memory vectors; use the progress feature recognition network layer to analyze the L real-time memory vectors to obtain the L iterative progress features.

[0050] Furthermore, the progress recognition module 14 is further configured to execute the following method: Match the L first historical speech texts according to a preset progress keyword library to obtain L first historical matching progress keyword sets, and add the preset progress keyword library and the L first historical matching progress keyword sets into an initially empty vector to obtain L first memory vectors.

[0051] It should be noted that the above sequence of embodiments of the present application is only for description and does not represent the advantages or disadvantages of the embodiments. And the above description of specific embodiments of this specification has been made. The processes depicted in the drawings do not necessarily require the specific order and 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 are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

[0053] This specification and the drawings are only exemplary descriptions of the present application and are considered to have covered any and all modifications, variations, combinations, or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications 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 equivalent technologies, the present application is intended to include these changes and modifications.

Claims

1. A remote video assistance task scheduling method combined with user portraits, characterized in that, The method includes: Collecting information of Q target users according to a preset portrait index set to obtain Q target user portrait index sets, where Q is a positive integer; Performing matter processing prototype matching analysis based on the target user portrait index sets and Q to-be-processed matters of Q target users to obtain Q predicted processing durations of the to-be-processed matters; Interacting with L processing windows of the remote video assistance platform to obtain L real-time processing matter video information and L window to-be-processed queues, where L is a positive integer; Traversing the L real-time processing matter video information for progress recognition to determine L real-time predicted completion durations; Scheduling the Q to-be-processed matters based on the Q predicted processing durations of the to-be-processed matters, the L window to-be-processed queues, and the L real-time predicted completion durations to obtain L updated to-be-processed queues.

2. The remote video errand scheduling method combined with user portraits according to claim 1, wherein Performing matter processing prototype matching analysis based on the target user portrait index sets and Q to-be-processed matters of Q target users to obtain Q predicted processing durations of the to-be-processed matters, including: Pre-constructing a matter processing prototype matching library; Performing similarity matching on the matter processing prototype matching library based on the Q to-be-processed matters and the target user portrait index sets, and obtaining the Q predicted processing durations of the to-be-processed matters according to the matching results.

3. The remote video errand scheduling method combined with user portraits according to claim 2, characterized in that, Pre-constructing a matter processing prototype matching library, including: Obtaining a set of historical assistance matter logs; Classifying the set of historical assistance matter logs according to the similarity degree of matter types and the similarity degree of user portrait indexes to obtain M classified historical assistance matter log sets, where M is a positive integer greater than or equal to Q; Constructing a matter processing prototype matching library based on the M classified historical assistance matter log sets.

4. The method for scheduling remote video assistance matters combined with user portraits according to claim 3, characterized in that, Including: Constructing a two-item classification function, where the two-item classification function is: ； Among them, is the two-way classification similarity, is the matter type of the j-th classified historical assistance matter log in the i-th classified historical assistance matter log set, is the user portrait index set of the j-th classified historical assistance matter log in the i-th classified historical assistance matter log set, is the union of the matter types in the i-th classified historical assistance matter log set, is the mean value of the user portrait index sets in the i-th classified historical assistance matter log set, is the weight for balancing the matter type similarity and the user portrait index similarity, is the total number of classified historical assistance matter logs in the i-th classified historical assistance matter log set, and M is the total number of classified historical assistance matter log sets; Using the two-item classification function to perform classification authentication on the M classified historical assistance matter log sets. If the output two-item classification similarity is greater than or equal to a preset similarity, the authentication passes.

5. The method for scheduling remote video assistance matters combined with user portraits according to claim 4, characterized in that, Including: Respectively using the M classified historical assistance matter log sets as training data to construct M matter processing prototypes; Summarizing the M matter processing prototypes to obtain the matter processing prototype matching library.

6. The method for scheduling remote video assistance matters combined with user portraits according to claim 1, characterized in that Traversing the L real-time processing matter video information for progress recognition to determine L real-time predicted completion durations, including: Performing speech-to-text conversion on the L real-time processing matter video information to obtain L real-time speech texts; Extracting L sets of historical speech texts of L processing windows, and performing keyword iteration analysis in combination with the L real-time speech texts to determine L iterative progress features; Performing progress recognition based on the L iterative progress features to obtain the L real-time predicted completion durations.

7. The method for scheduling remote video assistance matters combined with user portraits according to claim 6, wherein, Extracting L sets of historical speech texts of L processing windows, and performing keyword iteration analysis in combination with the L real-time speech texts to determine L iterative progress features, including: Extracting L first historical speech texts in the L sets of historical speech texts in the order from front to back in terms of time; Performing keyword matching on the L first historical speech texts to construct L first memory vectors; Extract the L second historical speech texts from the L historical speech text sets again in the order of time from front to back, and perform keyword similarity matching on the L second historical speech texts respectively based on the L first memory vectors to obtain L second historical matching progress keyword sets to update the L first memory vectors and obtain L second memory vectors; Perform iterative matching on the remaining historical speech texts in the L historical speech text sets in sequence based on the L second memory vectors to obtain L historical memory vectors; Perform keyword iterative analysis on the L real-time speech texts based on the L historical memory vectors to obtain L real-time memory vectors; Analyze the L real-time memory vectors by using the progress feature recognition network layer to obtain the L iterative progress features.

8. The method for scheduling remote video assistance matters combined with user portraits according to claim 7, wherein Match the L first historical speech texts according to the preset progress keyword library to obtain L first historical matching progress keyword sets, and add the preset progress keyword library and the L first historical matching progress keyword sets into an initially empty vector to obtain L first memory vectors.

9. The method for scheduling remote video assistance tasks combined with user portraits according to claim 1, characterized in that, The preset portrait index set includes user attribute indexes, user demand feature indexes, and user historical behavior feature indexes.

10. The remote video assistance matter scheduling system combined with user portraits is characterized in that, The system is used to execute the remote video help desk matter scheduling method combined with user portraits according to any one of claims 1-9, including: Information collection module: Collect information on Q target users according to the preset portrait index set to obtain Q target user portrait index sets, where Q is a positive integer; Matching analysis module: Perform matter processing prototype matching analysis on the Q to-be-processed matters of the Q target users based on the target user portrait index sets to obtain Q predicted durations of to-be-processed matters; Help desk platform interaction module; Interact with L processing windows of the remote video help desk platform to obtain L real-time processing matter video information and L window to-be-processed queues, where L is a positive integer; Progress recognition module: Traverse the L real-time processing matter video information for progress recognition to determine L real-time predicted completion durations; Matter scheduling module: Schedule the Q to-be-processed matters based on the Q predicted durations of to-be-processed matters, the L window to-be-processed queues, and the L real-time predicted completion durations to obtain L updated to-be-processed queues.

Images