Intelligent collaborative management method and platform based on artificial intelligence driving

Abstract

The invention discloses an intelligent collaborative management method and platform based on artificial intelligence driving. The method comprises the steps of obtaining a to-be-processed target post in a communication platform, and determining a multi-dimensional feature vector of the target post; constructing a graph neural network; analyzing the multi-dimensional feature vector and the graph neural network by using a deep Q network to obtain a target node used for processing the target post in the graph neural network; and forwarding the target post to the target node, obtaining a processing result of the target node for the target post, and forwarding the processing result to the information receiving area or the target communication address associated with the target post. According to the method and the device, the technical problem of low post management efficiency caused by the fact that the post in the communication platform cannot be efficiently and accurately distributed to the object suitable for processing the post in the prior art is solved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology. Specifically, it relates to an intelligent collaborative management method and platform driven by artificial intelligence. Background Art

[0002] Enterprise forum platforms in related technologies are mainly used for employee communication, supporting employees to post threads to share experiences, ask questions or discuss business-related topics, and other employees can comment and reply under the threads. Taking the enterprise forum of a certain bank as an example, employees can post a discussion thread about the promotion strategy of a new credit product, and colleagues can express their opinions in the comment area. However, the relevant platforms are not deeply integrated with the task management process. When valuable task requirements are generated in the forum, such as the task of optimizing a certain business process proposed based on the discussion, the administrator needs to manually screen the relevant thread content, export and organize it, and then reassign it to the corresponding department. During the task execution process, the forum cannot track the task progress and cannot perform closed-loop processing on the task, resulting in chaotic task management and making it difficult to ensure that tasks are completed on time and with high quality.

[0003] In response to the above problems, no effective solution has been proposed yet. Summary of the Invention

[0004] This application provides an intelligent collaborative management method and platform driven by artificial intelligence to at least solve the technical problem of low post management efficiency caused by the inability of related technologies to efficiently and accurately allocate the posts in the communication platform to the objects suitable for processing the posts.

[0005] According to one aspect of the present application, an artificial intelligence-driven intelligent collaborative management method is provided, including: obtaining a target post to be processed in a communication platform and determining a multi-dimensional feature vector of the target post, where the multi-dimensional feature vector includes: a theme classification feature, an emotion feature, and a keyword encoding feature; constructing a graph neural network, where the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-departmental collaboration edges, department-employee subordination edges, and employee-employee collaboration edges, and the department node attributes at least include: the number of posts to be processed, the average post processing duration, and a professional field encoding, and the employee node attributes at least include: the current status, a skill feature vector, and a historical satisfaction score; analyzing the multi-dimensional feature vector and the graph neural network by using a first deep Q-network to obtain a target node in the graph neural network for processing the target post, where the state in the first deep Q-network includes: the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and the node attributes, and the actions in the first deep Q-network include: selecting a target department node, selecting a target employee node, and performing a multi-hop decision along the graph edge; forwarding the target post to the target node, obtaining the processing result of the target node for the target post, and forwarding the processing result to the information receiving area or the target communication address associated with the target post.

[0006] Optionally, the reward function of the first deep Q-network is determined by the following method: determining a first sub-reward function according to a matching score, where the matching score is used to characterize the semantic matching degree between the post content and the node's professional field, and the matching score is obtained by calculating the similarity between the post keywords and the node's preset tags through natural language processing technology; determining a second sub-reward function according to a time delay, where the time delay is used to characterize the time difference from the post creation timestamp to the successful assignment timestamp; determining a third sub-reward function according to a load index, where the load index is used to characterize the ratio of the current task queue length of the node to the maximum processing capacity of the node; determining a fourth sub-reward function according to a satisfaction score, where the satisfaction score is used to characterize the feedback rating provided by the user after the post is processed; determining the reward function according to the first sub-reward function, the second sub-reward function, the third sub-reward function, and the fourth sub-reward function.

[0007] Optionally, determining a reward function according to the first sub-reward function, the second sub-reward function, the third sub-reward function, and the fourth sub-reward function includes: obtaining the initial weight coefficients of each sub-reward function; continuously detecting the post assignment success rate index, the average processing delay index, and the average user satisfaction index; when detecting a decrease in the post assignment success rate index, increasing the first initial weight coefficient corresponding to the first sub-reward function and the fourth initial weight coefficient corresponding to the fourth sub-reward function; when detecting an increase in the average processing delay index, increasing the second initial weight coefficient corresponding to the second sub-reward function and the third initial weight coefficient corresponding to the third sub-reward function; when detecting a decrease in the average user satisfaction index, increasing the fourth initial weight coefficient; performing a normalization operation after each weight coefficient update to make the sum of the updated weight coefficients equal to 1, obtaining the first weight coefficient, the second weight coefficient, the third weight coefficient, and the fourth weight coefficient; determining the reward function of the first deep Q-network according to the first sub-reward function, the first weight coefficient, the second sub-reward function, the second weight coefficient, the third sub-reward function, the third weight coefficient, the fourth sub-reward function, and the fourth weight coefficient.

[0008] Optionally, determining the multi-dimensional feature vector of the target post includes: extracting the semantic embedding vector of the target post using a pre-trained language model; calculating different probability distributions corresponding to different topic classifications through a topic attention layer; performing a normalization process on the different probability distributions to obtain the first feature component; calculating the lexical intensity score of the target post based on a financial sentiment dictionary, where the financial sentiment dictionary includes multiple different preset words for representing positive sentiment, neutral sentiment, and negative sentiment, and the lexical intensity score is used to quantify the intensity of the explicitly expressed sentiment tendency in the target post; determining the sentiment probability of the target post through a deep learning model, where the sentiment probability is used to represent the probability that the target post belongs to positive sentiment, neutral sentiment, and negative sentiment; dynamically weighting and fusing the lexical intensity score and the sentiment probability according to the financial term density to obtain the second feature component, where the financial term density is the ratio of the number of financial professional terms appearing in the target post to the total number of words in the target post; removing the common stop words and high-frequency financial function words in the target post and retaining the professional terms with a term frequency-inverse document frequency exceeding a preset threshold; screening the core words in the professional terms through the centrality metric index of the term co-occurrence graph and determining the third feature component according to the core words, where the term co-occurrence graph is used to represent the co-occurrence relationship of the retained professional terms in the target post, the nodes of the term co-occurrence graph are the screened professional terms, and the edges of the term co-occurrence graph are the semantic association degrees between the professional terms; splicing the first feature component, the second feature component, and the third feature component, and performing dimensionality reduction processing on the splicing result through an orthogonally constrained linear transformation layer to obtain the multi-dimensional feature vector.

[0009] Optionally, after obtaining the target post to be processed in the communication platform, the method further includes: encoding the target post using a pre-trained language model to obtain a semantic feature vector; determining the click-through rate normalization value and the rank weight of the post author of the target post; concatenating the semantic feature vector, the click-through rate normalization value, and the rank weight of the post author, and determining the concatenation result as the state vector; using a second deep Q-network to analyze the state vector to determine whether the target post is an excellent post, where the action space of the second deep Q-network includes: marking as an excellent post and not marking as an excellent post; the reward function of the second deep Q-network includes: compliance score, expert review passing rate, and mislabeling penalty, where the compliance score is the quantization result of scanning the post content through a preset financial compliance rule base, and the compliance score is the difference between 1 and the target ratio, and the target ratio is the ratio of the number of risk keywords to the total number of keywords; the expert review passing rate is the ratio of the number of excellent posts marked and confirmed by experts to the total number of marked excellent posts; the mislabeling penalty includes a first cost and a second cost, where the first cost is the product of the post exposure and the preset unit attention cost, and the second cost is the product of the post knowledge value coefficient and the preset decay factor, and the preset unit attention cost is determined according to the working hour loss of employees reading the preset content.

[0010] Optionally, if it is determined by the second deep Q-network that the target post is an excellent post, after obtaining the processing result of the target node for the target post, the processing result is sent to the target object for review, and the review result is obtained, and the review result is forwarded to the information receiving area or the target communication address associated with the target post.

[0011] Optionally, the processing result at least includes: target content, processing conclusion type, user privacy level identifier, and result sensitivity label; forwarding the processing result to the information receiving area or the target communication address associated with the target post includes: in the case where the processing conclusion type is the first type used to represent a public reply request, forwarding the target content to the information receiving area; in the case where the processing conclusion type is the second type used to represent personal affair processing, forwarding the target content to the target communication address; in the case where the user privacy level identifier is the first identifier used to represent public permission, forwarding the target content to the information receiving area; in the case where the user privacy level identifier is the second identifier used to represent private permission, forwarding the target content to the target communication address; in the case where the result sensitivity label is the first label used to represent private data, forwarding the target content to the target communication address.

[0012] According to another aspect of the present application, there is also provided an intelligent collaborative management platform driven by artificial intelligence, including: an acquisition unit for acquiring target posts to be processed in a communication platform and determining multi-dimensional feature vectors of the target posts, where the multi-dimensional feature vectors include: topic classification features, sentiment features, and keyword encoding features; a construction unit for constructing a graph neural network, where the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-departmental collaboration edges, department-employee affiliation edges, and employee-inter-employee collaboration edges. The department node attributes at least include: the number of posts to be processed, the average post processing duration, and the professional field encoding. The employee node attributes at least include: the current status, the skill feature vector, and the historical satisfaction score; an analysis unit for analyzing the multi-dimensional feature vectors and the graph neural network using a first deep Q-network to obtain target nodes in the graph neural network for processing the target posts, where the states in the first deep Q-network include: the multi-dimensional feature vectors of the current post, the topological structure of the graph neural network, and the node attributes. The actions in the first deep Q-network include: selecting a target department node, selecting a target employee node, and performing multi-hop decisions along the graph edges; a forwarding unit for forwarding the target post to the target node, obtaining the processing result of the target node for the target post, and forwarding the processing result to the information receiving area or the target communication address associated with the target post.

[0013] According to another aspect of the present application, there is also provided a non-volatile storage medium, where the storage medium includes a stored program, and when the program runs, it controls the device where the storage medium is located to execute the above-mentioned intelligent collaborative management method driven by artificial intelligence.

[0014] According to another aspect of the present application, there is also provided an electronic device, including: a memory and a processor, where the processor is used to run the program stored in the memory, and when the program runs, it executes the above-mentioned intelligent collaborative management method driven by artificial intelligence.

[0015] According to another aspect of the present application, there is also provided a computer program, where when the computer program is executed by a processor, it implements the above-mentioned intelligent collaborative management method driven by artificial intelligence.

[0016] According to another aspect of the present application, there is also provided a computer program product, where the computer program product includes a non-volatile computer-readable storage medium, and the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned intelligent collaborative management method driven by artificial intelligence.

[0017] In this application, the method includes obtaining a target post to be processed in a communication platform and determining a multi-dimensional feature vector of the target post. The multi-dimensional feature vector includes a topic classification feature, an emotion feature, and a keyword encoding feature. A graph neural network is constructed. The nodes in the graph neural network include department nodes and employee nodes, and the edges in the graph neural network include inter-department collaboration edges, department-employee affiliation edges, and employee-employee collaboration edges. The attributes of department nodes at least include the number of posts to be processed, the average processing duration of posts, and a professional field encoding. The attributes of employee nodes at least include the current status, a skill feature vector, and a historical satisfaction score. The first deep Q-network is used to analyze the multi-dimensional feature vector and the graph neural network to obtain a target node in the graph neural network for processing the target post. The state in the first deep Q-network includes the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and the node attributes. The actions in the first deep Q-network include selecting a target department node, selecting a target employee node, and performing multi-hop decisions along the graph edges. The target post is forwarded to the target node, and the processing result of the target node for the target post is obtained and forwarded to the information receiving area associated with the target post or the target communication address, achieving the purpose of efficiently and accurately allocating the posts in the communication platform to the target nodes in the graph neural network, thereby realizing the technical effect of improving the management efficiency of posts, and further solving the technical problem of low post management efficiency caused by the inability of related technologies to efficiently and accurately allocate the posts in the communication platform to the objects suitable for processing the posts. Description of the Drawings

[0018] The drawings described herein are used to provide a further understanding of the present application and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation to the present application. In the drawings:

[0019] Figure 1 is a flowchart of an intelligent collaborative management method based on artificial intelligence drive according to an embodiment of the present application;

[0020] Figure 2 is a structural diagram of an intelligent collaborative management platform based on artificial intelligence drive according to an embodiment of the present application;

[0021] Figure 3 is a hardware structure block diagram of a computer terminal of an intelligent collaborative management method based on artificial intelligence drive according to an embodiment of the present application. Detailed Embodiments

[0022] To enable those skilled in the art to better understand the solution of this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts shall fall within the protection scope of this application.

[0023] It should be noted that the terms "first", "second", etc. in the description and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of this application described here can be implemented in an order different from those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device 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 units not clearly listed or inherent to these processes, methods, products or devices.

[0024] According to the embodiments of this application, a method embodiment of an intelligent collaborative management method driven by artificial intelligence is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described here can be executed in an order different from that here.

[0025] Figure 1 is a flowchart of an intelligent collaborative management method driven by artificial intelligence according to the embodiments of this application. As Figure 1 shown, the method includes the following steps:

[0026] Step S102, obtain the target post to be processed in the communication platform and determine the multi-dimensional feature vector of the target post, where the multi-dimensional feature vector includes: topic classification feature, sentiment feature, keyword coding feature.

[0027] Among them, the communication platform includes but is not limited to: employee forum, instant messaging platform, document collaboration platform.

[0028] It is understandable that, in order to more accurately understand and process the massive post content in the communication platform, especially to identify information that contains potential task value or requires further action, each post that needs to be focused on, namely the target post, needs to be transformed into a structured mathematical representation. This mathematical representation is the "multi-dimensional feature vector". The multi-dimensional feature vector is essentially a list or array of numbers, where each position (dimension) represents the quantified feature value of the post in a specific aspect.

[0029] Specifically, the multi-dimensional feature vector consists of three core components. The topic classification feature refers to automatically determining the business domain or content category mainly discussed in the post through natural language processing techniques (such as topic models or text classifiers), such as "credit products", "internal process optimization", "IT system problems", or "customer service feedback", etc., and using specific numerical codes to represent the topic categories to which they belong. The sentiment feature reflects the emotional tendency or attitude intensity expressed by the post author, which can be a numerical value calculated through a sentiment analysis model, such as a score ranging from -1 (extremely negative) to +1 (extremely positive), or more fine-grained classifications such as the encodings corresponding to positive sentiment, neutral sentiment, and negative sentiment. The keyword encoding feature involves identifying and quantifying the core terms or phrases (such as "slow approval", "system freeze", "suggest adding function X") that appear in the post, and a numerical sequence can be generated through word embedding techniques, which can capture the key concepts mentioned in the post and their importance.

[0030] By calculating and combining the feature values of these three dimensions, a multi-dimensional feature vector is finally formed. This vector becomes a "digital fingerprint" of the post.

[0031] Step S104, construct a graph neural network. Among them, the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-departmental collaboration edges, department-employee subordination edges, and employee-interaction edges. The department node attributes at least include: the number of posts to be processed, the average processing duration of posts, and the professional field code. The employee node attributes at least include: the current status, the skill feature vector, and the historical satisfaction score.

[0032] A graph neural network consists of nodes (or vertices) and edges. Nodes represent entities, and edges represent the connection relationships between nodes. Both nodes and edges can have their own feature vectors to describe their attribute information, and the entire graph can also have global information. By iteratively transmitting and aggregating the feature information of nodes and their neighbors to learn the representations of nodes and the graph, a graph neural network can capture the dependency relationships and topological structure features among the nodes in the graph. A skill feature vector is a vector form used to represent the multi-faceted characteristics of a skill. Each dimension of a skill (such as the type of skill, proficiency level, application scenario, etc.) is quantified into a series of numerical values, and a vector composed of these numerical values is used to comprehensively describe a skill.

[0033] In the above step S104, the nodes in the graph neural network represent the basic constituent units of an enterprise, mainly divided into two categories: department nodes and employee nodes. Department nodes, such as the Credit Department, Risk Control Department, and IT Support Department, represent the functional units in the enterprise; employee nodes represent specific individual members. The connection relationships between nodes are reflected by edges. Among them, the inter-departmental collaboration edges depict the formal or informal collaboration relationships (such as the collaboration between the Credit Department and the Risk Control Department in the approval process) that occur between different departments due to projects or processes; the department-employee subordination edges clearly identify the subordination relationship between an employee and their affiliated department; and the employee-employee collaboration edges reflect the actual work network formed between individual employees based on historical project cooperation, knowledge sharing, or social interaction (such as two employees from different departments often collaborating to solve customer problems).

[0034] Furthermore, to enable the graph neural network to effectively learn and perform automatic task allocation, each node is assigned attribute information that describes its state and capabilities. For department nodes, their attributes at least include: the number of posts to be processed, which quantifies the backlog of pending task requests currently received by the department from the forum and is a key indicator for measuring the department's load; the average processing duration of posts; and the professional field code, which is used to describe the core responsibilities and the types of business that the department is good at handling (such as credit approval, system operation and maintenance, customer complaints) through a vectorized representation, facilitating the matching of tasks to the most suitable department. For employee nodes, their attributes at least include: the current status (such as idle, busy, on vacation); the skill feature vector, which is a multi-dimensional vector encoding the skills mastered by the employee and their proficiency levels (such as data analysis - advanced, customer communication - intermediate, Python programming - proficient), and is the core basis for accurately matching the skills required for tasks; and the historical satisfaction score, which is calculated based on the quality, timeliness of past task completion, or colleague evaluations, and is used to measure the reliability and work performance level of the employee.

[0035] Step S106: Analyze the multi-dimensional feature vector and the graph neural network using the first deep Q-network to obtain the target node in the graph neural network for processing the target post. Here, the state in the first deep Q-network includes: the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and the node attributes. The actions in the first deep Q-network include: selecting a target department node, selecting a target employee node, and performing multi-hop decisions along the graph edges.

[0036] Step S106 is used to combine the multi-dimensional feature vector representing the core information of the post to be processed in Step S102 and the graph neural network depicting the internal structure, capabilities, and relationships within the enterprise in Step S104, and to automatically and intelligently make decisions through reinforcement learning techniques (specifically, the deep Q-network, DQN) on which department or which employee should be responsible for handling the task requirements contained in the post.

[0037] In the above Step S106, the state of the first deep Q-network includes the multi-dimensional feature vector of the current target post. The multi-dimensional feature vector provides the core information of the task requirements, such as the topic under discussion (e.g., "Optimization of the credit approval process"), the expressed sentiment tendency (e.g., "Negative, indicating serious problems in the process"), and the key demand words (e.g., "Simplify", "Accelerate", "Automate"). The state also includes the current topological structure of the entire graph neural network and the real-time attributes of all nodes in the graph neural network. The current topological structure of the graph neural network reflects the collaborative relationships between departments, the subordination relationships between departments and employees, and the possible collaborative links between employees; the real-time attributes of all nodes, such as the current number of posts to be processed (load) of each department, the average processing duration (efficiency), the professional field (capability matching degree), as well as the current status of employees (whether they are idle), the skill vector (whether they possess the skills required to handle the task), and the historical satisfaction scores (reliability and performance). It can be understood that the above state comprehensively depicts "what the task is" and "what departments and employees can do currently".

[0038] The actions mainly include three categories: selecting a target department node (directly assigning the task to a certain department for processing), selecting a target employee node (directly assigning the task to a specific employee for processing), and performing multi-hop decisions along the graph edges. The multi-hop decision is particularly important and is used to simulate the process of the post (i.e., the task in reinforcement learning) flowing through the graph neural network. For example, deciding "Step 1: Select Department A -> Step 2: Along the collaboration edge between Department A and Department B, transfer to Department B -> Step 3: Select Employee C within Department B". Or "Step 1: Select Employee X -> Step 2: Along the collaboration edge between Employee X and Employee Y, transfer to Employee Y". This action design enables flexible handling of situations that require cross-departmental collaboration or finding specific experts, greatly enhancing the rationality and adaptability of the decision-making.

[0039] It is worth noting that multi-hop decision-making allows the agent to make continuous, path-based movements on a graph structure composed of department and employee nodes and various edges connecting them (inter-departmental collaboration edges, department-employee affiliation edges, employee- employee collaboration edges), and finally reach one or more target nodes. Specifically, each "hop" represents an action selected by the DQN in the current state. This action is not the final assignment, but rather indicates that the system moves along a specific "edge" in the graph to the next adjacent "node". For example: The first-hop action: Starting from a starting point (e.g., a certain department node A initially associated with the post topic), move along an inter-departmental collaboration edge to another department node B. This simulates that the task is transferred from department A to department B for processing due to the need for cross-departmental cooperation. The second-hop action: After reaching department node B, then decide to move along a department-employee affiliation edge from department node B to an employee node C affiliated with this department. This means that the task is assigned to the specific employee C within department B. The third-hop action: The DQN may also, at employee node C, decide to move along an employee-employee collaboration edge to another employee node D (even if D may belong to another department). This simulates that employee C believes that they cannot complete the task independently or believes that employee D is a more suitable expert, so they refer or send a collaboration request for the task to D.

[0040] The above "hop" process can continue until the DQN believes that the most suitable final processing node (target department or target employee) has been found, or the preset maximum hop count limit has been reached. The choice of each "hop" (which edge to take and which adjacent node to reach) depends on the DQN's analysis and evaluation of the current combined state (post feature vector + real-time structure / attributes of the entire graph), with the goal of maximizing the long-term cumulative reward.

[0041] Step S108, forward the target post to the target node, obtain the processing result of the target node for the target post, and forward the processing result to the information receiving area or target communication address associated with the target post.

[0042] According to the above steps, obtain the target post to be processed in the communication platform and determine the multi-dimensional feature vector of the target post, where the multi-dimensional feature vector includes: topic classification feature, sentiment feature, keyword encoding feature; construct a graph neural network, where the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-department collaboration edges, department-employee subordination edges, and employee-employee collaboration edges. The department node attributes at least include: the number of posts to be processed, the average post processing duration, and the professional field encoding. The employee node attributes at least include: the current status, the skill feature vector, and the historical satisfaction score; use the first deep Q-network to analyze the multi-dimensional feature vector and the graph neural network to obtain the target node in the graph neural network for processing the target post, where the state in the first deep Q-network includes: the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and the node attributes, and the actions in the first deep Q-network include: selecting a target department node, selecting a target employee node, and performing multi-hop decisions along the graph edges; forward the target post to the target node, obtain the processing result of the target node for the target post, and forward the processing result to the information receiving area or the target communication address associated with the target post, achieving the purpose of efficiently and accurately allocating the posts in the communication platform to the target nodes in the graph neural network, thereby realizing the technical effect of improving the management efficiency of the posts.

[0043] The following makes an exemplary illustration and explanation of Figure 1 the steps shown.

[0044] According to some optional embodiments of the present application, the reward function of the first deep Q-network in step S106 is determined by the following method: determine the first sub-reward function according to the matching score, where the matching score is used to characterize the semantic matching degree between the post content and the node's professional field, and the matching score is obtained by calculating the similarity between the post keywords and the node's preset labels through natural language processing techniques; determine the second sub-reward function according to the time delay, where the time delay is used to characterize the time difference from the post creation timestamp to the successful allocation timestamp; determine the third sub-reward function according to the load index, where the load index is used to characterize the ratio of the current task queue length of the node to the maximum processing capacity of the node; determine the fourth sub-reward function according to the satisfaction score, where the satisfaction score is used to characterize the feedback rating provided by the user after the post is processed; determine the reward function according to the first sub-reward function, the second sub-reward function, the third sub-reward function, and the fourth sub-reward function.

[0045] In this embodiment, the first sub-reward function is based on the matching score. The matching score is used to precisely measure the semantic fit between the task requirement content contained in the target post (reflected by keyword encoding, topic classification, etc. in its multi-dimensional feature vector) and the professional capabilities of potential processing nodes (departments or employees). Specifically, natural language processing techniques (such as word vector similarity calculation and topic model matching degree analysis) can be used to calculate the similarity score between the keywords or core semantics extracted from the post and the preset professional field labels of the nodes (such as the "professional field encoding" of the department or the core ability labels in the "skill feature vector" of the employee). The higher the matching score, the more expertise the node has in handling the task, and the greater the potential for task completion quality and efficiency.

[0046] The time delay refers to the time difference from the timestamp when the forum post was created to the timestamp when the task requirement represented by the post is successfully assigned to the target node. This metric directly reflects the efficiency of the task assignment process. Excessive delay will lead to a lag in problem response, affecting business operations and employee experience. The load index is used to evaluate the busyness and resource availability of the target node (department or employee) at the moment of task assignment. The load index is defined as the ratio of the length of the task queue to be processed by the node currently (such as the "number of posts to be processed" of the department or the number of tasks currently undertaken by the employee) to the maximum processing capacity of the node. The higher the ratio, the more overloaded the node is, and at this time, reassigning new tasks may lead to processing delays, quality degradation, or even system crashes. The satisfaction score is the final quality feedback after the task is closed, usually from the feedback rating provided by the post initiator or relevant users after the task is completed (such as a five-star rating or a satisfied / dissatisfied label). The satisfaction score directly reflects whether the result of task execution has truly solved the problem and met the requirements.

[0047] Specifically, the determination of the reward function based on the first sub-reward function, the second sub-reward function, the third sub-reward function, and the fourth sub-reward function can be achieved through the following method: Obtain the initial weight coefficients of each sub-reward function; Continuously detect the post allocation success rate index, the average processing delay index, and the average user satisfaction index; When it is detected that the post allocation success rate index decreases, increase the first initial weight coefficient corresponding to the first sub-reward function and the fourth initial weight coefficient corresponding to the fourth sub-reward function; When it is detected that the average processing delay index increases, increase the second initial weight coefficient corresponding to the second sub-reward function and the third initial weight coefficient corresponding to the third sub-reward function; When it is detected that the average user satisfaction index decreases, increase the fourth initial weight coefficient; Perform a normalization operation after each weight coefficient update so that the sum of the updated weight coefficients is 1, obtaining the first weight coefficient, the second weight coefficient, the third weight coefficient, and the fourth weight coefficient; Determine the reward function of the first deep Q-network according to the first sub-reward function, the first weight coefficient, the second sub-reward function, the second weight coefficient, the third sub-reward function, the third weight coefficient, and the fourth sub-reward function, the fourth weight coefficient.

[0048] In this embodiment, first, set an initial weight coefficient for each sub-reward function (for example, set based on business experience or preliminary experiments: matching coefficient , delay coefficient , load coefficient , satisfaction coefficient . Continuously monitor several core business operation indicators: post allocation success rate index (the proportion of successfully converting posts into tasks and allocating them), average processing delay index (the total average time from post creation to task completion), average user satisfaction index (the average score of user feedback on task processing results).

[0049] Next, dynamically adjust the corresponding weight coefficients according to the change trends of these monitoring indicators:

[0050] 1. When it is detected that the post allocation success rate index decreases, it means that there is a problem in the task allocation link. It may be that the task requirements do not match the capabilities of the processors (low matching degree), or the processing results do not satisfy the users (resulting in a decrease in the willingness to convert subsequent tasks). Therefore, it is necessary to increase the weight coefficient corresponding to the first sub-reward function (matching score) and increase the weight coefficient corresponding to the fourth sub-reward function (satisfaction score) . The purpose of doing this is to guide the DQN to pay more attention to selecting nodes with a high degree of professional ability matching and the allocation path that can ultimately obtain high satisfaction when making decisions, thereby increasing the probability of tasks being successfully received and effectively completed.

[0051] 2. When the average processing delay metric increases, it indicates that the overall cycle from task submission to completion has become longer and the efficiency has decreased. The delay may be due to the long time taken by the allocation process itself (large time delay), or the task being assigned to a node that is already too busy (high load), resulting in the task being backlogged in the queue. Therefore, it is necessary to increase the weight coefficient corresponding to the second sub-reward function (time delay) and the weight coefficient corresponding to the third sub-reward function (load index) . Furthermore, it prompts the DQN to preferentially select nodes that can respond quickly to allocation requests (low time delay) and have a light current load (low load index) to shorten the overall processing cycle.

[0052] 3. When the average user satisfaction metric decreases, it reflects that the final quality of the task processing result is poor and the user is dissatisfied. Although there may be multiple reasons for the decrease in satisfaction, increasing the weight of the fourth sub-reward is the most direct countermeasure. Therefore, it is necessary to increase the weight coefficient corresponding to the fourth sub-reward function (satisfaction score) .

[0053] Finally, after adjusting one or more weight coefficients according to the above rules, a normalization operation is performed, that is, it is ensured that the sum of all four updated weight coefficients is strictly equal to 1. The reward function of the first deep Q-network is obtained by multiplying the values of the corresponding sub-reward functions by these four dynamically adjusted weight coefficients and then adding them up.

[0054] For this dynamically constructed reward function, the relative importance (weight coefficient) of each internal dimension will change intelligently. When the allocation success rate is low, more emphasis is placed on matching and satisfaction; when the processing delay is high, more emphasis is placed on quick response and load balancing; when the user is dissatisfied, the ultimate satisfaction is most valued. This closed-loop feedback mechanism enables the DQN to continuously learn and optimize its allocation strategy to cope with changing business requirements, and ultimately drives the entire intelligent task allocation system to evolve towards the optimal overall performance.

[0055] According to some other alternative embodiments of the present application, determining the multi-dimensional feature vector of the target post can be achieved through the following method: extracting the semantic embedding vector of the target post using a pre-trained language model; calculating different probability distributions corresponding to different topic classifications through a topic attention layer; performing normalization processing on the different probability distributions to obtain the first feature component; calculating the lexical intensity score of the target post based on a financial sentiment dictionary, where the financial sentiment dictionary includes multiple different preset words for representing positive sentiment, neutral sentiment, and negative sentiment, and the lexical intensity score is used to quantify the intensity of the explicitly expressed sentiment tendency in the target post; determining the sentiment probability of the target post through a deep learning model, where the sentiment probability is used to represent the probability that the target post belongs to positive sentiment, neutral sentiment, and negative sentiment; dynamically weighting and fusing the lexical intensity score and the sentiment probability according to the financial term density to obtain the second feature component, where the financial term density is the ratio of the number of financial professional terms appearing in the target post to the total number of words in the target post; removing common stop words and high-frequency financial function words in the target post, and retaining professional terms with a term frequency-inverse document frequency exceeding a preset threshold; screening the core words in the professional terms through the centrality metric of the term co-occurrence graph, and determining the third feature component according to the core words, where the term co-occurrence graph is used to represent the co-occurrence relationship of the retained professional terms in the target post, the nodes of the term co-occurrence graph are the screened professional terms, and the edges of the term co-occurrence graph are the semantic association degrees between professional terms; splicing the first feature component, the second feature component, and the third feature component, and performing dimensionality reduction processing on the splicing result through an orthogonal constraint linear transformation layer to obtain the multi-dimensional feature vector.

[0056] In this embodiment, first, a pre-trained language model (such as BERT, RoBERTa, etc.) is used to encode the original text of the target post to generate a semantic embedding vector containing overall semantic information. Then, to capture the core business scope discussed in the post, a topic attention layer is introduced. This layer receives the semantic embedding vector and calculates the probability distribution of the post belonging to preset different topic classifications (such as "credit products", "risk control", "operation and maintenance", "customer service", "internal processes", etc.). After this probability distribution is normalized (ensuring that the sum of all probabilities is 1), it forms the first feature component representing the main discussion direction of the post, clearly quantifying the topic tendency of the post content.

[0057] Furthermore, in terms of sentiment analysis, a two-pronged strategy is adopted to improve accuracy in the context of the financial domain. On the one hand, based on a specially constructed financial sentiment dictionary (containing a large number of pre-set financial-related words labeled as positive, neutral, or negative), the lexical intensity score of the post is calculated. This score directly quantifies the intensity of the explicitly expressed sentiment tendency in the text by statistically analyzing the sentiment words and their intensities (possibly considering word frequency, position, modifiers, etc.) that appear in the post. On the other hand, a deep learning sentiment analysis model is used to predict the probability distribution of whether the overall post belongs to positive, neutral, or negative sentiment. Taking the recurrent neural network as an example, by introducing a recurrent structure in the network, information can be transmitted along the sequence, enabling the capture of sequential information in the text. For a post, the model processes each word (or character) one by one. Each word is transformed into a vector of a fixed dimension through an embedding layer, and this vector can represent the semantic information of the word. Then, the recurrent neural network influences the processing of subsequent words based on the information of the previous words, thus being able to understand the context and context relationship of the text. The model goes through a multi-layer neural network structure, gradually extracting higher-level text features. At the final stage of the model, there is an output layer, which can be a fully connected neural network layer. The number of output nodes corresponds to the sentiment categories, such as the three categories of positive, neutral, and negative. Each node outputs a probability value. These probability values are processed through the softmax function so that their sum is 1, thus forming a probability distribution. The probability distribution can intuitively represent the likelihood of the post belonging to each sentiment category. For example, a post may have a 0.7 probability of belonging to positive sentiment, a 0.2 probability of belonging to neutral sentiment, and a 0.1 probability of belonging to negative sentiment. In this way, the overall sentiment tendency of the post can be judged based on these probabilities, and at the same time, it also reflects the degree of uncertainty of sentiment classification.

[0058] It is worth noting that in order to integrate the advantages of these two methods and adapt to the characteristics of financial texts, the financial term density (i.e., the ratio of the number of financial professional terms appearing in the post to the total number of words in the post) is introduced as a dynamic weight factor. Specifically, when the financial term density is higher than the threshold, it indicates that the text is more professional and relies more on the lexical intensity score obtained from the financial sentiment dictionary method; when the financial term density is lower than the threshold, it is more inclined to the sentiment probability predicted by the deep learning sentiment analysis model. These two results are dynamically weighted and fused according to the term density, finally generating a second feature component reflecting the sentiment tendency and intensity of the post.

[0059] Furthermore, for identifying the core appeals and professional focuses in the post, focus on extracting key professional terms. Preprocess the post text to remove common stop words and high-frequency financial function words (such as certain high-frequency but low-information financial conjunctions), and then retain those professional terms whose term frequency-inverse document frequency values exceed a preset threshold. These screened terms form the candidate keyword set. To further identify the most central and representative words among them, a term co-occurrence graph is constructed. Among them, the nodes in the term co-occurrence graph are each screened professional term, and the edges between the nodes are established according to the frequency of co-occurrence or semantic association degree of the terms within the post context window (which can be calculated through word vectors). Calculate the centrality metric of each term node in the graph (such as degree centrality, closeness centrality, or eigenvector centrality). Terms with high centrality are considered to be in the core position in the post semantic network and can better represent the core topic of the post. These screened core words are used to form the third feature component, which can be a vector obtained by taking the average or weighted average of their word vectors (the weights can be determined by the centrality).

[0060] Finally, concatenate the first feature component representing the topic tendency, the second feature component representing the emotional intensity, and the third feature component containing the core professional appeal to form a higher-dimensional temporary vector. To compress this high-dimensional temporary vector into a final practical, information-dense, and dimensionally controllable multi-dimensional feature vector while reducing redundancy between features, a linear transformation layer with an orthogonality constraint (such as a linear layer whose weight matrix is constrained to be approximately orthogonal) is used for dimensionality reduction. The orthogonality constraint helps to preserve the information in the original feature space as much as possible and reduce the correlation between features during the dimensionality reduction process, and finally outputs a low-dimensional but high-information-retention multi-dimensional feature vector, which comprehensively encodes the information of the post in three key dimensions: topic, emotion, and core professional appeal.

[0061] In some alternative embodiments of the present application, after obtaining the target post to be processed in the communication platform, the method further includes: encoding the target post using a pre-trained language model to obtain a semantic feature vector; determining the click-through rate normalization value of the target post and the rank weight of the poster; concatenating the semantic feature vector, the click-through rate normalization value, and the rank weight of the poster, and determining the concatenation result as the state vector; using a second deep Q-network to analyze the state vector to determine whether the target post is an excellent post, where the action space of the second deep Q-network includes: marking as an excellent post and not marking as an excellent post; the reward function of the second deep Q-network includes: compliance score, expert review passing rate, and mislabeling penalty, where the compliance score is the quantitative result of risk scanning of the post content through a preset financial compliance rule library, and the compliance score is the difference between 1 and the target ratio, and the target ratio is the ratio of the number of risk keywords to the total number of keywords; the expert review passing rate is the ratio of the number of excellent posts marked and confirmed by experts to the total number of marked excellent posts; the mislabeling penalty includes a first cost and a second cost, where the first cost is the product of the post exposure and the preset unit attention cost, and the second cost is the product of the post knowledge value coefficient and the preset decay factor, and the preset unit attention cost is determined according to the working hour loss of employees reading the preset content.

[0062] In this embodiment, first, a pre-trained language model is used to perform deep semantic encoding on the target post to generate a semantic feature vector containing context information, which can capture the core value of the post in terms of business insights, knowledge depth, etc. At the same time, two key auxiliary indicators are calculated: the click-through rate normalization value (converting the historical click-through rate of the post into a relative popularity indicator in the [0,1] interval to eliminate the influence of traffic differences in different sections) and the rank weight of the poster (presetting a weight coefficient according to the rank of the poster in the enterprise architecture, such as "department director = 0.9, ordinary employee = 0.5" to reflect the difference in authority). These three types of information are fused into a comprehensive state vector through a concatenation operation to form a complete input representing the value potential of the post.

[0063] Secondly, the state vector is input into the second deep Q-network for decision-making analysis. The action space of this network includes a binary choice: marking a top-quality post or not marking it. Its reward function adopts a triple constraint mechanism to ensure the quality of decisions: the compliance score is calculated by scanning through a preset financial compliance rule base, and the specific formula is 1 - (the number of risk keywords / the total number of keywords). A score approaching 1 indicates full compliance, and a score below the threshold triggers a negative reward; the passing rate of expert review is used as a feedback supervision signal, calculated as the proportion of the number of top-quality posts marked and confirmed by domain experts in the total number of marked posts (for example, if 80 out of 100 marked posts are recognized by experts, the reward is 0.8); the mislabeling penalty quantifies the cost of wrong decisions and includes two key cost items: the first cost is the post exposure volume × the preset unit attention cost (the unit cost is converted into monetary value based on the man-hour loss of employees reading typical documents), which is used to punish the attention resources occupied by low-quality posts; the second cost is the post knowledge value coefficient × the preset decay factor (the knowledge value coefficient is retrospectively calculated based on indicators such as the number of references in the knowledge base, and the decay factor increases exponentially with the time delay of mislabeling discovery), which is used to punish the implicit loss of missing high-value content.

[0064] Finally, intelligent decision-making is achieved by dynamically balancing the reward factors. When the compliance score is low, the second deep Q-network avoids risky content; when the passing rate of expert review decreases, the second deep Q-network tightens the marking criteria; the dual-cost design of the mislabeling penalty ensures a balance between "reducing noise interference" and "avoiding the loss of knowledge assets". The second deep Q-network learns the optimal strategy during continuous training, that is, triggering the top-quality marking only when the comprehensive state vector of the post indicates that it can bring a positive net benefit (expected reward > 0), thus achieving the precise capture of knowledge value.

[0065] Furthermore, the above embodiment further includes the following steps: If it is determined by the second deep Q-network that the target post is a top-quality post, after obtaining the processing result of the target node for the target post, the processing result is sent to the target object for review, and the review result is obtained, and the review result is forwarded to the information receiving area or the target communication address associated with the target post.

[0066] As some alternative embodiments of the present application, the processing result at least includes: target content, processing conclusion type, user privacy level identifier, and result sensitivity label.

[0067] Further, forwarding the processing result to the information receiving area associated with the target post or the target communication address information receiving area can be achieved through the following methods: in the case where the processing conclusion type is the first type used to represent a public reply request, forwarding the target content to the information receiving area; in the case where the processing conclusion type is the second type used to represent personal affair processing, forwarding the target content to the target communication address; in the case where the user privacy level identifier is the first identifier used to represent public permission, forwarding the target content to the information receiving area; in the case where the user privacy level identifier is the second identifier used to represent private permission, forwarding the target content to the target communication address; in the case where the result sensitivity label is the first label used to represent private data, forwarding the target content to the target communication address.

[0068] Figure 2 is a structural diagram of an intelligent collaborative management platform driven by artificial intelligence according to an embodiment of the present application, as Figure 2 shown, the platform includes:

[0069] An acquisition unit 22, configured to acquire a target post to be processed in the communication platform and determine a multi-dimensional feature vector of the target post, where the multi-dimensional feature vector includes: a topic classification feature, an emotion feature, and a keyword coding feature.

[0070] A construction unit 24, configured to construct a graph neural network, where the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-department collaboration edges, department-employee subordination edges, and employee-interaction collaboration edges. The department node attributes at least include: the number of posts to be processed, the average post processing duration, and a professional field code. The employee node attributes at least include: the current status, a skill feature vector, and a historical satisfaction score.

[0071] An analysis unit 26, configured to analyze the multi-dimensional feature vector and the graph neural network by using a first deep Q network to obtain a target node in the graph neural network for processing the target post. The states in the first deep Q network include: the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and the node attributes. The actions in the first deep Q network include: selecting a target department node, selecting a target employee node, and performing a multi-hop decision along the graph edge.

[0072] A forwarding unit 28, configured to forward the target post to the target node, obtain a processing result of the target node for the target post, and forward the processing result to the information receiving area associated with the target post or the target communication address.

[0073] Optionally, the reward function of the first deep Q-network is determined as follows: determine a first sub-reward function according to a matching score, where the matching score is used to characterize the semantic matching degree between the post content and the node's professional field, and the matching score is obtained by calculating the similarity between the post keywords and the node's preset tags through natural language processing techniques; determine a second sub-reward function according to a time delay, where the time delay is used to characterize the time difference from the post creation timestamp to the successful assignment timestamp; determine a third sub-reward function according to a load index, where the load index is used to characterize the ratio of the current task queue length of the node to the maximum processing capacity of the node; determine a fourth sub-reward function according to a satisfaction score, where the satisfaction score is used to characterize the feedback rating provided by the user after the post is processed; determine the reward function according to the first sub-reward function, the second sub-reward function, the third sub-reward function, and the fourth sub-reward function.

[0074] Optionally, determining the reward function according to the first sub-reward function, the second sub-reward function, the third sub-reward function, and the fourth sub-reward function includes the following steps: obtain the initial weight coefficient of each sub-reward function; continuously detect the post assignment success rate index, the average processing delay index, and the user satisfaction mean index; when it is detected that the post assignment success rate index decreases, increase the first initial weight coefficient corresponding to the first sub-reward function and the fourth initial weight coefficient corresponding to the fourth sub-reward function; when it is detected that the average processing delay index increases, increase the second initial weight coefficient corresponding to the second sub-reward function and the third initial weight coefficient corresponding to the third sub-reward function; when it is detected that the user satisfaction mean index decreases, increase the fourth initial weight coefficient; perform a normalization operation after each weight coefficient update so that the sum of the updated weight coefficients is 1, and obtain the first weight coefficient, the second weight coefficient, the third weight coefficient, and the fourth weight coefficient; determine the reward function of the first deep Q-network according to the first sub-reward function, the first weight coefficient, the second sub-reward function, the second weight coefficient, the third sub-reward function, the third weight coefficient, the fourth sub-reward function, and the fourth weight coefficient.

[0075] Optionally, determining the multi-dimensional feature vector of the target post includes the following steps: extracting the semantic embedding vector of the target post using a pre-trained language model; calculating different probability distributions corresponding to different topic classifications through a topic attention layer; normalizing the different probability distributions to obtain the first feature component; calculating the lexical intensity score of the target post based on a financial sentiment dictionary, where the financial sentiment dictionary includes multiple different preset words for representing positive sentiment, neutral sentiment, and negative sentiment, and the lexical intensity score is used to quantify the intensity of the explicitly expressed sentiment tendency in the target post; determining the sentiment probability of the target post through a deep learning model, where the sentiment probability is used to represent the probability that the target post belongs to positive sentiment, neutral sentiment, and negative sentiment; dynamically weighted fusing the lexical intensity score and the sentiment probability according to the financial term density to obtain the second feature component, where the financial term density is the ratio of the number of financial professional terms appearing in the target post to the total number of words in the target post; removing the general stop words and high-frequency financial function words in the target post, and retaining the professional terms whose term frequency-inverse document frequency exceeds a preset threshold; screening the core words in the professional terms through the centrality metric of the term co-occurrence graph, and determining the third feature component according to the core words, where the term co-occurrence graph is used to represent the co-occurrence relationship of the retained professional terms in the target post, the nodes of the term co-occurrence graph are the screened professional terms, and the edges of the term co-occurrence graph are the semantic association degrees between the professional terms; concatenating the first feature component, the second feature component, and the third feature component, and performing dimensionality reduction processing on the concatenated result through an orthogonally constrained linear transformation layer to obtain the multi-dimensional feature vector.

[0076] Optionally, after obtaining the target post to be processed in the communication platform, the following steps are further included: encoding the target post using a pre-trained language model to obtain a semantic feature vector; determining the click-through rate normalization value and the post author rank weight of the target post; concatenating the semantic feature vector, the click-through rate normalization value, and the post author rank weight, and determining the concatenated result as the state vector; analyzing the state vector using a second deep Q-network to determine whether the target post is an excellent post, where the action space of the second deep Q-network includes: marking as an excellent post and not marking as an excellent post; the reward function of the second deep Q-network includes: compliance score, expert review passing rate, mislabeling penalty, where the compliance score is the quantization result of scanning the post content through a preset financial compliance rule library, and the compliance score is the difference between 1 and the target ratio, and the target ratio is the ratio of the number of risk keywords to the total number of keywords; the expert review passing rate is the ratio of the number of excellent posts marked and confirmed by experts to the total number of marked excellent posts; the mislabeling penalty includes a first cost and a second cost, where the first cost is the product of the post exposure and the preset unit attention cost, and the second cost is the product of the post knowledge value coefficient and the preset attenuation factor, and the preset unit attention cost is determined according to the working hour loss of employees reading the preset content.

[0077] Optionally, if it is determined by the second deep Q network that the target post is an excellent post, after obtaining the processing result of the target node for the target post, the processing result is sent to the target object for review, and the review result is obtained, and the review result is forwarded to the information receiving area or the target communication address associated with the target post.

[0078] Optionally, the processing result at least includes: target content, processing conclusion type, user privacy level identifier, and result sensitivity label. Forwarding the processing result to the information receiving area or the target communication address associated with the target post includes the following steps: when the processing conclusion type is the first type used to represent a public reply request, forwarding the target content to the information receiving area; when the processing conclusion type is the second type used to represent personal affair processing, forwarding the target content to the target communication address; when the user privacy level identifier is the first identifier used to represent public permission, forwarding the target content to the information receiving area; when the user privacy level identifier is the second identifier used to represent private permission, forwarding the target content to the target communication address; when the result sensitivity label is the first label used to represent private data, forwarding the target content to the target communication address.

[0079] It should be noted that the above Figure 2 Each of the modules can be a program module (for example, a set of program instructions for implementing a specific function), or a hardware module. For the latter, it can be presented in the following forms, but not limited to: the presentation form of each of the above modules is a processor, or the functions of each of the above modules are implemented by a processor.

[0080] It should be noted that Figure 2 The preferred implementation manners of the illustrated embodiments can be referred to Figure 1 the relevant descriptions of the illustrated embodiments, which will not be elaborated here.

[0081] Figure 3 shows a hardware structure block diagram of a computer terminal for implementing an intelligent collaborative management method driven by artificial intelligence. As Figure 3As shown, the computer terminal 30 may include one or more processors 302 (shown as 302a, 302b, ……, 302n in the figure) (the processor 302 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 304 for storing data, and a transmission module 306 for communication functions. In addition, it may further include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply, and / or a camera. Those of ordinary skill in the art can understand that Figure 3 the structure shown is only schematic and does not limit the structure of the above-mentioned electronic device. For example, the computer terminal 30 may further include more or fewer components than Figure 3 shown in, or have a different configuration from Figure 3 that shown.

[0082] It should be noted that the above one or more processors 302 and / or other data processing circuits are generally referred to as "data processing circuits" herein. The data processing circuit may be embodied in software, hardware, firmware, or any combination thereof, in whole or in part. In addition, the data processing circuit may be a single independent processing module, or be incorporated in whole or in part into any one of the other elements in the computer terminal 30. As involved in the embodiments of the present application, the data processing circuit is a kind of processor control (such as the selection of a variable resistor terminal path connected to an interface).

[0083] The memory 304 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the intelligent collaborative management method driven by artificial intelligence in the embodiments of the present application. The processor 302 executes various functional applications and data processing by running the software programs and modules stored in the memory 304, that is, implements the above-mentioned intelligent collaborative management method driven by artificial intelligence. The memory 304 may include a high-speed random access memory, and may further include a non-volatile memory, such as one or more magnetic storage devices, a flash memory, or other non-volatile solid-state memories. In some instances, the memory 304 may further include a memory remotely set relative to the processor 302, and these remote memories may be connected to the computer terminal 30 through a network. Examples of the above network include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0084] The transmission module 306 is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by the communication provider of the computer terminal 30. In one example, the transmission module 306 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission module 306 can be a Radio Frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0085] The display can be, for example, a touch-screen liquid crystal display (LCD), which enables the user to interact with the user interface of the computer terminal 30.

[0086] It should be noted here that in some alternative embodiments, the above Figure 3 illustrated computer terminal may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware elements and software elements. It should be pointed out that Figure 3 is only an example of a specific specific instance and is intended to show the types of components that may exist in the above computer terminal.

[0087] It should be noted that Figure 3 the illustrated computer terminal is used to execute Figure 1 the illustrated intelligent collaborative management method driven by artificial intelligence. Therefore, the relevant explanations in the above execution method of the command also apply to this electronic device, which will not be elaborated here.

[0088] The embodiment of the present application also provides a non-volatile storage medium. The non-volatile storage medium includes a stored program, wherein when the program runs, it controls the device where the storage medium is located to execute the above-mentioned intelligent collaborative management method driven by artificial intelligence.

[0089] A program for a non-volatile storage medium to perform the following functions: obtaining a target post to be processed in a communication platform, and determining a multi-dimensional feature vector of the target post, where the multi-dimensional feature vector includes: a topic classification feature, an emotion feature, and a keyword encoding feature; constructing a graph neural network, where the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-department collaboration edges, department-employee affiliation edges, and employee collaboration edges, and the department node attributes at least include: the number of posts to be processed, the average post processing duration, and a professional field encoding, and the employee node attributes at least include: the current status, a skill feature vector, and a historical satisfaction score; using a first deep Q network to analyze the multi-dimensional feature vector and the graph neural network to obtain a target node in the graph neural network for processing the target post, where the state in the first deep Q network includes: the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and node attributes, and the actions in the first deep Q network include: selecting a target department node, selecting a target employee node, and performing a multi-hop decision along the graph edge; forwarding the target post to the target node, obtaining the processing result of the target node for the target post, and forwarding the processing result to the information receiving area or the target communication address associated with the target post.

[0090] An embodiment of the present application also provides an electronic device, including: a memory and a processor, where the processor is used to run a program stored in the memory, and when the program runs, it executes the above-mentioned intelligent collaborative management method driven by artificial intelligence.

[0091] The processor is used to run a program that performs the following functions: obtaining a target post to be processed in a communication platform, and determining a multi-dimensional feature vector of the target post, where the multi-dimensional feature vector includes: a topic classification feature, an emotion feature, and a keyword encoding feature; constructing a graph neural network, where the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-department collaboration edges, department-employee affiliation edges, and employee collaboration edges, and the department node attributes at least include: the number of posts to be processed, the average post processing duration, and a professional field encoding, and the employee node attributes at least include: the current status, a skill feature vector, and a historical satisfaction score; using a first deep Q network to analyze the multi-dimensional feature vector and the graph neural network to obtain a target node in the graph neural network for processing the target post, where the state in the first deep Q network includes: the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and node attributes, and the actions in the first deep Q network include: selecting a target department node, selecting a target employee node, and performing a multi-hop decision along the graph edge; forwarding the target post to the target node, obtaining the processing result of the target node for the target post, and forwarding the processing result to the information receiving area or the target communication address associated with the target post.

[0092] The serial numbers of the embodiments of the present application above are only for description and do not represent the superiority or inferiority of the embodiments.

[0093] In the above embodiments of the present application, the descriptions of the respective embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0094] In the above embodiments of the present application, the information collected is information and data authorized by the user or fully authorized by all parties, and the processing of the relevant data, such as collection, storage, use, processing, transmission, provision, disclosure, and application, all comply with relevant laws, regulations, and standards, take necessary protection measures, do not violate public order and good customs, and provide corresponding operation entrances for the user to choose to authorize or refuse.

[0095] In several embodiments provided by the present application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only illustrative. For example, the division of the units can be a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of units or modules can be in an electrical or other form.

[0096] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0097] In addition, the functional units in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0098] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the related technology, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), mobile hard disks, magnetic disks, or optical discs.

[0099] The foregoing are only the preferred embodiments of this application. It should be noted that for those of ordinary skill in the art, without departing from the principle of this application, several improvements and refinements can still be made, and these improvements and refinements should also be regarded as the protection scope of this application.

Claims

1. An intelligent collaborative management method driven by artificial intelligence, characterized in that Including: Obtain a target post to be processed in the communication platform, and determine the multi-dimensional feature vector of the target post, where the multi-dimensional feature vector includes: topic classification feature, sentiment feature, keyword encoding feature; Construct a graph neural network, where the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-department collaboration edges, department-employee subordination edges, and employee-employee collaboration edges. The department node attributes at least include: the number of posts to be processed, the average post processing duration, and the professional field encoding. The employee node attributes at least include: the current status, the skill feature vector, and the historical satisfaction score; Use the first deep Q-network to analyze the multi-dimensional feature vector and the graph neural network to obtain the target node in the graph neural network for processing the target post. The state in the first deep Q-network includes: the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and the node attributes. The actions in the first deep Q-network include: selecting a target department node, selecting a target employee node, and performing multi-hop decisions along the graph edges; Forward the target post to the target node, obtain the processing result of the target node for the target post, and forward the processing result to the information receiving area or the target communication address associated with the target post.

2. The method according to claim 1, characterized in that, The reward function of the first deep Q-network is determined by the following method: Determine the first sub-reward function according to the matching score, where the matching score is used to characterize the semantic matching degree between the post content and the node's professional field, and the matching score is obtained by calculating the similarity between the post keywords and the node's preset tags through natural language processing techniques; Determine the second sub-reward function according to the time delay, where the time delay is used to characterize the time difference from the post creation timestamp to the successful assignment timestamp; Determine the third sub-reward function according to the load index, where the load index is used to characterize the ratio of the current task queue length of the node to the maximum processing capacity of the node; Determine the fourth sub-reward function according to the satisfaction score, where the satisfaction score is used to characterize the feedback rating provided by the user after the post is processed; Determine the reward function according to the first sub-reward function, the second sub-reward function, the third sub-reward function, and the fourth sub-reward function.

3. The method according to claim 2, characterized in that, Determine the reward function according to the first sub-reward function, the second sub-reward function, the third sub-reward function, and the fourth sub-reward function, including: Obtain the initial weight coefficients of each sub-reward function; Continuously detect the post assignment success rate index, the average processing delay index, and the user satisfaction mean index; In the case of detecting a decrease in the post assignment success rate metric, increase the first initial weight coefficient corresponding to the first sub-reward function and the fourth initial weight coefficient corresponding to the fourth sub-reward function; in the case of detecting an increase in the average processing delay metric, increase the second initial weight coefficient corresponding to the second sub-reward function and the third initial weight coefficient corresponding to the third sub-reward function; in the case of detecting a decrease in the user satisfaction mean metric, increase the fourth initial weight coefficient; Perform a normalization operation after each weight coefficient update so that the sum of the updated weight coefficients is 1, obtaining the first weight coefficient, the second weight coefficient, the third weight coefficient, and the fourth weight coefficient; Determine the reward function of the first deep Q-network according to the first sub-reward function, the first weight coefficient, the second sub-reward function, the second weight coefficient, the third sub-reward function, the third weight coefficient, the fourth sub-reward function, and the fourth weight coefficient; 4. The method according to claim 1, characterized in that Determine the multi-dimensional feature vector of the target post, including: Extract the semantic embedding vector of the target post using a pre-trained language model; calculate different probability distributions corresponding to different topic classifications through a topic attention layer; perform normalization processing on the different probability distributions to obtain the first feature component; Calculate the lexical intensity score of the target post based on a financial sentiment dictionary, where the financial sentiment dictionary includes multiple different preset words for representing positive sentiment, neutral sentiment, and negative sentiment, and the lexical intensity score is used to quantify the intensity of the explicitly expressed sentiment tendency in the target post; Determine the sentiment probability of the target post through a deep learning model, where the sentiment probability is used to represent the probability that the target post belongs to positive sentiment, neutral sentiment, and negative sentiment; Dynamically weight and fuse the lexical intensity score and the sentiment probability according to the financial term density to obtain the second feature component, where the financial term density is the ratio of the number of financial professional terms appearing in the target post to the total number of words in the target post; Remove the common stop words and high-frequency financial function words in the target post, and retain the professional terms with a term frequency-inverse document frequency exceeding a preset threshold; Screen the core words in the professional terms through the centrality metric of the term co-occurrence graph, and determine the third feature component according to the core words, where the term co-occurrence graph is used to represent the co-occurrence relationship of the retained professional terms in the target post, the nodes of the term co-occurrence graph are the screened professional terms, and the edges of the term co-occurrence graph are the semantic association degrees between the professional terms; Concatenate the first feature component, the second feature component, and the third feature component, and perform dimensionality reduction processing on the concatenation result through an orthogonal constraint linear transformation layer to obtain the multi-dimensional feature vector.

5. The method according to claim 1, wherein After obtaining the target post to be processed in the communication platform, the method further includes: Encode the target post using a pre-trained language model to obtain a semantic feature vector; determine the click-through rate normalization value and the poster rank weight of the target post; concatenate the semantic feature vector, the click-through rate normalization value, and the poster rank weight, and determine the concatenation result as the state vector. Use a second deep Q-network to analyze the state vector to determine whether the target post is an excellent post. The action space of the second deep Q-network includes: marking as an excellent post and not marking as an excellent post. The reward function of the second deep Q-network includes: compliance score, expert review pass rate, and mislabeling penalty. The compliance score is the quantization result of scanning the post content through a pre-set financial compliance rule library. The compliance score is the difference between 1 and the target ratio, where the target ratio is the ratio of the number of risk keywords to the total number of keywords. The expert review pass rate is the ratio of the number of excellent posts marked and confirmed by experts to the total number of marked excellent posts. The mislabeling penalty includes a first cost and a second cost. The first cost is the product of the post exposure and the pre-set unit attention cost, and the second cost is the product of the post knowledge value coefficient and the pre-set decay factor. The pre-set unit attention cost is determined based on the working hour loss of employees reading pre-set content.

6. The method according to claim 5, wherein The method further includes: if it is determined that the target post is an excellent post using the second deep Q-network, after obtaining the processing result of the target node for the target post, send the processing result to the target object for review, and obtain the review result, and forward the review result to the information receiving area associated with the target post or the target communication address.

7. The method according to claim 1, wherein The processing result at least includes: target content, processing conclusion type, user privacy level identifier, and result sensitivity label. Forwarding the processing result to the information receiving area associated with the target post or the target communication address information receiving area includes: In the case where the processing conclusion type is the first type indicating a public reply request, forward the target content to the information receiving area; in the case where the processing conclusion type is the second type indicating personal affair processing, forward the target content to the target communication address. In the case where the user privacy level identifier is the first identifier indicating public permission, forward the target content to the information receiving area; in the case where the user privacy level identifier is the second identifier indicating private permission, forward the target content to the target communication address. In the case where the result sensitivity label is the first label indicating private data, forward the target content to the target communication address.

8. An intelligent collaborative management platform driven by artificial intelligence, characterized in that, Includes: An acquisition unit for acquiring a target post to be processed in a communication platform and determining a multi-dimensional feature vector of the target post, where the multi-dimensional feature vector includes: topic classification feature, sentiment feature, and keyword encoding feature. Building unit for building a graph neural network, wherein the nodes in the graph neural network include: department nodes and employee nodes, and the edges in the graph neural network include: inter-department collaboration edges, department-employee affiliation edges, and employee collaboration edges. The department node attributes at least include: the number of posts to be processed, the average processing duration of posts, and the professional field code. The employee node attributes at least include: the current status, the skill feature vector, and the historical satisfaction score; Analysis unit for analyzing the multi-dimensional feature vector and the graph neural network using a first deep Q-network to obtain the target node in the graph neural network for processing the target post, wherein the state in the first deep Q-network includes: the multi-dimensional feature vector of the current post, the topological structure of the graph neural network, and the node attributes, and the actions in the first deep Q-network include: selecting a target department node, selecting a target employee node, and performing a multi-hop decision along the graph edges; Forwarding unit for forwarding the target post to the target node, obtaining the processing result of the target node for the target post, and forwarding the processing result to the information receiving area or the target communication address associated with the target post.

9. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein when the program runs, it controls the device where the non-volatile storage medium is located to execute the artificial intelligence-driven intelligent collaborative management method according to any one of claims 1 to 7.

10. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the artificial intelligence-driven intelligent collaborative management method according to any one of claims 1 to 7.

Images