Customer service system intelligent order sending method based on work order intention recognition
By constructing an intent recognition model and a multi-dimensional label classification model, and combining them with a reinforcement learning-based dispatch optimization model, we have achieved accurate dispatch of complaint work orders in the customer service system. This solves the problems of low efficiency and uneven resource utilization in existing technologies, and improves processing speed and accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-27
AI Technical Summary
In existing customer service systems, the dispatching of complaint work orders relies on fixed rules and human experience, resulting in low efficiency, unreasonable dispatching, unbalanced load, inaccurate intent recognition, insufficient granularity of classification, and inability to dynamically optimize the utilization of customer service resources.
An intelligent dispatching method based on work order intent recognition is adopted. By constructing an intent recognition model and a multi-dimensional label classification model, and combining them with a reinforcement learning dispatching optimization model, the method can achieve accurate understanding and multi-dimensional labeling of work order intent, and dynamically optimize dispatching decisions.
It improved the efficiency and quality of complaint ticket processing, enabled automatic parsing of complaint tickets and optimal dispatch decisions, improved processing speed and accuracy, and increased the utilization rate and efficiency of customer service resources.
Smart Images
Figure CN121745562A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of operator customer service systems, and particularly relates to a customer service system intelligent order allocation method based on order intention recognition. BACKGROUND
[0002] In operator customer service business, complaint orders are usually manually allocated to appropriate agents. The traditional order allocation method relies on fixed rules and manual experience, which is not only inefficient, but also prone to unreasonable order allocation and uneven load distribution.
[0003] Although the existing order allocation system introduces keyword matching and simple machine learning algorithms, it still has the following shortcomings:
[0004] 1. Inaccurate intention recognition: keyword matching cannot accurately understand the context of customer complaints, resulting in inaccurate order allocation.
[0005] 2. Insufficient classification granularity: most systems only classify by business category, lack multi-dimensional label information, and cannot fine-tune order allocation.
[0006] 3. Unable to dynamically optimize: lack of optimization capability based on real-time load, historical order allocation results, etc., resulting in low utilization of customer service resources. SUMMARY
[0007] In view of the above shortcomings of the prior art, the purpose of the application is to provide a customer service system intelligent order allocation method based on order intention recognition, which can accurately understand the intention of the order, automatically generate the best order allocation scheme, and improve the processing efficiency and quality of operator customer service orders.
[0008] The application provides a customer service system intelligent order allocation method based on order intention recognition, comprising:
[0009] S1, receiving original customer complaint orders from multiple channels, and preprocessing to obtain preprocessed order text;
[0010] S2, constructing an intention recognition model, using the intention recognition model to perform deep semantic analysis, entity recognition and intention reasoning on the preprocessed order text, and outputting a structured intention recognition result; wherein the structured intention recognition result includes core intention category, confidence and key entity information;
[0011] S3, combining the intention recognition result with the original customer complaint order and the preprocessed order text as the input of a multi-dimensional label classification model; using the multi-dimensional label classification model to predict multiple dimensions of labels for the current order;
[0012] S4, acquire the dynamic input of the agent skill map and real-time load information, input the intention recognition result, multi-dimensional label, agent skill map, real-time load information and historical order allocation effect data into the reinforcement learning order allocation optimization model for processing, acquire the characteristics of the current to-be-allocated work order according to the intention recognition result and multi-dimensional label, acquire the real-time state of all available agents according to the real-time load information and historical order allocation effect data, and analyze in combination with the overall load level of the system to generate an optimal target agent ID for the current work order;
[0013] S5, according to the optimal target agent ID, automatically push the original customer complaint work order, intention recognition result and multi-dimensional label to the workstation of the optimal target agent ID to realize order allocation;
[0014] S6, real-time collection of feedback data in the work order processing process, retraining and fine-tuning of the intention recognition model, multi-dimensional label classification model and order allocation optimization model according to the feedback data.
[0015] Further, in S1, the original customer complaint work order is preprocessed, including: removing noise characters, text standardization, error correction and extracting key information from the original customer complaint work order.
[0016] Further, in S2, the intention recognition model is constructed based on a pre-trained Sentence-BERT model.
[0017] Further, in S2, the intention recognition model converts the preprocessed work order text into a high-dimensional semantic work order vector through the Sentence-BERT model; then calculates the cosine similarity between each work order vector and the pre-defined intention vector library, and takes the highest similarity intention category as the core intention category.
[0018] Further, in S2, the intention recognition model constructs a knowledge base according to the historical work order experience, stores the average vector representation of each standardized intention, and constructs an intention vector library; and calculates the confidence of the intention recognition result based on the cosine similarity or the normalized similarity score.
[0019] Further, in S3, when the confidence of the intention recognition result is <0.8, or the work order text contains preset context information, the work order text of the original customer complaint work order is used for auxiliary classification.
[0020] Further, in S3, the multi-dimensional label includes: business type, urgency, customer registration, emotional tendency and complexity; and then outputs the structured multi-dimensional label.
[0021] Further, in S4, the order allocation optimization model is first trained offline on a large amount of historical order allocation data to learn the optimal order allocation strategy. After training, it can be deployed online for real-time decision-making, and continuously fine-tuned online based on feedback data during the order processing process to form a closed-loop feedback loop.
[0022] Further, in S6, the feedback data during the order processing process includes: the actual processing time of the order, the processing result of the order, the customer satisfaction evaluation and the feedback of the order allocation result of the agent.
[0023] Further, the order allocation optimization model calculates the reward of the corresponding order based on the feedback data during the order processing process.
[0024] The present application has the following advantages:
[0025] The customer service system intelligent order allocation method based on order intent recognition provided by the present application realizes automatic analysis and optimal order allocation decision of complaint orders through natural language processing, large model semantic understanding, multi-dimensional label classification and order allocation optimization algorithm, and improves processing speed and accuracy. The present application uses the powerful semantic understanding ability of the large model to accurately analyze the order intent, combines the multi-dimensional label with the order portrait, and dynamically optimizes the order allocation decision based on historical data and real-time state, to automatically generate an intelligent order allocation method of the best order allocation scheme, thereby improving the processing efficiency and quality of the operator customer service order.
[0026] 1. Accurate understanding of order content: The present application constructs an intent recognition model, uses a large model to accurately analyze the order intent, supports complex semantic understanding and multi-intent recognition. The large model intent recognition technology is used to accurately extract the order intent and key information from unstructured text.
[0027] 2. Fine-grained classification and labeling: The present application assigns business type, priority, emotional tendency and other labels to the order through a multi-dimensional label classifier. The multi-dimensional label classification model labels the order based on business type, urgency, customer level, emotional state and other dimensions to guide order allocation decision.
[0028] 3. Intelligent order allocation decision: The present application considers historical order allocation effect, agent skill map and real-time load state, and uses an order allocation optimization model to generate the best order allocation path. The order allocation optimization model combines historical order allocation data, agent skill map and real-time load information, and uses a pre-trained optimization model to generate the optimal order allocation path.
[0029] 4. Improved order allocation accuracy: The present application realizes that the one-time order allocation success rate is improved to more than 95%.
[0030] 5. Improved processing efficiency: The present application realizes that the average processing time is improved by more than 30%, and the customer waiting time is significantly shortened.
[0031] 6. Resource utilization optimization: the application realizes customer service agent utilization rate improvement of more than 20%, avoiding long-term idle or overload of part of personnel.
[0032] 7. Sustainable optimization: the application continuously optimizes the dispatching strategy through feedback data, and has self-evolution ability. BRIEF DESCRIPTION OF DRAWINGS
[0033] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:
[0034] Figure 1 A flow chart of a customer service system intelligent dispatching method based on work order intent recognition according to an embodiment of the application;
[0035] Figure 2 An architecture diagram of a customer service system intelligent dispatching method based on work order intent recognition according to an embodiment of the application;
[0036] Figure 3 A work order intent recognition and multi-dimensional label classification flow chart according to an embodiment of the application;
[0037] Figure 4 A dispatching optimization decision flow chart according to an embodiment of the application. DETAILED DESCRIPTION
[0038] In order for those skilled in the art to better understand the technical solutions in the embodiments of the application, the technical solutions of the application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the application, rather than all the embodiments of the application. It should be understood that these descriptions are only exemplary, and are not intended to limit the scope of the application. Based on the embodiments of the application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the application.
[0039] In addition, in the following description, the description of well-known structures and techniques is omitted to avoid unnecessary confusion of the concepts disclosed in the application.
[0040] In the description of the present application, it should be noted that, unless otherwise explicitly specified and limited, the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", "third" are only for descriptive purposes and cannot be understood as indicating or implying relative importance. The terms "mounting", "connecting", "connecting" should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.
[0041] The exemplary embodiments will be described in detail hereinbelow with reference to the drawings. In the following description, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of methods and systems consistent with some aspects of the present application as detailed in the appended claims.
[0042] The present application proposes a customer service system intelligent dispatching method based on work order intent recognition, which uses the powerful semantic understanding ability of large models to accurately analyze work order intent, combines fine-grained multi-dimensional labels to profile work orders, and dynamically optimizes dispatching decisions based on historical data and real-time state, to improve the processing efficiency and quality of operator customer service work orders.
[0043] As shown in Figure 1 and Figure 2 The present application proposes a customer service system intelligent dispatching method based on work order intent recognition, which includes:
[0044] S1, receiving original customer complaint work orders from multiple channels, and pre-processing to obtain pre-processed work order texts.
[0045] Specifically, first, receive original customer complaint work orders from multiple channels, and pre-process. Among them, multiple channels include: telephone, online, email, etc. That is, receive original customer complaint work orders (i.e. original text data) from telephone voice to text, online customer service chat records, email, APP feedback forms and other channels.
[0046] Then the original customer complaint work order is preprocessed, including: removing noise characters, text standardization, error correction and extracting key information from the original customer complaint work order. Among them, the extracted key information includes: user ID, time, region, etc.
[0047] Finally, the preprocessed work order text is used as the input of subsequent intent recognition.
[0048] S2, an intent recognition model is constructed, and the preprocessed work order text is subjected to deep semantic analysis, entity recognition and intent reasoning by using the intent recognition model, to output a structured intent recognition result.
[0049] In the present application, the intent recognition model is a large language model LLM constructed based on a pre-trained Sentence-BERT model. The intent recognition model is used to accurately analyze the intent of the work order, supporting complex semantic understanding and multi-intent recognition.
[0050] Reference Figure 3 The processing procedure of the intent recognition model is described
[0051] (1) The preprocessed work order text is input into the intent recognition model.
[0052] (2) The intent recognition model performs deep semantic analysis, entity recognition and intent reasoning on the preprocessed work order text, and the specific implementation is as follows:
[0053] (2.1) Vectorization process: the preprocessed work order text is converted into a high-dimensional semantic vector by the intent recognition model (i.e., the Sentence-BERT model), which is used as the work order vector.
[0054] (2.2) Intent matching: the cosine similarity (Consine Similarity) between each work order vector and the pre-defined intent vector library is calculated, and the intent category with the highest similarity is taken as the core intent category of the recognition result.
[0055] For example:
[0056] Work order text: "The network speed is so slow that it cannot open the webpage" -> vector A;
[0057] The vector library contains "network failure", "fee problem", "attitude complaint" and the corresponding vectors B1, B2 and B3, respectively.
[0058] The output matching result is: similarity (A, B1) = 0.92, similarity (A, B2) = 0.01, similarity (A, B3) = 0.01, the intent category with the highest similarity is taken as the core intent category of the recognition result, and the text intent recognition is inferred as "network failure".
[0059] (2.3) Construction of the intent vector library: The intent recognition model constructs a knowledge base based on historical work order experience, stores the average vector representation of various standardized intents, and constructs an intent vector library, as shown in Table 1.
[0060] Table 1 Intent vector library
[0061]
[0062] (2.4) Confidence calculation: Calculate the confidence of the intent recognition result based on the cosine similarity value or the normalized similarity score.
[0063] (3) Output structured intent recognition results, including core intent categories, confidence, and recognized key entity information.
[0064] Core intent categories: network failure, fee dispute, service change, complaint escalation, and service attitude, etc.
[0065] Confidence: based on the cosine similarity value or the normalized similarity score.
[0066] Key entity information: including product name (such as 5G package), work order service category, location, time, etc.
[0067] S3, combine the intent recognition result with the original customer complaint work order and the preprocessed work order text as the input of the multi-dimensional label classification model; use the multi-dimensional label classification model to predict multiple dimensions of labels for the current work order.
[0068] In the present application, based on the intent result and / or the original customer complaint work order and the preprocessed work order text, a multi-label classification model is used to output fine-grained multi-dimensional labels. Among them, the multi-label classification model labels the work order based on business type, urgency, customer level, emotional state, etc. Multi-dimension, guiding the dispatching decision.
[0069] Reference Figure 3 The label prediction process of the multi-dimensional label classification model is described.
[0070] This step combines the intent recognition result output by S2 with the original / preprocessed text as the input of the multi-dimensional label classification model. When the confidence of the intent recognition result is low (for example, confidence <0.8), or the text contains preset context information (such as time-sensitive feature words, "yesterday", "previous single", "repetition"), the multi-label classification model preferentially uses the work order text of the original customer complaint work order to assist classification.
[0071] The multi-label classification model is constructed based on a pre-training model BERT and adopts a multi-label classification (Multi-label Classification) architecture. After training, the model can simultaneously predict multiple dimensions of labels for a single work order.
[0072] In the present application, the multiple dimensions of labels include: business type, urgency level, customer registration, emotional tendency, and complexity.
[0073] The meanings of the various labels are described as follows:
[0074] Business type: a more subdivided business subclass. For example, "5G network coverage", "broadband fault", "international roaming cost", etc.
[0075] Urgency level: can be obtained by combining rules or model prediction, for example, "urgent", "high", "medium", "low", etc.
[0076] Customer level: usually obtained from the customer database association, and can also be identified and supplemented in the work order text. For example, "VIP diamond", "VIP gold card", "ordinary user", etc.
[0077] Region / area: the specific geographic location involved in the user complaint, such as province, city, county, and even cell / base station, etc.
[0078] Emotional tendency: obtained based on text sentiment analysis, for example, "angry", "anxious", "disappointed", "neutral", "satisfied", etc.
[0079] Problem complexity: obtained based on historical processing time, work order description length, keywords, etc. For example, "simple", "medium", "complex", etc.
[0080] Finally, the structured multi-dimensional labels are output as structured work order feature representations of intent+label, which are used for precise description of the attributes and features of the work order for dispatching decision-making.
[0081] S4, the dynamic input agent skill map and real-time load information are obtained, the intent recognition result, multi-dimensional label, agent skill map, real-time load information, and historical dispatching effect data are input into the reinforcement learning dispatching optimization model for processing, the characteristics of the current work order to be dispatched are obtained according to the intent recognition result and multi-dimensional label, the real-time state of all available agents is obtained according to the real-time load information and historical dispatching effect data, and the system overall load level is analyzed to generate the optimal target agent ID for the current work order.
[0082] In this step, the reinforcement learning dispatching optimization model combines historical dispatching data, agent skill map, and real-time load information to generate an optimal dispatching path using a pre-training optimization model, and outputs an optimal agent ID.
[0083] The intention recognition result of S2 and the multi-dimensional label of S3 are input into a dispatch optimization model together with dynamic information, wherein the dynamic information includes: a skill map of an agent, a real-time load state, and historical dispatch effect data.
[0084] The skill map of an agent: stores the skill information of each customer service agent, such as the business type that is good at (needs to be matched with the business type label in step three), the ability to handle a specific emergency level / customer level work order, language ability, historical performance, etc.
[0085] Real-time load state: the number of work orders currently to be processed by each agent / skill group, the estimated remaining processing time, online / busy / idle state, etc.
[0086] Historical dispatch effect data: dispatch records, processing timeliness, customer satisfaction, one-time resolution rate, etc. of historical work orders, which are used to evaluate the effect of different dispatch strategies.
[0087] The dispatch optimization model adopts a proximal policy optimization (PPO) reinforcement learning framework. This framework is selected because it is stable in training, not sensitive to hyperparameters, and can efficiently handle discrete decision problems. The model models the dispatch decision as a sequential decision problem, with the goal of maximizing long-term rewards (such as the shortest average processing time, the highest customer satisfaction, and the most balanced agent load).
[0088] The state of the dispatch optimization model: includes the features of the current work order to be dispatched (intention + multi-dimensional label), the real-time state of all available agents (skill matching degree, current load), the overall load level of the system, etc.
[0089] The action of the dispatch optimization model is a discrete selection of the target agent (or skill group) for the current work order.
[0090] The reward of the dispatch optimization model: according to the feedback data in the work order processing process, the reward of the corresponding work order is calculated. For example:
[0091] Reward = w1*(1 / processing time) + w2*customer satisfaction + w3*one-time resolution rate - w4*agent load imbalance
[0092] Wherein, w1-w4 are the weight coefficients corresponding to each data.
[0093] Finally, the dispatch optimization model outputs the optimal target agent ID (or skill group ID) generated for the current work order.
[0094] Reference Figure 4 The intelligent dispatching process of the reinforcement learning dispatch optimization model is described.
[0095] Specifically, the order allocation optimization model is first trained offline on a large amount of historical order allocation data to learn the optimal order allocation strategy. After training, it can be deployed online for real-time decision-making; and continuously fine-tuned online according to the feedback data in the order processing process, forming a closed-loop feedback loop.
[0096] (1) Receive structured order features (intention recognition results and multi-dimensional labels), query the agent skill atlas, obtain real-time load status, and call historical order allocation effect data.
[0097] (2) State construction: fuse the intention recognition results, multi-dimensional labels, agent skill atlas, real-time load status, and historical order allocation effect data information to construct the current state (State) of the order allocation optimization model.
[0098] (3) Model inference: input the current state into the trained order allocation optimization model (reinforcement learning Agent), and the model outputs the order allocation action (Action) according to the strategy, that is, selects the target agent (or skill group).
[0099] (4) Execute order allocation: output the optimal agent ID.
[0100] (5) Feedback collection (closed loop): collect the processing results (duration, satisfaction, etc.) of the order in the future, calculate the reward (Reward), and use it for model updating (offline or online learning), forming a closed-loop optimization.
[0101] S5, according to the optimal target agent ID of the intention recognition model, automatically push the original customer complaint order, intention recognition result and multi-dimensional label to the workbench of the optimal target agent ID to realize order allocation.
[0102] This step performs order allocation and execution. According to the optimal agent ID output by S4, the order information is automatically pushed to the workbench of the target agent through a predefined interface (API call, message queue, etc.). The agent receives a new order task notification in its work interface.
[0103] Among them, the order information includes original content, intention recognition result and multi-dimensional label, etc.
[0104] S6, real-time collection of feedback data in the order processing process, retraining and fine-tuning of the intention recognition model, multi-dimensional label classification model and order allocation optimization model according to the feedback data of the intention recognition model.
[0105] This step realizes feedback loop and model iterative optimization: collect feedback data in the order processing process to continuously update the intention recognition, label classification and order allocation optimization models. The agent skill atlas and real-time load information are provided as dynamic input to the order allocation optimization model. Reference Figure 2The database stores historical work orders, agent information, model parameters, etc.
[0106] In the present application, the feedback data in the work order processing process includes: the actual processing time of the work order, the work order processing result (solution, upgrade, transfer, etc.), customer satisfaction evaluation (such as post-research score) and feedback of agent assignment result (such as whether to match their skills). The feedback data is associated with the assignment decision of S4 to form a closed-loop feedback loop.
[0107] The present application supports timed batch update (once a day) or real-time incremental update (based on threshold triggering), where real-time update is given priority to ensure timely optimization, and the intent recognition model, multi-dimensional label classification model and assignment optimization model are retrained or fine-tuned with new feedback data, so that the system has the ability of continuous learning and self-optimization.
[0108] Among them, real-time incremental means: when the number of new work order samples or feedback data accumulates to 5% of the current training set size during system operation, the model fine-tuning task is automatically triggered.
[0109] In summary, the intelligent assignment method of the customer service system based on work order intent recognition of the present application uses an intent recognition model to accurately analyze the intent of the work order, supports complex semantic understanding and multi-intent recognition; uses a multi-dimensional label classification model to label the work order based on business type, urgency, customer level, emotional state and other dimensions, to guide assignment decision; uses an assignment optimization model to combine historical assignment data, agent skill map and real-time load information, and uses a pre-trained optimization model to generate the optimal assignment path.
[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present application, and not to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and any changes or replacements that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application.
Claims
1. A method for intelligent order dispatching in a customer service system based on work order intent recognition, characterized in that, include: S1 receives original customer complaint work orders from multiple channels, performs preprocessing, and obtains the preprocessed work order text. S2, Construct an intent recognition model, and use the intent recognition model to perform deep semantic parsing, entity recognition, and intent reasoning on the preprocessed work order text, and output a structured intent recognition result; wherein, the structured intent recognition result includes: core intent category, confidence level, and key entity information; S3, combine the intent recognition result with the original customer complaint work order and the preprocessed work order text as input to the multidimensional label classification model; use the multidimensional label classification model to predict labels of multiple dimensions for the current work order; S4. Obtain the dynamically input agent skill map and real-time load information. Input the intent recognition result, multi-dimensional labels, agent skill map, real-time load information and historical dispatch effect data into the reinforcement learning dispatch optimization model for processing. Obtain the features of the current work order to be dispatched based on the intent recognition result and multi-dimensional labels. Obtain the real-time status of all available agents based on the real-time load information and historical dispatch effect data. Analyze the data in conjunction with the overall system load level to generate the optimal target agent ID for the current work order. S5, based on the optimal target agent ID, automatically push the original customer complaint work order, intent recognition result and multi-dimensional tags to the workbench of the optimal target agent ID to realize order dispatch; S6. Collect feedback data in real time during the work order processing, and retrain and fine-tune the intent recognition model, multi-dimensional label classification model and dispatch optimization model based on the feedback data.
2. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, In S1, the original customer complaint work orders are preprocessed, including: removing noisy characters, standardizing text, correcting errors, and extracting key information.
3. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, In S2, the intent recognition model is built based on the pre-trained Sentence-BERT model.
4. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, In S2, the intent recognition model converts the preprocessed work order text into a high-dimensional semantic work order vector using the Sentence-BERT model; then, it calculates the cosine similarity between each work order vector and a predefined intent vector library, and takes the intent category with the highest similarity as the core intent category.
5. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, In S2, the intent recognition model constructs a knowledge base based on historical work order experience, stores the average vector representation of various standardized intents, and builds an intent vector library; and calculates the confidence level of the intent recognition result based on cosine similarity or normalized similarity score.
6. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, In S3, when the confidence level of the intent recognition result is <0.8, or when the work order text contains preset contextual information, the original customer complaint work order text is used as the auxiliary classification tool.
7. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, In S3, the labels for the multiple dimensions include: business type, urgency, customer registration, sentiment tendency, and complexity; then, structured multi-dimensional labels are output.
8. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, In S4, the dispatch optimization model is first trained offline on a large amount of historical dispatch data to learn the optimal dispatch strategy. After training, it can be deployed online for real-time decision-making; and it is continuously fine-tuned online based on feedback data during the work order processing, forming a closed-loop feedback loop.
9. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, In S6, the feedback data during the work order processing includes: actual work order processing time, work order processing result, customer satisfaction evaluation, and feedback on agent dispatch results.
10. The intelligent order dispatching method for a customer service system based on work order intent recognition according to claim 1, characterized in that, The dispatch optimization model calculates the reward for each work order based on feedback data during the work order processing.