A method for generating instruction chains based on contextual intent recognition

Abstract

This invention relates to the field of artificial intelligence conversation processing technology, and discloses a method for generating instruction chains based on contextual intent recognition. The method includes: extracting event types and determining event risk levels from user conversation requests; for low-risk or medium-risk events, establishing an intelligent conversation window; dynamically capturing user input and contextual information to form basic conversation features; inputting these features into an explicit intent recognition network to obtain first intent information and confidence level; generating and displaying a response based on the first intent, while simultaneously scoring it in real time using an emotion perception model; when the user's emotion or intent confidence level is below a threshold, constructing enhanced conversation features and inputting them into a implicit intent recognition network to obtain second intent information; finally, combining the first and second intents to generate an ordered instruction chain. This achieves accurate intent recognition, intelligent response, and automated business operations in multi-turn conversations.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence conversation processing technology, and more specifically to a method for generating instruction chains based on contextual intent recognition. Background Technology

[0002] With the rapid development of artificial intelligence and natural language processing technologies, conversation-based intelligent interaction systems have been widely applied in scenarios such as customer service, intelligent question answering, and task execution within the State Grid Corporation of China. In existing technologies, intelligent conversation systems typically identify the intent of user-input text to match corresponding business operations or generate responses. However, existing intelligent conversation systems often suffer from the following technical problems:

[0003] First, existing intelligent conversation systems struggle to accurately understand users' potential needs and emotional changes, and their grasp of user intent is insufficient in multi-round interactions, leading to inaccurate responses and inefficient processing.

[0004] Second, existing intelligent conversations typically rely solely on single-turn user input for intent recognition, lacking full utilization of multi-turn contexts and failing to effectively allocate appropriate human customer service representatives in handling high-risk events. This results in inaccurate intent recognition, unreasonable instruction generation, and difficulty in guaranteeing efficiency and service quality in handling high-risk events.

[0005] Third, existing technologies assess user emotions through single-modal information, leading to inaccurate emotion judgments and affecting the timely response and accurate judgment of user status during intelligent conversations. Summary of the Invention

[0006] The summary section of this invention provides a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] This invention proposes an instruction chain generation method based on contextual intent recognition to solve one or more of the technical problems mentioned in the background section above.

[0008] This invention provides a method for generating instruction chains based on contextual intent recognition, comprising: upon receiving a user session request, extracting the event type from the user session request, and determining the event risk level based on the event type;

[0009] If the event risk level is low or medium, an intelligent conversation window is established and the context data corresponding to the user's current input information is dynamically captured, and basic conversation features are generated; the basic conversation features are input into the explicit intent recognition network to obtain the first intent information and the first confidence level.

[0010] Based on the initial intent information, a response to the user's question is generated and displayed in the intelligent conversation window;

[0011] During the conversation, the user's emotion is scored in real time using an emotion perception model;

[0012] When a user's emotion score is lower than a preset score threshold; or when the first confidence score is less than a preset threshold and the current conversation round in the smart conversation window reaches the first preset round threshold, then an enhanced conversation feature is constructed and input into the implicit intent recognition network to obtain the second intent information;

[0013] Based on the first intent information and the second intent information, an instruction chain is generated.

[0014] Optionally, an enhanced session feature is constructed and fed into the implicit intent recognition network, including:

[0015] User features are obtained by extracting features from user emotion ratings and user attribute information; user features are then concatenated with current basic conversation features to obtain enhanced conversation features, which are then input into the implicit intent recognition network.

[0016] Optionally, the current basic session features are extracted through the following steps:

[0017] Extract the conversation data within a preset time window in the smart conversation window, and perform a redundancy deletion operation on the conversation data to obtain the redundant deleted conversation data;

[0018] If the number of complete dialogue rounds contained in the session data after redundancy removal is less than the second preset round threshold, then the time window is updated based on the preset time window until the number of non-redundant complete dialogue rounds contained in the updated time window is greater than or equal to the second preset round threshold.

[0019] Word embedding is performed on the session data within the update time window to obtain the current basic session features.

[0020] Optionally, based on the first intent information and the second intent information, an instruction chain is generated, including:

[0021] The first intent information and the second intent information are matched in a pre-defined set of operation items to obtain multiple matching operation items; each operation item corresponds to a set of instructions.

[0022] Based on the dependencies and time sequence between instructions, multiple sets of instructions corresponding to multiple operation items are sorted to form an instruction chain.

[0023] Optionally, the instruction chain generation method based on contextual intent recognition of the present invention further includes:

[0024] If the event risk level is high risk, a human chat window is created and the corresponding customer service personnel are matched with the user's chat request. The human chat window is used to display the chat between the customer service personnel and the user.

[0025] Customer service personnel are matched through the following steps:

[0026] Match the appropriate customer service representative to the user's session request based on user attribute information or the user's historical session records.

[0027] Optionally, by using user attribute information or user's historical session records, a corresponding customer service representative can be matched for the user's session request, including:

[0028] If a user has multiple historical conversation records, they are sorted from highest to lowest according to the user's feedback rating for each historical conversation record, and the customer service representative with the highest feedback rating is selected as the candidate customer service representative; if the candidate customer service representative is currently available, then the match is determined.

[0029] If no historical session records are found for the user, then a matching customer service representative will be selected from those currently available based on the user's attribute information.

[0030] If a user has historical conversation records, but the user has not rated these records, the customer service representative with the highest overall rating will be selected from among the multiple customer service representatives associated with the historical conversation records to be the matching representative.

[0031] Optional, user attribute information includes user geolocation and account type, and

[0032] Based on user attribute information, select matching customer service personnel from those currently available, including:

[0033] Based on the customer service personnel who are currently available, determine the corresponding customer service information, which includes the customer service identifier and the skill tag corresponding to the customer service identifier;

[0034] Based on the user's geographical location, the system will match the available customer service personnel to the region to obtain at least one primary candidate customer service personnel.

[0035] If there are multiple first-selection customer service personnel, business matching is performed from the first-selection customer service personnel according to the account type, and at least one second-selection customer service personnel is selected.

[0036] If there are multiple second-selection customer service personnel, the skill tags of the customer service personnel are matched with the event type selected by the user to obtain at least one target customer service personnel; the target customer service personnel are then identified as the matched customer service personnel.

[0037] Optionally, the overall customer service rating is determined through the following steps:

[0038] Collect historical service data for each of the multiple customer service personnel. The historical service data includes historical response time, problem resolution rate, user satisfaction rating, and complaint rate.

[0039] The various indicators in the historical service data are weighted according to preset weights to obtain the weighted total score for each customer service representative; the weighted total score is used as the overall customer service score.

[0040] The present invention has the following beneficial effects:

[0041] 1. Improve the accuracy and efficiency of intent recognition. Specifically, through explicit and implicit intent recognition networks, it can not only understand users' explicit needs but also capture potential or implicit needs, improving multi-turn conversation understanding capabilities. It also provides a real-time emotion perception and feedback mechanism, adjusting strategies to improve user satisfaction when users experience negative emotions. Furthermore, through instruction chain generation logic, it can automatically sort and schedule operation items, achieving fast and accurate business processing and reducing manual intervention. An event risk classification mechanism ensures that high-risk events are promptly diverted to human customer service, while low / medium-risk events follow intelligent paths, achieving resource optimization.

[0042] 2. Improve the understanding and response capabilities of multi-turn conversations, and ensure that high-risk events are handled promptly and appropriately. Specifically, through redundancy removal, time window updates, and word embedding, multi-turn conversation data is transformed into concise and sufficient basic conversation features, ensuring the accuracy of explicit and implicit intent recognition. Multi-turn intent information is matched with a preset set of operation items, and instruction chains are generated based on dependencies and time sequence to automate and accurately execute multi-step business operations. By using user attribute information and historical conversation records, combined with feedback scores or comprehensive scores, available customer service personnel are intelligently matched, enabling high-risk events to be handled promptly and accurately, while reducing the blindness of manual intervention and improving service efficiency and user satisfaction.

[0043] 3. Improve the accuracy and real-time performance of emotion recognition. Specifically, text, facial, and speech features are extracted using semantic emotion recognition, facial expression recognition, and speech emotion analysis models, respectively. Pre-trained models ensure accurate feature extraction. Emotional information from the three modalities is quantified into comparable scores (facial emotion score, speech emotion score, and semantic emotion score), providing a foundation for fusion. The three emotion scores are merged into a fusion model by weighted summation or feature concatenation using preset weights to obtain the final user emotion score. The fusion method can adaptively adjust the importance of different modalities, improving the accuracy and robustness of emotion assessment. The obtained user emotion score can be used in intelligent conversation windows to dynamically adjust response strategies or trigger enhanced intent recognition, thereby improving interaction quality and user satisfaction. Attached Figure Description

[0044] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0045] Figure 1 This is a flowchart of an instruction chain generation method based on contextual intent recognition according to the present invention;

[0046] Figure 2 This is a flowchart illustrating the creation of an artificial session window using a method for generating instruction chains based on contextual intent recognition, as described in this invention.

[0047] Figure 3 This is a flowchart illustrating the matching process for customer service personnel in a method for generating instruction chains based on contextual intent recognition, as described in this invention. Detailed Implementation

[0048] The invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the drawings and embodiments of the invention are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0049] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0050] It should be noted that the concepts of "first" and "second" mentioned in this invention are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0051] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0052] The names of messages or information exchanged between the various devices of this invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0053] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0054] like Figure 1 The diagram illustrates a flowchart of an instruction chain generation method based on contextual intent recognition according to the present invention, which specifically includes the following steps:

[0055] Step 101: Upon receiving a user session request, extract the event type from the user session request and determine the event risk level based on the event type;

[0056] In some embodiments, the execution entity of the instruction chain generation method based on contextual intent recognition of the present invention is a backend server. Based on this, after receiving a user session request, the backend server receives the event type actively selected by the user in the session interface. Then, according to the event type, it queries the corresponding risk level from a risk grading table pre-stored on the backend server. The risk grading table includes event types and corresponding event risk levels. For example, if the event type is "repair request," the corresponding event risk level is medium risk. If the event type is "query," the corresponding event risk level is low risk. The user session request refers to the interactive request initiated by the user in the session interface. The session interface provides several preset event type buttons, and the user actively selects the corresponding event type during the session. The event type refers to the business processing category label actively selected by the user, used to quickly clarify the business background and processing priority of the session. Event types include, but are not limited to: queries, such as electricity bill inquiries and electricity consumption inquiries; complaints, such as service attitude complaints and billing disputes; repair requests, such as meter damage and power outages; and emergency event reporting, such as transformer sparks and substation fires. The event risk level refers to the preset risk level for different event types, used to determine processing strategies and triage paths.

[0057] Step 102: If the event risk level is low or medium risk, establish an intelligent conversation window and dynamically capture the context data corresponding to the user's current input information, and generate basic conversation features; input the basic conversation features into the explicit intent recognition network to obtain the first intent information and the first confidence level;

[0058] In some embodiments, when the event risk level is low or medium, the backend server creates an intelligent conversation window on the user terminal to display multi-turn dialogue interactions with the user and supports real-time responses to user input. During this process, the backend server dynamically captures the user's current input information and its related contextual data; that is, it automatically and in real-time acquires conversation history, user attribute information, and business status information related to the current user input during the conversation. The dynamically captured contextual data, along with the user's current input, undergoes text cleaning and encoding to generate basic conversation features. The backend server then inputs these basic conversation features into an explicit intent recognition network for inference. The explicit intent recognition network employs a relatively simple and computationally efficient natural language understanding model, suitable for quickly identifying the user's direct intent in real-time conversation scenarios. For example, TextCNN (a text classification model based on convolutional neural networks) can be used. This type of model, after the input text is segmented and vectorized, extracts and classifies features through a lightweight network structure, and uses a Softmax classification layer in the output layer to transform the results into probability distributions for each intent category, thereby obtaining the first intent information and the corresponding first confidence level. For example, the model outputs the first intent information (such as "electricity bill inquiry") and the corresponding first confidence score (such as 0.92). The intelligent conversation window refers to the interactive interface displayed on the user's terminal (such as a mobile app or webpage). This interface supports real-time information exchange with the backend server and dynamically displays responses based on user input. The intelligent conversation window can load context data, display automatically generated responses, and provide multi-turn dialogue functions. Context data refers to historical conversation content, user attribute information, and business status information related to the user's current input. This mainly includes: messages previously sent by the user in the current conversation, previous replies, and necessary historical business records (such as historical payment records, repair records, etc.). Basic conversation features refer to the vectorized representation obtained after feature extraction processing of the user's current input information and its context data. This feature is used as input to the explicit intent recognition network to determine the user's direct intent. The explicit intent recognition network refers to a deep learning-based natural language understanding model used to identify direct and explicit intents in user input. Explicit intents can usually be directly determined through textual expression, such as "I want to check my electricity bill" or "I want to file a complaint." The first intent information refers to the classification result of the explicit intent recognition network on the user's current input intent. Examples include "electricity bill inquiry," "equipment repair request," and "account recharge." The first confidence level refers to the degree of confidence the explicit intent recognition network has in the first intent information; it typically ranges from 0 to 1, with higher values ​​indicating greater confidence in the correctness of the intent classification. The Softmax classification layer is a commonly used output layer in deep learning for multi-class classification tasks. Its core function is to transform the neural network's output into a probability distribution for each class.TextCNN is a text classification model commonly used in Natural Language Processing (NLP), based on a Convolutional Neural Network (CNN) architecture.

[0059] Step 103: Based on the first intent information, generate a response to the user's question and display it in the smart conversation window;

[0060] Step 104: During the conversation, the user's emotion is scored in real time using an emotion perception model;

[0061] In some embodiments, the executing entity matches the corresponding reply content in a predefined reply template library based on the first intent information. For example, if the first intent information is "electricity bill inquiry," the corresponding reply in the template library is "Your electricity bill for July is 325 yuan, and the payment deadline is August 31st." The generated reply information is sent to the user through a smart conversation window to complete the information display. As an example, if the first intent information is "reporting meter repair," the backend server matches a template or generates a dynamic reply: "The meter malfunction in your area has been registered. We will arrange for staff to come and handle it as soon as possible. Please keep your phone accessible." During the conversation, the latest input text from the user is captured and segmented and vectorized. Then, it is input into a trained emotion perception model, which outputs a semantic emotion score. The emotion perception model can be an LSTM (Long Short-Term Memory) emotion classification model. Based on this, the latest input text from the user is input into the LSTM emotion classification model, and the output hidden state is processed through a fully connected layer and a Softmax classification layer to obtain the probability distribution of each emotion category (such as positive, negative, neutral). The user's emotion score can be the probability value of the corresponding category or a continuous score generated by combining weighted methods. Reply messages refer to targeted text or multimedia information generated based on user intent, used to answer user questions or guide further actions; they can be predefined template text. Emotion perception models are artificial intelligence-based analysis models used to judge and quantify a user's emotional state during a conversation. They can perform emotion recognition based on multimodal information such as text sentiment analysis, voice tone, and facial expression recognition. User emotion scores are quantitative indicators output by the emotion perception model, representing the degree of positive or negative or intense emotion of the user's current emotional state. Values ​​are generally in the range of 0 to 1, with lower values ​​indicating more negative or dissatisfied emotions. LSTM is a commonly used variant of recurrent neural networks, adept at processing sequential data and capable of capturing long-distance dependencies. In text sentiment analysis, LSTM can understand the relationships between words in a sentence, thereby determining the overall emotional tendency. As an example, a user inputs: "The electricity meter is broken, why hasn't it been repaired yet?!" The text sentiment analysis model determines the text semantics to be negative, outputting a user emotion score of 0.3 (below the preset threshold of 0.5), indicating that the user's emotion is relatively intense.

[0062] Step 105: When the user's emotion score is lower than the preset score threshold; or the first confidence score is less than the preset threshold and the current conversation round in the smart conversation window reaches the first preset round threshold, then construct the enhanced conversation features and input them into the implicit intent recognition network to obtain the second intent information;

[0063] Step 106: Generate an instruction chain based on the first intent information and the second intent information.

[0064] In some embodiments, when a user's sentiment score is lower than a preset score threshold; or when the first confidence level is less than a preset threshold and the current session round in the smart session window reaches a first preset round threshold, features are extracted from the user's sentiment score and user attribute information to obtain a unified user feature vector. The user features are then concatenated with the current basic session features to obtain enhanced session features, which are then input into the implicit intent recognition network. Concatenation is a vector-level operation, referring to directly connecting the user feature vector and the current basic session feature vector along one dimension to form a new vector, namely the enhanced session feature vector. User attribute information refers to the user's basic information and business-related information, such as account type, historical repair records, user geographical location, meter number, etc. Feature extraction refers to converting the original information (user sentiment score, user attribute information) into a numerical vector or high-dimensional feature representation for input into the neural network for processing. For example, user attributes are encoded as numerical normalization, and the sentiment score is directly used as a numerical feature. Enhanced session features refer to concatenating user features with basic session features to form a new feature vector, which includes both the user's semantic context and sentiment and personalized information, used for implicit intent recognition. Implicit intent recognition networks are deep learning-based natural language understanding models. For example, they build upon BERT pre-trained models, incorporating multi-layer attention mechanisms or deep Transformer stacking, and adding a Softmax classification layer to the output layer for reasoning under longer contexts and multi-dimensional features. Compared to explicit intent recognition networks, implicit networks are more complex and slower inference, but they can capture latent needs or unexpressed intentions beyond explicit statements. They input enhanced conversation features and output secondary intent information. Secondary intent information refers to the intent category output by the implicit intent recognition network, used to assist in generating more accurate or comprehensive instruction chains. A preset scoring threshold is a numerical limit set to determine whether a user's emotion is negative or dissatisfied. When a user's emotion score is below this threshold, it indicates that the user's emotion may be intense or dissatisfied, requiring further processing measures (such as triggering implicit intent recognition or human intervention). A preset threshold is a confidence level limit set to determine the reliability of the first intent prediction output by the explicit intent recognition network. When the first confidence level is below this threshold, it indicates that there may be uncertainty in explicit intent recognition, requiring further confirmation of the user's intent through enhanced conversation features or implicit intent recognition. The current session round refers to the number of interaction rounds between the user and the backend server in the intelligent session window (each round includes user input and response), used to determine whether the depth or number of rounds of the session has reached the point where further processing is required. The first preset round threshold refers to the pre-set limit on the number of session rounds to determine whether enhanced intent recognition is still needed after multiple rounds of dialogue.BERT (Bidirectional Encoder Representations from Transformers) is a well-known pre-trained language model in the field of natural language processing. Based on the Transformer encoder structure, it employs a bidirectional attention mechanism, allowing it to simultaneously consider the dependencies between preceding and following words in the context, thereby generating word vectors or sentence vectors containing rich semantic information. In specific tasks, BERT can be fine-tuned by adding task-related classifiers (such as Softmax classification layers) to the output layer for purposes such as intent recognition, sentiment analysis, and question answering. The Transformer is a neural network architecture based on a self-attention mechanism, consisting of an encoder and a decoder. The Softmax classification layer is a commonly used output layer in deep learning models, typically used for multi-class classification tasks. Its function is to map the model's output real-valued vectors to a probability distribution, with each class corresponding to a probability value. In intent recognition tasks, the Softmax classification layer can transform the input feature vector (such as the BERT encoding result) into predicted probabilities for each intent class, thus obtaining the most probable intent class and its corresponding confidence level.

[0065] Based on this, the first and second intent information are matched in the operation item library to obtain the corresponding operation items and instruction sets. Instructions are then sorted according to dependency and priority to form an instruction chain. The instruction chain refers to a sequence of operation instructions generated based on the user's intent, used for backend business processing or automated execution. Each instruction corresponds to a specific operation item (such as querying a bill, generating a work order, or assigning staff).

[0066] These embodiments improve the accuracy and efficiency of intent recognition. Specifically, through explicit and implicit intent recognition networks, not only can explicit user needs be understood, but also potential or implicit needs can be captured, improving multi-turn conversation understanding capabilities. A real-time emotion perception and feedback mechanism is provided, adjusting strategies to improve user satisfaction when users experience negative emotions. Furthermore, instruction chain generation logic can automatically sort and schedule operation items, achieving fast and accurate business processing and reducing manual intervention. An event risk classification mechanism ensures that high-risk events are promptly diverted to human customer service, while low / medium-risk events follow intelligent paths, achieving resource optimization.

[0067] In some embodiments, to further address the second technical problem described in the background section, namely, "existing intelligent conversations typically rely solely on single-turn user input for intent recognition, lacking full utilization of multi-turn context and failing to effectively allocate appropriate human customer service representatives in high-risk event handling, resulting in inaccurate intent recognition, unreasonable instruction generation, and difficulty in guaranteeing efficiency and service quality in high-risk event handling," in some embodiments of the present invention, the current basic conversation features are extracted through the following steps:

[0068] Step 1: Extract the conversation data within the preset time window in the smart conversation window, and perform a redundancy deletion operation on the conversation data to obtain the conversation data after redundancy deletion;

[0069] In some embodiments, by removing duplicate or invalid information from the session data, noise interference is avoided, the quality of the input data is improved, and the results of subsequent feature extraction are ensured to be more accurate and efficient. When the number of session rounds is insufficient, the time window is expanded or historical messages are introduced to supplement the context, ensuring that the generated session features contain sufficient contextual information, thereby enhancing semantic integrity. Furthermore, pre-trained models or deep learning methods are used to transform the text into a high-dimensional semantic vector representation, which can accurately represent the semantics of user input and context, providing more effective input for intent recognition. Specifically, the current basic session features are numerical vectors extracted from the intelligent session window, used to represent the user's current input and multi-round session context information, serving as the input basis for explicit and implicit intent recognition. The preset time window refers to the limitation of extracting session data within a specific time range during multi-round sessions, rather than retaining all dialogue indefinitely. The time window can be a fixed duration (e.g., sessions within the past 5 minutes), a fixed number of rounds (e.g., the most recent 5 rounds), or a fixed number of messages (e.g., the most recent 10 messages). Based on this, a window size is set, such as "the most recent 5 rounds." Each time a user inputs data, the backend server extracts only the most recent five rounds of conversation data as context input. Through redundancy removal operations (such as removing duplicate questions, irrelevant short dialogues, or merging similar statements), a more refined conversation context is obtained. Conversation data refers to the interaction information that occurs within the intelligent conversation window, including user input and backend server responses. Redundancy removal removes duplicate, useless, or irrelevant content from the conversation to avoid information redundancy during model processing and affecting feature quality. For example, it removes repeatedly entered identical content (such as "Are you there?", "Are you there?"), removes stop words (such as "Ah", "Hmm"), or filters out small dialogues unrelated to the business (such as "The weather is nice"). The redundancy-removed conversation data refers to the processed, more concise, and valuable conversation data used for subsequent feature extraction; it retains the core semantic information relevant to the task.

[0070] Step 2: If the number of complete dialogue rounds contained in the session data after redundancy removal is less than the second preset round threshold, then the time window is updated based on the preset time window until the number of non-redundant complete dialogue rounds contained in the updated time window is greater than or equal to the second preset round threshold.

[0071] Step 3: Embed words into the conversation data within the update time window to obtain the current basic conversation features.

[0072] In some embodiments, when the number of complete dialogue rounds after redundancy removal is less than a second preset threshold (e.g., the most recent 3 rounds), the backend server can expand the retrieval scope of historical sessions in the time dimension (e.g., if only 2 valid dialogue rounds are identified in the most recent 5 minutes, it can be expanded forward to 10 minutes to supplement 3 complete dialogue rounds, meeting the second preset threshold). Alternatively, it can increase the number of historical messages retrieved in the message dimension (e.g., if only 2 valid business dialogue rounds are included in the user's most recent 5 dialogue rounds, it can be increased to the most recent 10 rounds to supplement 3 complete dialogue rounds). This ensures that the final number of non-redundant complete dialogue rounds is not less than the second preset threshold. A complete dialogue round refers to a complete interaction where the user inputs a sentence and the backend server responds with a sentence. For example, user: "I want to check my electricity bill." Backend server: "Which month's electricity bill do you want to check?" This is one complete round. The second preset threshold is a set lower limit to ensure that there is at least enough context before entering intent recognition. For example, if the second preset threshold is set to 3, at least 3 complete dialogue rounds are required. Time window updates refer to the process where, if the number of effective conversation rounds within the initial time window is too small, the backend server expands to earlier historical conversations until a threshold requirement is met. Time window updates can be implemented by: expanding the time range forward (e.g., from the most recent 5 minutes to the most recent 10 minutes); or increasing the number of conversation rounds (e.g., from the most recent 5 rounds to the most recent 10 rounds). Non-redundant complete conversation rounds refer to the effective conversation rounds after redundancy removal, ensuring they are business-relevant and non-repetitive. Subsequently, the multi-round conversation text within the update time window is encoded into vectors using a pre-trained word embedding model (such as BERT), and then concatenated or average pooled to form the current basic conversation features. For example, if a user continuously sends "The meter isn't working," "Did anyone come to repair it last night?" and a reply, the above processing can generate basic feature vectors for intent recognition, providing reliable input for intent judgment and command chain generation in multi-round conversations. Word embedding refers to mapping words or sentences in text into continuous vector representations, ensuring that semantically similar words are also close in distance in the vector space. The goal of word embedding is to transform discrete linguistic symbols into numerical representations that can be processed by neural networks. The current basic session features refer to the vectorized session features extracted by the backend server within the "update time window" and formed after processing.

[0073] The instruction chain is generated based on the first intent information and the second intent information, including:

[0074] Step 1: The first intent information and the second intent information are matched in a pre-defined set of operation items to obtain multiple matching operation items; each operation item corresponds to a set of instructions.

[0075] Step two: Sort the multiple sets of instructions corresponding to multiple operation items according to the dependencies and time order between instructions to form an instruction chain.

[0076] In some embodiments, the operation item set refers to a business operation mapping table pre-established by the backend server. Each operation item corresponds to a set of specific executable instructions used to complete a specific business processing task. The first intent information and the second intent information are matched with the tags or semantic representations in the operation item set to filter out operation items related to the user's intent. For each matched operation item, its corresponding set of instructions is selected to execute the actual business operation. The matching method can be based on keyword matching or vector similarity calculation, such as calculating the cosine similarity between the intent vector and the operation item vector, and selecting operation items with a similarity higher than a threshold. As an example, a user explicitly requests "report meter repair" in the smart conversation window, while implicit intent recognition infers that the user may want to "query historical repair status." The backend server matches two operation items in the operation item set: ① "generate repair work order" and ② "query historical work order status." Each operation item corresponds to a set of instructions: for example, the instructions for the "generate repair work order" operation item include creating a work order record, assigning repair personnel, and notifying relevant departments; the instructions for the "query historical work order status" operation item include retrieving historical repair records and returning processing status information. Through this matching process, the backend server can generate a complete, multi-dimensional set of business instructions for the user, laying the foundation for the subsequent formation of an ordered instruction chain. The instruction corresponding to an operation item refers to the specific steps or command sequence to implement that operation item, such as generating a work order, querying records, or assigning staff.

[0077] In some embodiments, the dependency relationship between instructions refers to the sequential or logical constraints between different instructions during execution, such as certain instructions requiring the completion of other instructions before execution. Temporal order refers to arranging the execution order of multiple instructions according to the actual business process or processing sequence to ensure reasonable operation and business continuity. Multiple sets of instructions corresponding to matched operation items are analyzed, and the dependency fields of each instruction are extracted, such as the operation item ID (Identifier) ​​that must be executed first and the priority field (e.g., the larger the value, the higher the priority). Then, an instruction dependency graph is constructed: each instruction is treated as a node. If instruction B depends on instruction A for completion, a directed edge A→B is added to the graph. The dependency graph is then topologically sorted to generate an instruction execution sequence, ensuring that each instruction is executed after its dependent instructions are completed. Based on the topological sorting, a secondary sorting can be performed using instruction priority or timestamps to achieve the optimal execution order of business logic. The backend server packages the sorted instruction sequence into an instruction chain object, containing instruction ID, execution order, and dependency information. The instruction chain is a sequence of sorted instructions used to automate the execution of the complete business process, ensuring that each instruction is executed according to its dependency relationship and temporal order. An instruction dependency graph is a Directed Acyclic Graph (DAG) structure used to represent the execution order and dependencies between instructions in a generated instruction chain. For example, a user reports a meter repair and queries its historical status. The backend server matches the operations "Generate Repair Order" and "Query Historical Records". Instruction dependencies are resolved as follows: Creating a repair order must be executed first; assigning repair personnel and notifying departments can be executed in parallel; querying historical records is executed after user confirmation. The DAG is constructed as follows: nodes are Create Repair Order (A), Assign Personnel (B), Notify Department (C), and Query Historical Records (D); edges are A→B, A→C. The instruction chain is generated by topological sorting: A→(B,C in parallel)→D. Finally, the instructions are executed sequentially to ensure that repair order generation, personnel assignment, and historical record querying are completed smoothly according to business logic.

[0078] The method for generating instruction chains based on contextual intent recognition according to the present invention further includes:

[0079] If the event risk level is high, a human chat window is created and the corresponding customer service personnel are matched with the user's chat request. The human chat window is used to display the chat between the customer service personnel and the user.

[0080] In some embodiments, such as Figure 2The flowchart shown illustrates the creation of a human customer service chat window. When an event is classified as high-risk, the backend server creates this window to display the interaction between customer service personnel and the user, ensuring timely handling of high-risk events. To match suitable customer service personnel, the backend server selects based on user attribute information or historical chat records, assigning the most appropriate personnel to high-risk chats and improving the efficiency and quality of human intervention. The human customer service chat window is the interactive interface displayed on the user's terminal (such as a mobile app or webpage), presenting the real-time conversation between the user and a human customer service representative. It supports text, voice, or other interactive formats, ensuring timely human intervention for high-risk events. Therefore, when an event is classified as high-risk, creating a human customer service chat window transfers the user's chat request to a human representative, ensuring human intervention in emergencies or critical situations. This effectively reduces the risks caused by misjudgments or processing delays by intelligent customer service, protecting the power supply system and user safety.

[0081] Customer service personnel are matched through the following steps:

[0082] Matching user session requests with appropriate customer service personnel based on user attribute information or user session history records, including:

[0083] Step 1: If multiple historical conversation records are found for a user, sort them in descending order of the user's feedback rating for each historical conversation record, and select the customer service representative of the historical conversation record with the highest feedback rating as the candidate customer service representative; if the candidate customer service representative is currently available, then determine the matching customer service representative.

[0084] In some embodiments, such as Figure 3The flowchart shown illustrates the customer service matching process. Historical conversation records refer to the user's past interactions with customer service personnel, including conversation content, processing results, and user feedback. Feedback ratings are the user's subjective evaluation of customer service quality after each conversation, typically in numerical form (e.g., 1-5 points), reflecting the efficiency and satisfaction of customer service in handling issues. Based on this, the backend server first queries the database to retrieve all of the user's historical conversation records. If multiple historical conversation records are found, the customer service ID and corresponding user feedback rating for each record are extracted. Multiple historical records are sorted from highest to lowest rating. The customer service personnel corresponding to the record with the highest rating are selected as candidate customer service personnel. The status of the candidate customer service personnel is checked: if the candidate customer service personnel is currently available (not serving other users), they are directly assigned to establish a human customer service conversation with the user; if the candidate customer service personnel is busy, the next matching step is initiated (e.g., continuing to search for the next high-scoring customer service personnel, or switching to attribute matching). Here, candidate customer service personnel refer to the set of customer service personnel who have interacted with the user in the historical conversation records, selected as the highest priority customer service personnel after sorting by feedback rating. "Idle state" refers to a state where customer service personnel are not currently handling other user sessions and can immediately take over new user requests. For example, Mr. Zhang has previously reported meter problems multiple times through the intelligent customer service system: the first time was handled by Customer Service A, with a feedback rating of 3 points; the second time by Customer Service B, with a feedback rating of 5 points (very satisfied); and the third time by Customer Service C, with a feedback rating of 4 points. When Mr. Zhang initiates a high-risk repair request again (such as "meter smoking"), the backend server will query his historical session records and sort the feedback ratings from highest to lowest as follows: Customer Service B (5 points), Customer Service C (4 points), and Customer Service A (3 points). The backend server will consider Customer Service B as a candidate customer service representative. If Customer Service B is currently idle, it will immediately match Mr. Zhang with Customer Service B and generate a human chat window on the user's terminal, allowing Mr. Zhang to communicate directly with the customer service representative who previously provided the best service, improving processing efficiency and user experience.

[0085] Step 2: If no historical session records are found for the user, then select matching customer service personnel from those currently available based on the user's attribute information.

[0086] In some embodiments, if no historical session records are found for the user, all currently available customer service representatives are filtered. Then, matching customer service representatives are selected from the currently available representatives based on user attribute information: first, the candidate customer service representatives are narrowed down based on the user's geographical location; then, customer service representatives skilled in handling that type of business are selected from the candidates based on account type. Finally, based on the skill tags stored by the customer service representatives in the backend server and their correspondence with user attributes, the matching customer service representatives are determined.

[0087] Step 3: If a user has historical conversation records, but the user has not rated the historical conversation records, then select the customer service representative with the highest overall customer service rating from among the multiple customer service representatives corresponding to the historical conversation records and determine them as the matching customer service representative.

[0088] In some embodiments, if a user has historical conversation records but has not provided a rating, the system retrieves the corresponding comprehensive rating from a pre-stored comprehensive rating table based on the customer service ID among multiple customer service personnel corresponding to the historical conversation records. The person with the highest comprehensive rating is selected as the matched customer service personnel. If the highest-rated customer service representative is busy, the next one is selected in turn. The comprehensive rating table includes the customer service ID and the corresponding comprehensive rating. The comprehensive rating refers to the service quality score calculated based on the multi-dimensional performance of the customer service personnel. The comprehensive rating is generally a numerical indicator (e.g., 0 to 100 points) used to measure the overall service level of customer service. For example, user B repeatedly uses the intelligent customer service to inquire about electricity issues but has not previously rated the customer service personnel. The backend server queries historical conversations and finds that the customer service personnel involved are Customer Service Zhang, Customer Service Li, and Customer Service Wang. Because the user has not rated, the backend server retrieves the comprehensive ratings of the three customer service personnel: Customer Service Zhang: 97 (high resolution rate, low complaint rate), Customer Service Li: 95, and Customer Service Wang: 92. The backend server determines that Customer Service Zhang has the highest comprehensive rating and is currently idle, so it matches him as the customer service personnel for user B in this service. If customer service representative Zhang is busy, the system will prioritize selecting customer service representative Li. In short, if a user has multiple historical conversation records, the system selects the customer service representative who previously provided the best service experience based on feedback ratings. This enhances user trust and service continuity, avoids users repeatedly explaining the same issue, and improves problem-solving efficiency. If a user has no historical conversation records, the system precisely filters available customer service representatives based on attributes such as geographical location, account type, and skill tags. This achieves person-job matching, ensuring customer service representatives have relevant business experience and improving the first-response resolution rate. If a user has historical conversation records but has not rated them, a comprehensive rating table maintained in the backend is used for matching. This rating is based on multi-dimensional indicators (such as resolution rate, complaint rate, response speed, etc.) to ensure that the assigned customer service representative has a high overall service level, improving the professionalism and stability of the handling. In addition to considering ratings or matching degrees, the system also checks the availability / busyness of customer service representatives in real time, enabling dynamic scheduling. This avoids user queuing and ensures that high-risk events are addressed immediately.

[0089] The user attribute information includes the user's geographical location and account type, as well as...

[0090] Based on user attribute information, select matching customer service personnel from those currently available, including:

[0091] Based on the customer service personnel who are currently available, determine the corresponding customer service information, which includes the customer service identifier and the skill tag corresponding to the customer service identifier;

[0092] Based on the user's geographical location, the system will match the available customer service personnel to the region to obtain at least one primary candidate customer service personnel.

[0093] If there are multiple first-selection customer service personnel, business matching is performed from the first-selection customer service personnel according to the account type, and at least one second-selection customer service personnel is selected.

[0094] If there are multiple second-selection customer service personnel, the skill tags of the customer service personnel are matched with the event type selected by the user to obtain at least one target customer service personnel; the target customer service personnel are then identified as the matched customer service personnel.

[0095] In some embodiments, user attribute information includes the user's geographic location (e.g., region or power supply bureau jurisdiction) and account type (e.g., residential electricity, industrial electricity, commercial electricity). The user's geographic location refers to their actual location or service area, used to determine the power supply bureau, city, or region to which the user belongs. For example, it can be represented by specific administrative divisions, latitude and longitude coordinates, or power supply bureau codes. Geographic location is mainly used for regional matching, ensuring that user requests are assigned to customer service representatives covering the same area. Account type refers to the user's electricity consumption category or account identity type in the power system, used for business matching. For example, common account types include: residential electricity: electricity for daily household use; commercial electricity: electricity for commercial business premises; industrial electricity: electricity for production plants and enterprises. Based on this, the customer service representative ID and corresponding skill tags of each customer service representative are obtained from the list of currently available customer service representatives. Based on the geographic location information in the user attributes, customer service representatives located in the same region or service area are filtered to obtain the first candidate customer service representatives. If there are multiple customer service representatives among the first candidate representatives, customer service representatives are filtered according to the user's account type, i.e., those skilled in handling business for that account type are selected to obtain the second candidate customer service representatives. If multiple customer service representatives remain in the second pool, their skill tags are matched against the user's current event type, and the representative best suited to handle the event is selected as the target representative. This target representative is then assigned to the user's session request to ensure appropriate human assistance. Customer service information refers to the relevant information recorded on the backend server, including a unique customer service identifier and skill tags (the representative's professional skills or areas of expertise, such as repair requests, complaints, inquiries, etc.). The first pool of potential customer service representatives is the set of candidates selected from currently available representatives based on user geographic location. These representatives cover the user's region and can handle user requests within that region, but may require further business or skill matching. The second pool of potential customer service representatives is the set obtained from the first pool of potential representatives based on account type. These representatives not only cover the user's region but also specialize in handling the user's business type (e.g., residential electricity, industrial electricity). The target customer service representative is the final representative selected from the second pool of potential representatives by matching their skill tags against the event type selected by the user. The target customer service personnel meet the requirements in terms of region, business, and skills, and are identified as the customer service personnel to establish a live chat window with the user. As an example, Mr. Zhang, a user, initiates a repair request on the app, with the event type being "Repair Request," the account type being "Residential Electricity," and the user located in District A of City A. First, all available customer service IDs and skill tags are retrieved from the currently available customer service personnel. Based on Mr. Zhang's geographical location, region matching is performed, resulting in the first candidate customer service personnel A, B, and C. Because Mr. Zhang's account type is "Residential Electricity," further filtering is performed on customer service personnel specializing in residential services, resulting in the second candidate customer service personnel A and B.The skill tags of the second candidate customer service representatives were matched with the event type "repair report." It was found that Customer Service Representative B was skilled in reporting meter repairs, while Customer Service Representative A mainly handled inquiries. Ultimately, Customer Service Representative B was selected as the match, and a live chat window was generated on the user's terminal, allowing Mr. Zhang to communicate directly with Customer Service Representative B for efficient repair processing.

[0096] The overall customer service rating is determined through the following steps:

[0097] Collect historical service data for each of the multiple customer service personnel. The historical service data includes historical response time, problem resolution rate, user satisfaction rating, and complaint rate.

[0098] The various indicators in the historical service data are weighted according to preset weights to obtain the weighted total score for each customer service representative; the weighted total score is used as the overall customer service score.

[0099] In some embodiments, historical service data for each customer service representative is extracted from the database of the backend server. Predefined weights are assigned to each indicator; for example, user satisfaction rating has the highest weight, response time has the second highest weight, complaint rate is a negative indicator, and problem resolution rate also has a weight. Each customer service representative's indicators are multiplied by their corresponding weights and summed to obtain a weighted total score. This weighted total score is directly used as the overall customer service representative score for subsequent matching decisions. Historical service data refers to relevant data generated by customer service representatives in the past when handling user requests, including response time (time required to process a request), problem resolution rate (the proportion of successfully resolved issues), user satisfaction rating (the user's subjective evaluation of the service, which can be 1 to 5 points), complaint rate (the frequency of user complaints), etc. For example, Mr. Zhang, a user, initiates a "meter malfunction" repair request. Historical service data for three potential customer service representatives were collected from the database: Customer Service A, with a response time of 12 minutes, a problem resolution rate of 95%, a user satisfaction rating of 4.8, and a complaint rate of 2%; Customer Service B, with a response time of 8 minutes, a problem resolution rate of 90%, a user satisfaction rating of 4.5, and a complaint rate of 1%; and Customer Service C, with a response time of 15 minutes, a problem resolution rate of 98%, a user satisfaction rating of 4.7, and a complaint rate of 0%. Weights were assigned as follows: response time 10%, problem resolution rate 30%, user satisfaction rating 40%, and complaint rate 20%. A weighted total score was calculated based on these weights. The results showed that Customer Service A had the highest overall score and was currently available. Therefore, Customer Service A was assigned as Mr. Zhang's customer service representative for handling his repair request, ensuring an efficient and satisfactory service experience.

[0100] These embodiments improve the understanding and response capabilities of multi-turn conversations and ensure that high-risk events are handled promptly and appropriately. Specifically, by redundancy removal, time window updates, and word embedding, multi-turn conversation data is transformed into concise and sufficient basic conversation features, ensuring the accuracy of explicit and implicit intent recognition. Multi-turn intent information is matched with a preset set of operation items, and instruction chains are generated based on dependencies and time sequence to automate and accurately execute multi-step business operations. By combining user attribute information and historical conversation records with feedback scores or comprehensive scores, available customer service personnel are intelligently matched, enabling timely and accurate handling of high-risk events while reducing the blindness of manual intervention and improving service efficiency and user satisfaction.

[0101] In some embodiments, to further address the third technical problem described in the background section, namely, "existing technologies evaluate user emotions through single-modal information, leading to inaccurate emotion judgments, thereby affecting timely responses and accurate judgments of user states during intelligent conversations," some embodiments of the present invention include an emotion perception model comprising a semantic emotion recognition model, a facial expression recognition model, and a voice emotion analysis model, and...

[0102] Real-time user sentiment scoring is performed using a sentiment perception model, including:

[0103] During the conversation, a semantic emotion recognition model is used to generate a semantic emotion score;

[0104] Collect user facial expression information and user voice information; analyze facial expression features through a facial expression recognition model to generate facial emotion scores; and generate voice emotion scores through a voice emotion analysis model.

[0105] By integrating facial emotion scores, voice emotion scores, and semantic emotion scores, a user emotion score is generated.

[0106] In some embodiments, an emotion perception model, used for real-time assessment of a user's emotional state, comprises three parts: a semantic emotion recognition model, which analyzes potential emotional information in user-input text (such as text or chat content); a facial expression recognition model, which captures the user's facial expressions via a camera and extracts emotion-related features; and a speech emotion analysis model, which analyzes features such as tone, volume, and speech rate in the user's speech to identify the emotional state from the speech. Specifically, the semantic emotion recognition model can use a pre-trained BERT model to extract text features and then output an emotion category (such as positive, negative, or neutral) or emotion score through a classification layer. The facial expression recognition model, such as the OpenFace model, can map facial keypoint detection and action unit recognition to emotional states (happy, angry, sad, etc.). The speech emotion analysis model, such as OpenSMILE, an open-source audio feature extraction tool, can extract features such as tone, volume, and speech rate. An LSTM based on the extracted features is used for emotion classification or emotion score prediction. Facial emotion scoring refers to quantifying the emotional state reflected by the user's facial expressions into a numerical emotion score. Voice emotion scoring refers to quantifying a user's emotional state into a numerical score by analyzing features such as tone, volume, and speech rate in their voice. Semantic emotion scoring, on the other hand, quantifies a user's emotional state by analyzing the semantic information of the user's input text (such as text content and the emotional tone of words). Based on this, facial emotion scoring, voice emotion scoring, and semantic emotion scoring are weighted and fused to obtain a user's overall emotion score. This weighted fusion can be achieved through methods such as pre-set weighted summation or feature concatenation input fusion models to generate a more accurate user emotion evaluation.

[0107] These embodiments improve the accuracy and real-time performance of emotion recognition. Specifically, text, facial, and speech features are extracted using a semantic emotion recognition model, a facial expression recognition model, and a speech emotion analysis model, respectively. Pre-trained models ensure accurate feature extraction. Emotional information from the three modalities is quantified into comparable scores (facial emotion score, speech emotion score, and semantic emotion score), providing a foundation for fusion. The three emotion scores are fused into a fusion model by weighted summation or feature concatenation using preset weights to obtain the final user emotion score. The fusion method can adaptively adjust the importance of different modalities, improving the accuracy and robustness of emotion assessment. The obtained user emotion score can be used in intelligent conversation windows to dynamically adjust response strategies or trigger enhanced intent recognition, thereby improving interaction quality and user satisfaction.

[0108] The above description is merely a selection of preferred embodiments of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to specific combinations of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A method for generating instruction chains based on contextual intent recognition, characterized in that, include: Upon receiving a user session request, the event type is extracted from the user session request, and the event risk level is determined based on the event type; If the event risk level is low or medium risk, an intelligent conversation window is established and the context data corresponding to the user's current input information is dynamically captured, and basic conversation features are generated. The basic session features are input into the explicit intent recognition network to obtain the first intent information and the first confidence level; Based on the first intent information, a response to the user's question is generated and displayed in the smart conversation window; During the conversation, the user's emotion is scored in real time using an emotion perception model; When a user's sentiment score is lower than a preset score threshold; If the first confidence level is less than a preset threshold and the current session round in the smart session window reaches the first preset round threshold, then an enhanced session feature is constructed and input into the implicit intent recognition network to obtain the second intent information; Based on the first intent information and the second intent information, an instruction chain is generated; The construction of enhanced session features and input into the implicit intent recognition network includes: User features are obtained by extracting features from user emotion ratings and user attribute information; the user features are then concatenated with the current basic conversation features to obtain enhanced conversation features, which are then input into the implicit intent recognition network. The current basic session features are extracted through the following steps: Extract the session data within a preset time window in the smart session window, and perform a redundancy deletion operation on the session data to obtain the session data after redundancy deletion; If the number of complete dialogue rounds contained in the session data after the redundancy removal is less than the second preset round threshold, then the time window is updated based on the preset time window until the number of non-redundant complete dialogue rounds contained in the updated time window is greater than or equal to the second preset round threshold. Word embedding is performed on the conversation data within the update time window to obtain the current basic conversation features; The step of generating an instruction chain based on the first intent information and the second intent information includes: The first intent information and the second intent information are matched in a pre-defined set of operation items to obtain multiple matching operation items; wherein, each operation item corresponds to a set of instructions; Based on the dependencies and time sequence between instructions, multiple sets of instructions corresponding to multiple operation items are sorted to form an instruction chain.

2. The instruction chain generation method based on contextual intent recognition according to claim 1, characterized in that, Also includes: If the risk level of the event is high risk, a human chat window is created and a corresponding customer service representative is matched for the user's chat request. The human chat window is used to display the chat between the customer service representative and the user. The customer service personnel are matched through the following steps: The system matches the user's session request with a corresponding customer service representative based on the user's attribute information or the user's historical session records.

3. The instruction chain generation method based on contextual intent recognition according to claim 2, characterized in that, The step of matching the user's session request with a corresponding customer service representative based on user attribute information or the user's historical session records includes: If a user has multiple historical conversation records, they are sorted from highest to lowest according to the user's feedback rating for each historical conversation record, and the customer service representative with the highest feedback rating is selected as a candidate customer service representative; if the candidate customer service representative is currently available, then the candidate customer service representative is determined to be the matching customer service representative. If no historical session records are found for the user, then a matching customer service representative will be selected from those currently available based on the user's attribute information. If a user has historical conversation records, but the user has not rated these records, the customer service representative with the highest overall rating will be selected from among the multiple customer service representatives associated with the historical conversation records to be the matching representative.

4. The instruction chain generation method based on contextual intent recognition according to claim 3, characterized in that, The user attribute information includes the user's geographical location and account type, and The step of filtering and matching customer service personnel from those currently available based on user attribute information includes: Based on the customer service personnel who are currently available, determine the corresponding customer service information, which includes the customer service identifier and the skill tag corresponding to the customer service identifier; Based on the user's geographical location, the system will match the available customer service personnel to the region to obtain at least one primary candidate customer service personnel. If there are multiple first-selection customer service personnel, business matching is performed from the first-selection customer service personnel according to the account type, and at least one second-selection customer service personnel is selected. If there are multiple second-selection customer service personnel, the skill tags of the customer service personnel are matched with the event type selected by the user to obtain at least one target customer service personnel; the target customer service personnel are then identified as the matched customer service personnel.

5. The instruction chain generation method based on contextual intent recognition according to claim 4, characterized in that, The overall customer service rating is determined through the following steps: Collect historical service data for each customer service representative from multiple customer service personnel. The historical service data includes historical response time, problem resolution rate, user satisfaction rating, and complaint rate. The various indicators in the historical service data are weighted according to preset weights to obtain the weighted total score for each customer service representative; the weighted total score is used as the overall customer service score.

Images